test

Signed-off-by: Roger Wang <hey@rogerw.me>

.buildkite/lm-eval-harness/configs/DeepSeek-V2-Lite-Chat.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash ./run-lm-eval-gsm-vllm-baseline.sh -m deepseek-ai/DeepSeek-V2-Lite-Chat -b "auto" -l 1000 -f 5 -t 2
 model_name: "deepseek-ai/DeepSeek-V2-Lite-Chat"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform.yaml

@@ -1,3 +1,4 @@
+# For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform -b auto -l 1000 -f 5
 model_name: "nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-70B-Instruct.yaml

@@ -1,3 +1,4 @@
+# For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-70B-Instruct -b 32 -l 250 -f 5
 model_name: "meta-llama/Meta-Llama-3-70B-Instruct"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-W8A8-FP8-Channelwise-compressed-tensors -b auto -l 1000 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-Instruct-W8A8-FP8-Channelwise-compressed-tensors"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform -b auto -l 1000 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-Instruct-FBGEMM-nonuniform"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test -b 32 -l 1000 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-FP8.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Meta-Llama-3-8B-Instruct-FP8 -b 32 -l 250 -f 5 -t 1
 model_name: "neuralmagic/Meta-Llama-3-8B-Instruct-FP8"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Asym-Per-Token-Test -b "auto" -l 250 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Asym-Per-Token-Test"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-INT8-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Per-Token-Test -b "auto" -l 250 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-Instruct-W8-Channel-A8-Dynamic-Per-Token-Test"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Meta-Llama-3-8B-Instruct-nonuniform-test -b auto -l 1000 -f 5 -t 1
 model_name: "nm-testing/Meta-Llama-3-8B-Instruct-nonuniform-test"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml

@@ -1,4 +1,5 @@
-# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-8B-Instruct -b 32 -l 250 -f 5 -t 1
+# For hf script, without -t option (tensor parallel size).
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m meta-llama/Meta-Llama-3-8B-Instruct -b 32 -l 250 -f 5
 model_name: "meta-llama/Meta-Llama-3-8B-Instruct"
 tasks:
 - name: "gsm8k"

.buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-QQQ.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m HandH1998/QQQ-Llama-3-8b-g128 -b 32 -l 1000 -f 5 -t 1
 model_name: "HandH1998/QQQ-Llama-3-8b-g128"
 tasks:

.buildkite/lm-eval-harness/configs/Meta-Llama-3.2-1B-Instruct-FP8-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m RedHatAI/Llama-3.2-1B-Instruct-FP8 -b "auto" -l 1319 -f 5 -t 1
+model_name: "RedHatAI/Llama-3.2-1B-Instruct-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.335
+  - name: "exact_match,flexible-extract"
+    value: 0.323
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Llama-3.2-1B-Instruct-quantized.w8a8 -b "auto" -l 1000 -f 5 -t 1
 model_name: "neuralmagic/Llama-3.2-1B-Instruct-quantized.w8a8"
 tasks:

.buildkite/lm-eval-harness/configs/Minitron-4B-Base-FP8.yaml

@@ -1,11 +1,12 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m mgoin/Minitron-4B-Base-FP8 -b auto -l 1000 -f 5 -t 1
 model_name: "mgoin/Minitron-4B-Base-FP8"
 tasks:
 - name: "gsm8k"
   metrics:
   - name: "exact_match,strict-match"
-    value: 0.233
+    value: 0.231
   - name: "exact_match,flexible-extract"
-    value: 0.236
+    value: 0.22
 limit: 1000
 num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x22B-Instruct-v0.1-FP8-Dynamic.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic -b "auto" -l 250 -f 5 -t 8
 model_name: "neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic"
 tasks:

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1-FP8.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8 -b "auto" -l 250 -f 5 -t 4
 model_name: "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8"
 tasks:

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1.yaml

@@ -1,4 +1,5 @@
-# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1 -b 32 -l 250 -f 5 -t 4
+# For hf script, without -t option (tensor parallel size).
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-hf-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1 -b 32 -l 250 -f 5
 model_name: "mistralai/Mixtral-8x7B-Instruct-v0.1"
 tasks:
 - name: "gsm8k"

.buildkite/lm-eval-harness/configs/Qwen1.5-MoE-W4A16-compressed-tensors.yaml

@@ -0,0 +1,12 @@
+# For vllm script, with -t option (tensor parallel size).
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16 -b auto -l 1319 -f 5 -t 1
+model_name: "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.30
+  - name: "exact_match,flexible-extract"
+    value: 0.465
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-FP8W8.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-FP8W8 -b auto -l 1000 -f 5 -t 1
 model_name: "nm-testing/Qwen2-1.5B-Instruct-FP8W8"
 tasks:

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Qwen2-1.5B-Instruct-quantized.w8a8 -b "auto" -l 1000 -f 5 -t 1
 model_name: "neuralmagic/Qwen2-1.5B-Instruct-quantized.w8a8"
 tasks:

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-W8A16-compressed-tensors.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise -b "auto" -l 1000 -f 5 -t 1
 model_name: "nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise"
 tasks:

.buildkite/lm-eval-harness/configs/Qwen2-57B-A14-Instruct.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash ./run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen2-57B-A14B-Instruct -b "auto" -l 250 -f 5 -t 4
 model_name: "Qwen/Qwen2-57B-A14B-Instruct"
 tasks:

.buildkite/lm-eval-harness/configs/Qwen2.5-1.5B-Instruct.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen2.5-1.5B-Instruct -b auto -l 1319 -f 5 -t 1
+model_name: "Qwen/Qwen2.5-1.5B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.54
+  - name: "exact_match,flexible-extract"
+    value: 0.59
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m RedHatAI/Qwen2.5-VL-3B-Instruct-FP8-Dynamic -b auto -l 1319 -f 5 -t 1
+model_name: "RedHatAI/Qwen2.5-VL-3B-Instruct-FP8-Dynamic"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.47
+  - name: "exact_match,flexible-extract"
+    value: 0.64
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/SparseLlama3.1_2of4_fp8_compressed.yaml

@@ -1,3 +1,4 @@
+# For vllm script, with -t option (tensor parallel size).
 # bash ./run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM -b "auto" -t 2
 model_name: "nm-testing/SparseLlama-3.1-8B-gsm8k-pruned.2of4-chnl_wts_per_tok_dyn_act_fp8-BitM"
 tasks:

.buildkite/lm-eval-harness/configs/models-large.txt

@@ -3,3 +3,4 @@ Meta-Llama-3-70B-Instruct.yaml
 Mixtral-8x7B-Instruct-v0.1.yaml
 Qwen2-57B-A14-Instruct.yaml
 DeepSeek-V2-Lite-Chat.yaml
+Meta-Llama-3-8B-QQQ.yaml

.buildkite/lm-eval-harness/configs/models-small.txt

@@ -1,10 +1,6 @@
-Meta-Llama-3-8B-Instruct.yaml
-Meta-Llama-3-8B-Instruct-FP8-compressed-tensors.yaml
+Qwen2.5-1.5B-Instruct.yaml
 Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
-Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
-Minitron-4B-Base-FP8.yaml
-Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
-Qwen2-1.5B-Instruct-FP8W8.yaml
-Meta-Llama-3-8B-QQQ.yaml
+Qwen2.5-VL-3B-Instruct-FP8-dynamic.yaml
+Qwen1.5-MoE-W4A16-compressed-tensors.yaml

.buildkite/lm-eval-harness/conftest.py

@@ -0,0 +1,39 @@
+# SPDX-License-Identifier: Apache-2.0
+from pathlib import Path
+
+import pytest
+
+
+def pytest_addoption(parser):
+    parser.addoption(
+        "--config-list-file",
+        action="store",
+        help="Path to the file listing model config YAMLs (one per line)")
+    parser.addoption("--tp-size",
+                     action="store",
+                     default="1",
+                     help="Tensor parallel size to use for evaluation")
+
+
+@pytest.fixture(scope="session")
+def config_list_file(pytestconfig, config_dir):
+    rel_path = pytestconfig.getoption("--config-list-file")
+    return config_dir / rel_path
+
+
+@pytest.fixture(scope="session")
+def tp_size(pytestconfig):
+    return pytestconfig.getoption("--tp-size")
+
+
+def pytest_generate_tests(metafunc):
+    if "config_filename" in metafunc.fixturenames:
+        rel_path = metafunc.config.getoption("--config-list-file")
+        config_list_file = Path(rel_path).resolve()
+        config_dir = config_list_file.parent
+        with open(config_list_file, encoding="utf-8") as f:
+            configs = [
+                config_dir / line.strip() for line in f
+                if line.strip() and not line.startswith("#")
+            ]
+        metafunc.parametrize("config_filename", configs)

.buildkite/lm-eval-harness/run-tests.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-
-usage() {
-    echo``
-    echo "Runs lm eval harness on GSM8k using vllm and compares to "
-    echo "precomputed baseline (measured by HF transformers.)"
-    echo
-    echo "usage: ${0} <options>"
-    echo
-    echo "  -c    - path to the test data config (e.g. configs/small-models.txt)"
-    echo "  -t    - tensor parallel size"
-    echo
-}
-
-SUCCESS=0
-
-while getopts "c:t:" OPT; do
-  case ${OPT} in
-    c ) 
-        CONFIG="$OPTARG"
-        ;;
-    t )
-        TP_SIZE="$OPTARG"
-        ;;
-    \? )
-        usage
-        exit 1
-        ;;
-  esac
-done
-
-# Parse list of configs.
-IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < "$CONFIG"
-
-for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
-do
-    LOCAL_SUCCESS=0
-    
-    echo "=== RUNNING MODEL: $MODEL_CONFIG WITH TP SIZE: $TP_SIZE==="
-
-    export LM_EVAL_TEST_DATA_FILE=$PWD/configs/${MODEL_CONFIG}
-    export LM_EVAL_TP_SIZE=$TP_SIZE
-    pytest -s test_lm_eval_correctness.py || LOCAL_SUCCESS=$?
-
-    if [[ $LOCAL_SUCCESS == 0 ]]; then
-        echo "=== PASSED MODEL: ${MODEL_CONFIG} ==="
-    else
-        echo "=== FAILED MODEL: ${MODEL_CONFIG} ==="
-    fi
-
-    SUCCESS=$((SUCCESS + LOCAL_SUCCESS))
-
-done
-
-if [ "${SUCCESS}" -eq "0" ]; then
-    exit 0
-else
-    exit 1
-fi

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -3,34 +3,25 @@
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml
 
-* export LM_EVAL_TEST_DATA_FILE=configs/Meta-Llama-3-70B-Instruct.yaml
-* export LM_EVAL_TP_SIZE=4 
-* pytest -s test_lm_eval_correctness.py
+pytest -s -v test_lm_eval_correctness.py \
+    --config-list-file=configs/models-small.txt \
+    --tp-size=1
 """
 
-import os
-from pathlib import Path
-
 import lm_eval
-import numpy
+import numpy as np
 import yaml
 
-RTOL = 0.05
-TEST_DATA_FILE = os.environ.get(
-    "LM_EVAL_TEST_DATA_FILE",
-    ".buildkite/lm-eval-harness/configs/Meta-Llama-3-8B-Instruct.yaml")
-
-TP_SIZE = os.environ.get("LM_EVAL_TP_SIZE", 1)
+RTOL = 0.08
 
 
-def launch_lm_eval(eval_config):
+def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get('trust_remote_code', False)
-
     model_args = f"pretrained={eval_config['model_name']}," \
-                 f"tensor_parallel_size={TP_SIZE}," \
+                 f"tensor_parallel_size={tp_size}," \
+                 f"enforce_eager=true," \
                  f"add_bos_token=true," \
                  f"trust_remote_code={trust_remote_code}"
-
     results = lm_eval.simple_evaluate(
         model="vllm",
         model_args=model_args,
@@ -38,18 +29,14 @@ def launch_lm_eval(eval_config):
         num_fewshot=eval_config["num_fewshot"],
         limit=eval_config["limit"],
         batch_size="auto")
-
     return results
 
 
-def test_lm_eval_correctness():
-    eval_config = yaml.safe_load(
-        Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
+def test_lm_eval_correctness_param(config_filename, tp_size):
+    eval_config = yaml.safe_load(config_filename.read_text(encoding="utf-8"))
 
-    # Launch eval requests.
-    results = launch_lm_eval(eval_config)
+    results = launch_lm_eval(eval_config, tp_size)
 
-    # Confirm scores match ground truth.
     success = True
     for task in eval_config["tasks"]:
         for metric in task["metrics"]:
@@ -57,8 +44,7 @@ def test_lm_eval_correctness():
             measured_value = results["results"][task["name"]][metric["name"]]
             print(f'{task["name"]} | {metric["name"]}: '
                   f'ground_truth={ground_truth} | measured={measured_value}')
-            success = success and numpy.isclose(
+            success = success and np.isclose(
                 ground_truth, measured_value, rtol=RTOL)
 
-    # Assert at the end, print all scores even on failure for debugging.
     assert success

.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh

@@ -426,7 +426,7 @@ main() {
 
   pip install -U transformers
 
-  pip install -r requirements-dev.txt
+  pip install -r requirements/dev.txt
   which genai-perf
 
   # check storage

.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh

@@ -10,15 +10,24 @@ set -x
 set -o pipefail
 
 check_gpus() {
-  # check the number of GPUs and GPU type.
-  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if command -v nvidia-smi; then
+    # check the number of GPUs and GPU type.
+    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  elif command -v amd-smi; then
+    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
+  fi
+
   if [[ $gpu_count -gt 0 ]]; then
     echo "GPU found."
   else
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  if command -v nvidia-smi; then
+    declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  elif command -v amd-smi; then
+    declare -g gpu_type=$(amd-smi static -g 0 -a | grep 'MARKET_NAME' | awk '{print $2}')
+  fi
   echo "GPU type is $gpu_type"
 }
 
@@ -90,9 +99,15 @@ kill_gpu_processes() {
 
 
   # wait until GPU memory usage smaller than 1GB
-  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
-    sleep 1
-  done
+  if command -v nvidia-smi; then
+    while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
+      sleep 1
+    done
+  elif command -v amd-smi; then
+    while [ "$(amd-smi metric -g 0 | grep 'USED_VRAM' | awk '{print $2}')" -ge 1000 ]; do
+      sleep 1
+    done
+  fi
 
   # remove vllm config file
   rm -rf ~/.config/vllm
@@ -361,7 +376,7 @@ main() {
   # get the current IP address, required by benchmark_serving.py
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
   # turn of the reporting of the status of each request, to clean up the terminal output
-  export VLLM_LOG_LEVEL="WARNING"
+  export VLLM_LOGGING_LEVEL="WARNING"
 
   # prepare for benchmarking
   cd benchmarks || exit 1

.buildkite/nightly-benchmarks/tests/serving-tests.json

@@ -63,10 +63,12 @@
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",
             "disable_log_requests": "", 
             "tensor_parallel_size": 4,
-            "swap_space": 16, 
-            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
-            "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1
+            "swap_space": 16,
+            "speculative_config": {
+                "model": "turboderp/Qwama-0.5B-Instruct",
+                "num_speculative_tokens": 4,
+                "draft_tensor_parallel_size": 1
+            }
         },
         "client_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",

.buildkite/release-pipeline.yaml

@@ -1,23 +1,23 @@
 steps:
-  - label: "Build wheel - CUDA 12.4"
+  - label: "Build wheel - CUDA 12.8"
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
-  - label: "Build wheel - CUDA 12.1"
+  - label: "Build wheel - CUDA 12.6"
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.6.3 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -31,10 +31,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=11.8.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -48,7 +48,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Build and publish TPU release image"
@@ -57,7 +57,9 @@ steps:
     agents:
       queue: tpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f Dockerfile.tpu ."
+      - "yes | docker system prune -a"
+      - "git fetch --all"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"
       - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
     plugins:
@@ -82,7 +84,22 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"
+
+  - block: "Build Neuron release image"
+    key: block-neuron-release-image-build
+    depends_on: ~
+
+  - label: "Build and publish Neuron release image"
+    depends_on: block-neuron-release-image-build
+    agents:
+      queue: neuron-postmerge
+    commands:
+      - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest --progress plain -f docker/Dockerfile.neuron ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version)"
+    env:
+      DOCKER_BUILDKIT: "1"

.buildkite/run-openvino-test.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-# This script build the OpenVINO docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Try building the docker image
-docker build -t openvino-test -f Dockerfile.openvino .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f openvino-test || true; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic/generate.py --model facebook/opt-125m

.buildkite/run-tpu-test.sh

@@ -1,26 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# Build the docker image.
-docker build -f Dockerfile.tpu -t vllm-tpu .
-
-# Set up cleanup.
-remove_docker_container() { docker rm -f tpu-test || true; }
-trap remove_docker_container EXIT
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
-
-# For HF_TOKEN.
-source /etc/environment
-# Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it \
-    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
-    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py \
-    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
-    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
-    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/run-xpu-test.sh

@@ -1,19 +0,0 @@
-#!/bin/bash
-
-# This script build the CPU docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Try building the docker image
-docker build -t xpu-test -f Dockerfile.xpu .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f xpu-test || true; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Run the image and test offline inference/tensor parallel
-docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
-'

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -75,42 +75,91 @@ HF_MOUNT="/root/.cache/huggingface"
 commands=$@
 echo "Commands:$commands"
 #ignore certain kernels tests
-if [[ $commands == *" kernels "* ]]; then
+if [[ $commands == *" kernels/core"* ]]; then
   commands="${commands} \
-  --ignore=kernels/test_attention_selector.py \
-  --ignore=kernels/test_blocksparse_attention.py \
-  --ignore=kernels/test_causal_conv1d.py \
-  --ignore=kernels/test_cutlass.py \
-  --ignore=kernels/test_encoder_decoder_attn.py \
-  --ignore=kernels/test_flash_attn.py \
-  --ignore=kernels/test_flashinfer.py \
-  --ignore=kernels/test_int8_quant.py \
-  --ignore=kernels/test_machete_gemm.py \
-  --ignore=kernels/test_mamba_ssm.py \
-  --ignore=kernels/test_marlin_gemm.py \
-  --ignore=kernels/test_moe.py \
-  --ignore=kernels/test_prefix_prefill.py \
-  --ignore=kernels/test_rand.py \
-  --ignore=kernels/test_sampler.py \
-  --ignore=kernels/test_cascade_flash_attn.py \
-  --ignore=kernels/test_mamba_mixer2.py \
-  --ignore=kernels/test_aqlm.py \
-  --ignore=kernels/test_machete_mm.py \
-  --ignore=kernels/test_mha_attn.py \
-  --ignore=kernels/test_block_fp8.py \
-  --ignore=kernels/test_permute_cols.py"
+  --ignore=kernels/core/test_fused_quant_layernorm.py \
+  --ignore=kernels/core/test_permute_cols.py"
 fi
 
-#ignore certain Entrypoints tests
+if [[ $commands == *" kernels/attention"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/attention/stest_attention_selector.py \
+  --ignore=kernels/attention/test_blocksparse_attention.py \
+  --ignore=kernels/attention/test_encoder_decoder_attn.py \
+  --ignore=kernels/attention/test_attention_selector.py \
+  --ignore=kernels/attention/test_flash_attn.py \
+  --ignore=kernels/attention/test_flashinfer.py \
+  --ignore=kernels/attention/test_prefix_prefill.py \
+  --ignore=kernels/attention/test_cascade_flash_attn.py \
+  --ignore=kernels/attention/test_mha_attn.py \
+  --ignore=kernels/attention/test_lightning_attn.py \
+  --ignore=kernels/attention/test_attention.py"
+fi
+
+if [[ $commands == *" kernels/quantization"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/quantization/test_int8_quant.py \
+  --ignore=kernels/quantization/test_aqlm.py \
+  --ignore=kernels/quantization/test_machete_mm.py \
+  --ignore=kernels/quantization/test_block_fp8.py \
+  --ignore=kernels/quantization/test_block_int8.py \
+  --ignore=kernels/quantization/test_marlin_gemm.py \
+  --ignore=kernels/quantization/test_cutlass_scaled_mm.py \
+  --ignore=kernels/quantization/test_int8_kernel.py"
+fi
+
+if [[ $commands == *" kernels/mamba"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/mamba/test_mamba_mixer2.py \
+  --ignore=kernels/mamba/test_causal_conv1d.py \
+  --ignore=kernels/mamba/test_mamba_ssm_ssd.py"
+fi
+
+if [[ $commands == *" kernels/moe"* ]]; then
+  commands="${commands} \
+  --ignore=kernels/moe/test_moe.py \
+  --ignore=kernels/moe/test_cutlass_moe.py \
+  --ignore=kernels/moe/test_triton_moe_ptpc_fp8.py"
+fi
+
+#ignore certain Entrypoints/openai tests
 if [[ $commands == *" entrypoints/openai "* ]]; then
   commands=${commands//" entrypoints/openai "/" entrypoints/openai \
-  --ignore=entrypoints/openai/test_accuracy.py \
   --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_encoder_decoder.py \
-  --ignore=entrypoints/openai/test_embedding.py \
-  --ignore=entrypoints/openai/test_oot_registration.py "}
+  --ignore=entrypoints/openai/test_shutdown.py \
+  --ignore=entrypoints/openai/test_completion.py \
+  --ignore=entrypoints/openai/test_sleep.py \
+  --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_lora_adapters.py \
+  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+  --ignore=entrypoints/openai/test_root_path.py \
+  --ignore=entrypoints/openai/test_tokenization.py \
+  --ignore=entrypoints/openai/test_prompt_validation.py "}
 fi
 
+#ignore certain Entrypoints/llm tests
+if [[ $commands == *" entrypoints/llm "* ]]; then
+  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
+  --ignore=entrypoints/llm/test_chat.py \
+  --ignore=entrypoints/llm/test_accuracy.py \
+  --ignore=entrypoints/llm/test_init.py \
+  --ignore=entrypoints/llm/test_generate_multiple_loras.py \
+  --ignore=entrypoints/llm/test_prompt_validation.py "}
+fi
+
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
+# --ignore=entrypoints/openai/test_encoder_decoder.py \
+# --ignore=entrypoints/openai/test_embedding.py \
+# --ignore=entrypoints/openai/test_oot_registration.py
+# --ignore=entrypoints/openai/test_accuracy.py \
+# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
+
+
 PARALLEL_JOB_COUNT=8
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
@@ -120,9 +169,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     # assign shard-id for each shard
     commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
+    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
-        --device /dev/kfd --device /dev/dri \
-        --network host \
+        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+        --network=host \
         --shm-size=16gb \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
@@ -149,9 +199,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     fi
   done
 else
+  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd --device /dev/dri \
-          --network host \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --network=host \
           --shm-size=16gb \
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+
+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Setup cleanup
+remove_docker_container() {
+  if [[ -n "$container_id" ]]; then
+      podman rm -f "$container_id" || true
+  fi
+  podman system prune -f
+}
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Try building the docker image
+podman build -t cpu-test-ubi9-ppc -f docker/Dockerfile.ppc64le .
+
+# Run the image
+container_id=$(podman run -itd --entrypoint /bin/bash -v /tmp/:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN cpu-test-ubi9-ppc)
+
+function cpu_tests() {
+
+  # offline inference
+  podman exec -it "$container_id" bash -c "
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
+
+  # Run basic model test
+  podman exec -it "$container_id" bash -c "
+    set -e
+    pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
+    pip install sentence-transformers datamodel_code_generator
+    pytest -v -s tests/models/embedding/language/test_cls_models.py::test_classification_models[float-jason9693/Qwen2.5-1.5B-apeach]
+    pytest -v -s tests/models/embedding/language/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model"
+}
+
+# All of CPU tests are expected to be finished less than 40 mins.
+
+export container_id
+export -f cpu_tests
+timeout 40m bash -c cpu_tests
+

.buildkite/scripts/hardware_ci/run-cpu-test-s390x.sh

@@ -10,5 +10,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-docker build -t cpu-test -f Dockerfile.ppc64le .
-
+docker build -t cpu-test -f docker/Dockerfile.s390x .

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -8,24 +8,29 @@ set -ex
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
-# Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .
-
 # Setup cleanup
-remove_docker_container() { set -e; docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; }
+remove_docker_container() { 
+    set -e; 
+    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
+    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
+}
 trap remove_docker_container EXIT
 remove_docker_container
 
+# Try building the docker image
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+
 # Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
- --cpuset-mems="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
+ --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
 docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE" \
- --cpuset-mems="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
+ --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
+  export BUILDKITE_BUILD_NUMBER=$3
 
   # offline inference
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
@@ -35,7 +40,8 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements-test.txt
+    pytest -v -s tests/kernels/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
     pytest -v -s tests/models/decoder_only/language -m cpu_model
     pytest -v -s tests/models/embedding/language -m cpu_model
     pytest -v -s tests/models/encoder_decoder/language -m cpu_model
@@ -85,4 +91,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE $BUILDKITE_BUILD_NUMBER"

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -9,11 +9,13 @@ python3 use_existing_torch.py
 
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
+  --file docker/Dockerfile \
   --target vllm-openai \
   --platform "linux/arm64" \
   -t gh200-test \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
+  --build-arg RUN_WHEEL_CHECK=false \
   --build-arg torch_cuda_arch_list="9.0+PTX" \
   --build-arg vllm_fa_cmake_gpu_arches="90-real"
 
@@ -23,6 +25,6 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and test offline inference
-docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+docker run -e HF_TOKEN -e VLLM_WORKER_MULTIPROC_METHOD=spawn -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
     python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '

.buildkite/scripts/hardware_ci/run-hpu-test.sh

@@ -5,7 +5,7 @@
 set -ex
 
 # Try building the docker image
-docker build -t hpu-test-env -f Dockerfile.hpu .
+docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can

.buildkite/scripts/hardware_ci/run-neuron-test.sh

@@ -35,7 +35,7 @@ else
     date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
 
-docker build -t "${image_name}" -f Dockerfile.neuron .
+docker build -t "${image_name}" -f docker/Dockerfile.neuron .
 
 # Setup cleanup
 remove_docker_container() {
@@ -44,11 +44,11 @@ remove_docker_container() {
 trap remove_docker_container EXIT
 
 # Run the image
-docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
+docker run --rm -it --device=/dev/neuron0 --network bridge \
        -v "${HF_CACHE}:${HF_MOUNT}" \
        -e "HF_HOME=${HF_MOUNT}" \
        -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
        -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
        --name "${container_name}" \
        ${image_name} \
-       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/ -v --capture=tee-sys"
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys && python3 -m pytest /workspace/vllm/tests/neuron/2_core/ -v --capture=tee-sys"

.buildkite/scripts/hardware_ci/run-tpu-v1-test.sh

@@ -0,0 +1,103 @@
+#!/bin/bash
+
+set -xu
+
+# Build the docker image.
+docker build -f docker/Dockerfile.tpu -t vllm-tpu .
+
+# Set up cleanup.
+remove_docker_container() { docker rm -f tpu-test || true; }
+trap remove_docker_container EXIT
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+# For HF_TOKEN.
+source /etc/environment
+# Run a simple end-to-end example.
+docker run --privileged --net host --shm-size=16G -it \
+    -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
+    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
+    && python3 -m pip install pytest pytest-asyncio tpu-info \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && export VLLM_XLA_CACHE_PATH= \
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo HARDWARE \
+    && tpu-info \
+    && { \
+        echo TEST_0: Running test_perf.py; \
+        pytest -s -v /workspace/vllm/tests/tpu/test_perf.py; \
+        echo TEST_0_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_1: Running test_compilation.py; \
+        pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py; \
+        echo TEST_1_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_2: Running test_basic.py; \
+        pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py; \
+        echo TEST_2_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_3: Running test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
+        pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
+        echo TEST_3_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_4: Running test_quantization_accuracy.py; \
+        pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py; \
+        echo TEST_4_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_5: Running examples/offline_inference/tpu.py; \
+        python3 /workspace/vllm/examples/offline_inference/tpu.py; \
+        echo TEST_5_EXIT_CODE: \$?; \
+    } & \
+    { \
+        echo TEST_6: Running test_tpu_model_runner.py; \
+        pytest -s -v /workspace/vllm/tests/tpu/worker/test_tpu_model_runner.py; \
+        echo TEST_6_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_7: Running test_sampler.py; \
+        pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py; \
+        echo TEST_7_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_8: Running test_topk_topp_sampler.py; \
+        pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py; \
+        echo TEST_8_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_9: Running test_multimodal.py; \
+        pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py; \
+        echo TEST_9_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_10: Running test_pallas.py; \
+        pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py; \
+        echo TEST_10_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_11: Running test_struct_output_generate.py; \
+        pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py; \
+        echo TEST_11_EXIT_CODE: \$?; \
+    } & \
+    && { \
+        echo TEST_12: Running test_moe_pallas.py; \
+        pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py; \
+        echo TEST_12_EXIT_CODE: \$?; \
+    } & \
+    # Disable the TPU LoRA tests until the feature is activated
+    # && { \
+    #     echo TEST_13: Running test_moe_pallas.py; \
+    #     pytest -s -v /workspace/vllm/tests/tpu/lora/; \
+    #     echo TEST_13_EXIT_CODE: \$?; \
+    # } & \
+    wait \
+    && echo 'All tests have attempted to run. Check logs for individual test statuses and exit codes.' \
+"
+
+# TODO: This test fails because it uses RANDOM_SEED sampling
+# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+# This script build the CPU docker image and run the offline inference inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
+container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+# Try building the docker image
+docker build -t ${image_name} -f docker/Dockerfile.xpu .
+
+# Setup cleanup
+remove_docker_container() { 
+  docker rm -f "${container_name}" || true; 
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
+}
+trap remove_docker_container EXIT
+
+# Run the image and test offline inference/tensor parallel
+docker run \
+    --device /dev/dri \
+    -v /dev/dri/by-path:/dev/dri/by-path \
+    --entrypoint="" \
+    --name "${container_name}" \
+    "${image_name}" \
+    sh -c '
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
+    VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
+'

.buildkite/scripts/run-benchmarks.sh

@@ -5,8 +5,8 @@
 set -ex
 set -o pipefail
 
-# cd into parent directory of this file
-cd "$(dirname "${BASH_SOURCE[0]}")/.."
+# cd 2 levels into the working directory
+cd "$(dirname "${BASH_SOURCE[0]}")/../.."
 
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 

.buildkite/scripts/run-multi-node-test.sh

@@ -3,7 +3,7 @@
 set -euox pipefail
 
 if [[ $# -lt 4 ]]; then
-    echo "Usage: .buildkite/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
+    echo "Usage: .buildkite/scripts/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
     exit 1
 fi
 

.buildkite/scripts/upload-wheels.sh

@@ -50,11 +50,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu121"* ]]; then
-    # if $normal_wheel matches cu121, do not upload the index.html
-    echo "Skipping index files for cu121 wheels"
+elif [[ $normal_wheel == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
 else
-    # only upload index.html for cu124 wheels (default wheels)
+    # only upload index.html for cu128 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@@ -66,12 +66,12 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
-elif [[ $normal_wheel == *"cu121"* ]]; then
-    # if $normal_wheel matches cu121, do not upload the index.html
-    echo "Skipping index files for cu121 wheels"
+elif [[ $normal_wheel == *"cu126"* ]]; then
+    # if $normal_wheel matches cu126, do not upload the index.html
+    echo "Skipping index files for cu126 wheels"
 else
-    # only upload index.html for cu124 wheels (default wheels)
+    # only upload index.html for cu128 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 
-aws s3 cp "$wheel" "s3://vllm-wheels/$version/"
+aws s3 cp "$wheel" "s3://vllm-wheels/$version/"

.buildkite/test-pipeline.yaml

@@ -8,6 +8,7 @@
 # Documentation
 # label(str): the name of the test. emoji allowed.
 # fast_check(bool): whether to run this on each commit on fastcheck pipeline.
+# torch_nightly(bool): whether to run this on vllm against torch nightly pipeline.
 # fast_check_only(bool): run this test on fastcheck pipeline only
 # optional(bool): never run this test by default (i.e. need to unblock manually) unless it's scheduled nightly run.
 # command(str): the single command to run for tests. incompatible with commands.
@@ -35,13 +36,12 @@ steps:
   fast_check: true
   no_gpu: True
   commands:
-  - pip install -r requirements-docs.txt
+  - pip install -r ../../requirements/docs.txt
   - SPHINXOPTS=\"-W\" make html
   # Check API reference (if it fails, you may have missing mock imports)
-  - grep \"sig sig-object py\" build/html/api/inference_params.html
+  - grep \"sig sig-object py\" build/html/api/vllm/vllm.sampling_params.html
 
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
-  fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/mq_llm_engine
@@ -71,6 +71,7 @@ steps:
 - label: Basic Correctness Test # 30min
   #mirror_hardwares: [amd]
   fast_check: true
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
@@ -78,6 +79,7 @@ steps:
   - tests/basic_correctness/test_preemption
   - tests/basic_correctness/test_cumem.py
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
@@ -104,7 +106,8 @@ steps:
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
   fast_check: true
-  mirror_hardwares: [amd]
+  torch_nightly: true
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
@@ -112,19 +115,19 @@ steps:
   - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
-  - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_openai_schema.py
   - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
-  fast_check: true
   source_file_dependencies:
   - vllm/distributed/
   - vllm/core/
@@ -135,38 +138,41 @@ steps:
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/test_async_llm_dp.py
   commands:
-  - VLLM_USE_V1=1 python3 ../examples/offline_inference/data_parallel.py
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with internal dp
+  - python3 ../examples/offline_inference/data_parallel.py
+  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
-  - python3 ../examples/offline_inference/rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/rlhf_colocate.py
+  - pushd ../examples/offline_inference
+  - python3 rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - popd
 
 - label: Metrics, Tracing Test # 10min
+  mirror_hardwares: [amd]
   num_gpus: 2
-  fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/metrics
   - tests/tracing
   commands:
   - pytest -v -s metrics
-  - "pip install \
-      'opentelemetry-sdk>=1.26.0,<1.27.0' \
-      'opentelemetry-api>=1.26.0,<1.27.0' \
-      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -196,15 +202,20 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - VLLM_USE_V1=1 pytest -v -s v1/core
-    - VLLM_USE_V1=1 pytest -v -s v1/engine
-    - VLLM_USE_V1=1 pytest -v -s v1/sample
-    - VLLM_USE_V1=1 pytest -v -s v1/worker
-    - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
-    - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/core
+    - pytest -v -s v1/engine
+    - pytest -v -s v1/entrypoints
+    - pytest -v -s v1/sample
+    - pytest -v -s v1/worker
+    - pytest -v -s v1/structured_output
+    - pytest -v -s v1/spec_decode
+    - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/test_stats.py
+    - pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/test_oracle.py
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
-    - VLLM_USE_V1=1 pytest -v -s v1/e2e
+    - pytest -v -s v1/e2e
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-neuralmagic/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -222,14 +233,17 @@ steps:
     - python3 offline_inference/basic/chat.py
     - python3 offline_inference/prefix_caching.py
     - python3 offline_inference/llm_engine_example.py
-    - python3 offline_inference/vision_language.py
-    - python3 offline_inference/vision_language_multi_image.py
-    - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_embedding.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - VLLM_USE_V1=0 python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference/encoder_decoder.py
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
-    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
   mirror_hardwares: [amd]
@@ -271,15 +285,25 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py  --ignore=lora/test_transfomers_model.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
   parallelism: 4
 
+- label: PyTorch Compilation Unit Tests
+  torch_nightly: true
+  source_file_dependencies:
+    - vllm/
+    - tests/compile
+  commands:
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+    - pytest -v -s compile/test_sequence_parallelism.py
+
 - label: PyTorch Fullgraph Smoke Test # 9min
-  fast_check: true
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -290,24 +314,61 @@ steps:
   - pytest -v -s compile/piecewise/test_toy_llama.py
 
 - label: PyTorch Fullgraph Test # 18min
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
   commands:
   - pytest -v -s compile/test_full_graph.py
 
-- label: Kernels Test %N # 1h each
+- label: Kernels Core Operation Test
   mirror_hardwares: [amd]
   source_file_dependencies:
   - csrc/
-  - vllm/attention
-  - tests/kernels
+  - tests/kernels/core
   commands:
-    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 4
+    - pytest -v -s kernels/core
+
+- label: Kernels Attention Test %N
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - csrc/attention/
+  - vllm/attention
+  - vllm/v1/attention
+  - tests/kernels/attention
+  commands:
+    - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
+
+- label: Kernels Quantization Test %N
+  mirror_hardwares: [amd]
+  source_file_dependencies:
+  - csrc/quantization/
+  - vllm/model_executor/layers/quantization
+  - tests/kernels/quantization
+  commands:
+    - pytest -v -s kernels/quantization  --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 2
+
+- label: Kernels MoE Test
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - csrc/moe/
+  - tests/kernels/moe
+  - vllm/model_executor/layers/fused_moe/
+  commands:
+    - pytest -v -s kernels/moe
+
+- label: Kernels Mamba Test
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - csrc/mamba/
+  - tests/kernels/mamba
+  commands:
+    - pytest -v -s kernels/mamba
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -323,14 +384,22 @@ steps:
   source_file_dependencies:
   - benchmarks/
   commands:
-  - bash run-benchmarks.sh
+  - bash scripts/run-benchmarks.sh
 
-- label: Quantization Test # 33min
+- label: Benchmarks CLI Test # 10min
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+  - pytest -v -s benchmarks/
+
+- label: Quantization Test
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
   - tests/quantization
-  command: VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
+  commands:
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
@@ -339,7 +408,7 @@ steps:
   - vllm/model_executor/layers/quantization
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-small.txt -t 1
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
 - label: OpenAI API correctness
   source_file_dependencies:
@@ -358,96 +427,101 @@ steps:
 
 - label: OpenAI-Compatible Tool Use # 20 min
   fast_check: false
-  mirror_hardwares: [ amd ]
+  #mirror_hardwares: [ amd ]
   source_file_dependencies:
     - vllm/
     - tests/tool_use
+    - tests/mistral_tool_use
   commands:
     - pytest -v -s tool_use
+    - pytest -v -s mistral_tool_use
 
 #####  models test  #####
 
 - label: Basic Models Test # 24min
+  torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/models
   commands:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
-    - pytest -v -s models/test_initialization.py
+    - pytest -v -s models/test_utils.py
+    - pytest -v -s models/test_vision.py
+    # V1 Test: https://github.com/vllm-project/vllm/issues/14531
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'plamo2'
 
-- label: Language Models Test (Standard) # 32min
+- label: Language Models Test (Standard)
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/language
-  - tests/models/embedding/language
-  - tests/models/encoder_decoder/language
+  - tests/models/language
   commands:
-    - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
-    - pytest -v -s models/embedding/language -m core_model
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language -m core_model
 
-- label: Language Models Test (Extended) # 1h10min
+- label: Language Models Test (Extended)
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/language
-  - tests/models/embedding/language
-  - tests/models/encoder_decoder/language
+  - tests/models/language
   commands:
-    - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
-    - pytest -v -s models/embedding/language -m 'not core_model'
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language -m 'not core_model'
 
-- label: Multi-Modal Models Test (Standard) # 40min
+- label: Multi-Modal Models Test (Standard)
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/audio_language
-  - tests/models/decoder_only/vision_language
-  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/audio_language
-  - tests/models/encoder_decoder/vision_language
+  - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal
-    - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
-    - pytest -v -s models/embedding/vision_language -m core_model
-    - pytest -v -s models/encoder_decoder/audio_language -m core_model
-    - pytest -v -s models/encoder_decoder/language -m core_model
-    - pytest -v -s models/encoder_decoder/vision_language -m core_model
+    - pytest -v -s models/multimodal/processing
+    - pytest -v -s --ignore models/multimodal/generation/test_whisper.py models/multimodal -m core_model
+    - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
-- label: Multi-Modal Models Test (Extended) 1 # 48m
+- label: Multi-Modal Models Test (Extended) 1
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/audio_language
-  - tests/models/decoder_only/vision_language
-  - tests/models/embedding/vision_language
-  - tests/models/encoder_decoder/vision_language
+  - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
-    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
-    # HACK - run phi3v tests separately to sidestep this transformers bug
-    # https://github.com/huggingface/transformers/issues/34307
-    - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
-    - pytest -v -s models/embedding/vision_language -m 'not core_model'
-    - pytest -v -s models/encoder_decoder/language -m 'not core_model'
-    - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
+    - pytest -v -s --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing models/multimodal -m 'not core_model'
 
-- label: Multi-Modal Models Test (Extended) 2 # 38m
+- label: Multi-Modal Models Test (Extended) 2
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/decoder_only/vision_language
+  - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=1) and not core_model and not quant_model'
+    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
+
+- label: Multi-Modal Models Test (Extended) 3
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
+
+- label: Quantized Models Test
+  #mirror_hardwares: [amd]
+  source_file_dependencies:
+  - vllm/model_executor/layers/quantization
+  - tests/models/quantization
+  commands:
+    - pytest -v -s models/quantization
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
+  mirror_hardwares: [amd]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -459,6 +533,7 @@ steps:
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
+  mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -501,29 +576,31 @@ steps:
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - vllm/v1/engine/
   commands:
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - VLLM_USE_V1=1 torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
-  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/language -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)'
+  # test sequence parallel
+  - pytest -v -s distributed/test_sequence_parallel.py
   # this test fails consistently.
   # TODO: investigate and fix
   # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
-  fast_check: true
   source_file_dependencies:
   - vllm/plugins/
   - tests/plugins/
@@ -582,14 +659,10 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    # This test runs llama 13B, so it is required to run on 4 GPUs.
-    - pytest -v -s -x lora/test_long_context.py
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
-    - pytest -v -s -x lora/test_minicpmv_tp.py
-    - pytest -v -s -x lora/test_transfomers_model.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min
@@ -640,4 +713,4 @@ steps:
   - vllm/model_executor/layers/quantization
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - bash ./run-tests.sh -c configs/models-large.txt -t 4
+  - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4

.github/CODEOWNERS

@@ -10,27 +10,33 @@
 /vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth
-/vllm/model_executor/guided_decoding @mgoin
+/vllm/model_executor/guided_decoding @mgoin @russellb
 /vllm/multimodal @DarkLight1337 @ywang96
+/vllm/vllm_flash_attn @LucasWilkinson
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
+/vllm/v1/structured_output @mgoin @russellb
 
 # Test ownership
-/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
-/tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
-/tests/models @DarkLight1337 @ywang96
-/tests/multimodal @DarkLight1337 @ywang96
-/tests/prefix_caching @comaniac @KuntaiDu
-/tests/spec_decode @njhill @LiuXiaoxuanPKU
-/tests/kernels @tlrmchlsmth @WoosukKwon
-/tests/quantization @mgoin @robertgshaw2-redhat
 /.buildkite/lm-eval-harness @mgoin @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
+/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
+/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb
+/tests/kernels @tlrmchlsmth @WoosukKwon
+/tests/model_executor/test_guided_processors.py @mgoin @russellb
+/tests/models @DarkLight1337 @ywang96
 /tests/multi_step @alexm-redhat @comaniac
+/tests/multimodal @DarkLight1337 @ywang96
+/tests/prefix_caching @comaniac @KuntaiDu
+/tests/quantization @mgoin @robertgshaw2-redhat
+/tests/spec_decode @njhill @LiuXiaoxuanPKU
+/tests/test_inputs.py @DarkLight1337 @ywang96
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb
+/tests/v1/structured_output @mgoin @russellb
 /tests/weight_loading @mgoin @youkaichao
-/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -14,19 +14,19 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```
       It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
     value: |
       <details>
-      <summary>The output of `python collect_env.py`</summary>
+      <summary>The output of <code>python collect_env.py</code></summary>
 
       ```text
       Your output of `python collect_env.py` here
       ```
-      
+
       </details>
   validations:
     required: true
@@ -75,7 +75,7 @@ body:
       ```
 
       ```
-      The error message you got, with the full traceback.
+      The error message you got, with the full traceback and the error logs with [dump_input.py:##] if present.
       ```
   validations:
     required: true

.github/ISSUE_TEMPLATE/600-new-model.yml

@@ -9,7 +9,7 @@ body:
     value: >
       #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
 
-      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/adding_model.html first to understand how to add a new model.
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/index.html first to understand how to add a new model.
 - type: textarea
   attributes:
     label: The model to consider.

.github/ISSUE_TEMPLATE/700-performance-discussion.yml

@@ -35,7 +35,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/800-misc-discussion.yml

@@ -1,28 +0,0 @@
-name: 🎲 Misc/random discussions that do not fit into the above categories.
-description: Submit a discussion as you like. Note that developers are heavily overloaded and we mainly rely on community users to answer these issues.
-title: "[Misc]: "
-labels: ["misc"]
-
-body:
-- type: markdown
-  attributes:
-    value: >
-      #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
-- type: textarea
-  attributes:
-    label: Anything you want to discuss about vllm.
-    description: >
-      Anything you want to discuss about vllm.
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
-- type: checkboxes
-  id: askllm
-  attributes:
-    label: Before submitting a new issue...
-    options:
-      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
-        required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1 +1,5 @@
 blank_issues_enabled: false
+contact_links:
+  - name: Questions
+    url: https://discuss.vllm.ai
+    about: Ask questions and discuss with other vLLM community members

.github/PULL_REQUEST_TEMPLATE.md

@@ -3,4 +3,4 @@ FILL IN THE PR DESCRIPTION HERE
 FIX #xxxx (*link existing issues this PR will resolve*)
 
 <!--- pyml disable-next-line no-emphasis-as-heading -->
-**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>**
+**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>** (anything written below this line will be removed by GitHub Actions)

.github/mergify.yml

@@ -19,7 +19,7 @@ pull_request_rules:
       - files~=\.buildkite/
       - files~=^cmake/
       - files=CMakeLists.txt
-      - files~=^Dockerfile
+      - files~=^docker/Dockerfile
       - files~=^requirements.*\.txt
       - files=setup.py
   actions:
@@ -36,15 +36,38 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-multi-modality
+  description: Automatically apply multi-modality label
+  conditions:
+    - or:
+      - files~=^vllm/multimodal/
+      - files~=^tests/multimodal/
+      - files~=^tests/models/multimodal/
+      - files~=^tests/models/*/audio_language/
+      - files~=^tests/models/*/vision_language/
+      - files=tests/models/test_vision.py
+  actions:
+    label:
+      add:
+        - multi-modality
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
     - or:
+      - files~=^benchmarks/structured_schemas/
+      - files=benchmarks/benchmark_serving_structured_output.py
+      - files=benchmarks/run_structured_output_benchmark.sh
+      - files=docs/source/features/structured_outputs.md
+      - files=examples/offline_inference/structured_outputs.py
+      - files=examples/online_serving/openai_chat_completion_structured_outputs.py
+      - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
       - files~=^vllm/model_executor/guided_decoding/
       - files=tests/model_executor/test_guided_processors.py
       - files=tests/entrypoints/llm/test_guided_generate.py
-      - files=benchmarks/benchmark_serving_guided.py
-      - files=benchmarks/benchmark_guided.py
+      - files~=^tests/v1/structured_output/
+      - files=tests/v1/entrypoints/llm/test_guided_generate.py
+      - files~=^vllm/v1/structured_output/
   actions:
     label:
       add:
@@ -73,6 +96,58 @@ pull_request_rules:
       add:
         - v1
 
+- name: label-tpu
+  description: Automatically apply tpu label
+  # Keep this list in sync with `label-tpu-remove` conditions
+  conditions:
+    - or:
+      - files~=tpu.py
+      - files~=_tpu
+      - files~=tpu_
+      - files~=/tpu/
+      - files~=pallas
+  actions:
+    label:
+      add:
+        - tpu
+
+- name: label-tpu-remove
+  description: Automatically remove tpu label
+  # Keep this list in sync with `label-tpu` conditions
+  conditions:
+    - and:
+      - -files~=tpu.py
+      - -files~=_tpu
+      - -files~=tpu_
+      - -files~=/tpu/
+      - -files~=pallas
+  actions:
+    label:
+      remove:
+        - tpu
+
+- name: label-tool-calling
+  description: Automatically add tool-calling label
+  conditions:
+    - or:
+      - files~=^tests/tool_use/
+      - files~=^tests/mistral_tool_use/
+      - files~=^tests/entrypoints/openai/tool_parsers/
+      - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/
+      - files=docs/source/features/tool_calling.md
+      - files=docs/source/getting_started/examples/openai_chat_completion_client_with_tools.md
+      - files=docs/source/getting_started/examples/chat_with_tools.md
+      - files~=^examples/tool_chat_*
+      - files=examples/offline_inference/chat_with_tools.py
+      - files=examples/online_serving/openai_chat_completion_client_with_tools_required.py
+      - files=examples/online_serving/openai_chat_completion_tool_calls_with_reasoning.py
+      - files=examples/online_serving/openai_chat_completion_client_with_tools.py
+  actions:
+    label:
+      add:
+        - tool-calling
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
       - conflict

.github/workflows/lint-and-deploy.yaml

@@ -50,7 +50,7 @@ jobs:
         uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
 
       - name: Build the Docker image vllm cpu
-        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
+        run: docker buildx build -f docker/Dockerfile.cpu -t vllm-cpu-env .
 
       - name: Configuration of docker images, network and namespace for the kind cluster
         run: |
@@ -66,7 +66,7 @@ jobs:
           export AWS_SECRET_ACCESS_KEY=minioadmin
           sleep 30 && kubectl -n ns-vllm logs -f "$(kubectl -n ns-vllm get pods | awk '/deployment/ {print $1;exit}')" &
           helm install --wait --wait-for-jobs --timeout 5m0s --debug --create-namespace --namespace=ns-vllm test-vllm examples/online_serving/chart-helm -f examples/online_serving/chart-helm/values.yaml --set secrets.s3endpoint=http://minio:9000 --set secrets.s3bucketname=testbucket --set secrets.s3accesskeyid=$AWS_ACCESS_KEY_ID --set secrets.s3accesskey=$AWS_SECRET_ACCESS_KEY --set resources.requests.cpu=1 --set resources.requests.memory=4Gi --set resources.limits.cpu=2 --set resources.limits.memory=5Gi --set image.env[0].name=VLLM_CPU_KVCACHE_SPACE --set image.env[1].name=VLLM_LOGGING_LEVEL --set-string image.env[0].value="1" --set-string image.env[1].value="DEBUG" --set-string extraInit.s3modelpath="opt-125m/" --set-string 'resources.limits.nvidia\.com/gpu=0' --set-string 'resources.requests.nvidia\.com/gpu=0' --set-string image.repository="vllm-cpu-env"
-    
+
       - name: curl test
         run: |
           kubectl -n ns-vllm port-forward service/test-vllm-service 8001:80 &
@@ -79,4 +79,4 @@ jobs:
                           "max_tokens": 7,
                           "temperature": 0
                   }'):$CODE"
-          echo "$CODE"
+          echo "$CODE"

.github/workflows/publish.yml

@@ -39,7 +39,7 @@ jobs:
             const script = require('.github/workflows/scripts/create_release.js')
             await script(github, context, core)
 
-  # NOTE(simon): No longer build wheel using Github Actions. See buildkite's release workflow. 
+  # NOTE(simon): No longer build wheel using GitHub Actions. See buildkite's release workflow. 
   # wheel:
   #   name: Build Wheel
   #   runs-on: ${{ matrix.os }}
@@ -50,7 +50,7 @@ jobs:
   #     matrix:
   #         os: ['ubuntu-20.04']
   #         python-version: ['3.9', '3.10', '3.11', '3.12']
-  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
+  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements/cuda.txt.
   #         cuda-version: ['11.8', '12.1']
 
   #   steps:

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-$python_executable -m pip install -r requirements-build.txt -r requirements-cuda.txt
+$python_executable -m pip install -r requirements/build.txt -r requirements/cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1

.github/workflows/scripts/create_release.js

@@ -1,4 +1,4 @@
-// Uses Github's API to create the release and wait for result.
+// Uses GitHub's API to create the release and wait for result.
 // We use a JS script since github CLI doesn't provide a way to wait for the release's creation and returns immediately.
 
 module.exports = async (github, context, core) => {

.gitignore

@@ -2,7 +2,7 @@
 /vllm/_version.py
 
 # vllm-flash-attn built from source
-vllm/vllm_flash_attn/
+vllm/vllm_flash_attn/*
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -80,6 +80,7 @@ instance/
 # Sphinx documentation
 docs/_build/
 docs/source/getting_started/examples/
+docs/source/api/vllm
 
 # PyBuilder
 .pybuilder/
@@ -197,8 +198,11 @@ _build/
 hip_compat.h
 
 # Benchmark dataset
-benchmarks/*.json
+benchmarks/**/*.json
 
 # Linting
 actionlint
 shellcheck*/
+
+# Ingore moe/marlin_moe gen code
+csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -1,3 +1,6 @@
+default_install_hook_types:
+  - pre-commit
+  - commit-msg
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
@@ -8,31 +11,30 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
-    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
 - repo: https://github.com/astral-sh/ruff-pre-commit
-  rev: v0.9.3
+  rev: v0.11.7
   hooks:
   - id: ruff
     args: [--output-format, github, --fix]
 - repo: https://github.com/codespell-project/codespell
-  rev: v2.4.0
+  rev: v2.4.1
   hooks:
   - id: codespell
     additional_dependencies: ['tomli']
     args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
-  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
+  rev: 6.0.1
   hooks:
   - id: isort
 - repo: https://github.com/pre-commit/mirrors-clang-format
-  rev: v19.1.7
+  rev: v20.1.3
   hooks:
   - id: clang-format
     exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/jackdewinter/pymarkdown
-  rev: v0.9.27
+  rev: v0.9.29
   hooks:
   - id: pymarkdown
     args: [fix]
@@ -41,11 +43,11 @@ repos:
   hooks:
   - id: actionlint
 - repo: https://github.com/astral-sh/uv-pre-commit
-  rev: 0.6.2
+  rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements-test.in, -o, requirements-test.txt]
-      files: ^requirements-test\.(in|txt)$
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
+      files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
   - id: mypy-local
@@ -53,7 +55,7 @@ repos:
     entry: tools/mypy.sh 0 "local"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests]
     stages: [pre-commit] # Don't run in CI
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
@@ -99,8 +101,8 @@ repos:
     args:
       - -c
       - |
-        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" .git/COMMIT_EDITMSG; then
-          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> .git/COMMIT_EDITMSG
+        if ! grep -q "^Signed-off-by: $(git config user.name) <$(git config user.email)>" "$(git rev-parse --git-path COMMIT_EDITMSG)"; then
+          printf "\nSigned-off-by: $(git config user.name) <$(git config user.email)>\n" >> "$(git rev-parse --git-path COMMIT_EDITMSG)"
         fi
     language: system
     verbose: true
@@ -119,6 +121,10 @@ repos:
     language: system
     always_run: true
     pass_filenames: false
+  - id: update-dockerfile-graph
+    name: Update Dockerfile dependency graph
+    entry: tools/update-dockerfile-graph.sh
+    language: script
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -18,4 +18,4 @@ formats: []
 # Optionally declare the Python requirements required to build your docs
 python:
   install:
-    - requirements: docs/requirements-docs.txt
+    - requirements: requirements/docs.txt

CMakeLists.txt

@@ -15,7 +15,6 @@ project(vllm_extensions LANGUAGES CXX)
 
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
-
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
 
@@ -34,7 +33,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -44,10 +43,10 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
 #
 # Note: the CUDA torch version is derived from pyproject.toml and various
 # requirements.txt files and should be kept consistent.  The ROCm torch
-# versions are derived from Dockerfile.rocm
+# versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.5.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.5.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.7.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -230,25 +229,28 @@ set(VLLM_EXT_SRC
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
+  "csrc/attention/merge_attn_states.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
   "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/gguf/gguf_kernel.cu"
+  "csrc/quantization/activation_kernels.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
+  "csrc/custom_all_reduce.cu"
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
-  # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
-  # Please keep this in sync with FetchContent_Declare line below.
-  set(CUTLASS_REVISION "v3.8.0" CACHE STRING "CUTLASS revision to use")
+  # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
+  set(CUTLASS_REVISION "v3.9.2" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -266,7 +268,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG v3.8.0
+        GIT_TAG ${CUTLASS_REVISION}
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -282,13 +284,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
     "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
-    "csrc/cutlass_extensions/common.cpp")
+    "csrc/cutlass_extensions/common.cpp"
+    "csrc/attention/mla/cutlass_mla_entry.cu")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -299,8 +301,52 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # are not supported by Machete yet.
   cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
+
+    #
+    # For the Marlin kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
+    file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=$PYTHONPATH
+          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE marlin_generation_result
+        OUTPUT_VARIABLE marlin_generation_result
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
+      )
+
+      if (NOT marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin generation failed."
+                            " Result: \"${marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
+      else()
+        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin generate script hash" FORCE)
+        message(STATUS "Marlin generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
+      CUDA_ARCHS "${MARLIN_ARCHS}")
+
+    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
+
     set(MARLIN_SRCS
-       "csrc/quantization/fp8/fp8_marlin.cu"
        "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
        "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
        "csrc/quantization/marlin/qqq/marlin_qqq_gemm_kernel.cu"
@@ -319,7 +365,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # Only build AllSpark kernels if we are building for at least some compatible archs.
   cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
-  if (${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND ALLSPARK_ARCHS)
+  if (ALLSPARK_ARCHS)
     set(ALLSPARK_SRCS
        "csrc/quantization/gptq_allspark/allspark_repack.cu"
        "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
@@ -330,39 +376,68 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
   else()
     message(STATUS "Not building AllSpark kernels as no compatible archs found"
-                   " in CUDA target architectures, or CUDA not >= 12.0")
+                   " in CUDA target architectures")
   endif()
 
+
+  set(SCALED_MM_3X_ARCHS)
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a;10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
+  # CUDA 12.0 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C3X=1")
-    message(STATUS "Building scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x as CUDA Compiler version is "
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
                      "Hopper.")
     else()
-      message(STATUS "Not building scaled_mm_c3x as no compatible archs found "
+      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
+  endif()
 
-    # clear SCALED_MM_3X_ARCHS so the scaled_mm_c2x kernels know we didn't
-    # build any 3x kernels
-    set(SCALED_MM_3X_ARCHS)
+  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
+  # CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm100 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm100 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
   endif()
 
   #
@@ -394,17 +469,18 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 2:4 Sparse Kernels
 
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
-  # require CUDA 12.2 or later (and only work on Hopper and Blackwell).
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+  # require CUDA 12.2 or later (and only work on Hopper).
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
-    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -432,22 +508,52 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(FP4_ARCHS)
   endif()
 
-  # FP8 Blackwell Archs 
-  cuda_archs_loose_intersection(BLACKWELL_ARCHS "10.0;10.1;12.0" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND BLACKWELL_ARCHS)
-    set(SRCS 
-      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
-    )
+  # CUTLASS MLA Archs and flags
+  cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND MLA_ARCHS)
+    set(SRCS
+      "csrc/attention/mla/cutlass_mla_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${BLACKWELL_ARCHS}")
+      CUDA_ARCHS "${MLA_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    message(STATUS "Building FP8 for archs: ${BLACKWELL_ARCHS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MLA=1")
+    # Add MLA-specific include directories only to MLA source files
+    set_source_files_properties(${SRCS}
+      PROPERTIES INCLUDE_DIRECTORIES "${CUTLASS_DIR}/examples/77_blackwell_fmha;${CUTLASS_DIR}/examples/common")
+    message(STATUS "Building CUTLASS MLA for archs: ${MLA_ARCHS}")
   else()
-    # clear BLACKWELL_ARCHS
-    set(BLACKWELL_ARCHS)
+    message(STATUS "Not building CUTLASS MLA as no compatible archs were found.")
+    # clear MLA_ARCHS
+    set(MLA_ARCHS)
   endif()
-  
+
+  # CUTLASS MoE kernels
+
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
+  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
+  # to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
+             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
+    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building grouped_mm_c3x kernels as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper.")
+    else()
+      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
   #
   # Machete kernels
 
@@ -548,28 +654,70 @@ set(VLLM_MOE_EXT_SRC
   "csrc/moe/moe_align_sum_kernels.cu"
   "csrc/moe/topk_softmax_kernels.cu")
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
+endif()
+
 set_gencode_flags_for_srcs(
   SRCS "${VLLM_MOE_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(VLLM_MOE_WNA16_SRC
+    "csrc/moe/moe_wna16.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${VLLM_MOE_WNA16_SRC}"
+    CUDA_ARCHS "${CUDA_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
-    set(MARLIN_MOE_SRC
-        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu"
-        "csrc/moe/marlin_moe_ops.cu")
 
+    #
+    # For the Marlin MOE kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MOE_MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
+    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=$PYTHONPATH
+          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE moe_marlin_generation_result
+        OUTPUT_VARIABLE moe_marlin_generation_output
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+      )
+
+      if (NOT moe_marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin MOE generation failed."
+                            " Result: \"${moe_marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
+      else()
+        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
+        message(STATUS "Marlin MOE generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
+      SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
 
-    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_SRC}")
+    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
+
     message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
   else()
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
@@ -577,6 +725,17 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 endif()
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  set(MOE_PERMUTE_SRC
+      "csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.cu"
+      "csrc/moe/moe_permute_unpermute_op.cu")
+
+  set_gencode_flags_for_srcs(
+    SRCS "${MARLIN_PERMUTE_SRC}"
+    CUDA_ARCHS "${MOE_PERMUTE_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${MOE_PERMUTE_SRC}")
+endif()
 message(STATUS "Enabling moe extension.")
 define_gpu_extension_target(
   _moe_C
@@ -585,6 +744,8 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -594,6 +755,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   #
   set(VLLM_ROCM_EXT_SRC
     "csrc/rocm/torch_bindings.cpp"
+    "csrc/rocm/skinny_gemms.cu"
     "csrc/rocm/attention.cu")
 
   define_gpu_extension_target(

Dockerfile.cpu

@@ -1,69 +0,0 @@
-# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
-
-FROM ubuntu:22.04 AS cpu-test-1
-
-ENV CCACHE_DIR=/root/.cache/ccache
-
-ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
-
-RUN --mount=type=cache,target=/var/cache/apt \
-    apt-get update -y \
-    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
-    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
-
-# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
-# intel-openmp provides additional performance improvement vs. openmp
-# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install intel-openmp==2025.0.1
-
-ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
-
-RUN echo 'ulimit -c 0' >> ~/.bashrc
-
-RUN pip install intel_extension_for_pytorch==2.5.0
-
-WORKDIR /workspace
-
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
-ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
-    pip install --upgrade pip && \
-    pip install -r requirements-build.txt
-
-FROM cpu-test-1 AS build
-
-WORKDIR /workspace/vllm
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
-    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
-    pip install -v -r requirements-cpu.txt
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
-ARG VLLM_CPU_DISABLE_AVX512
-ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
-    pip install dist/*.whl && \
-    rm -rf dist
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -e tests/vllm_test_utils
-
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.openvino

@@ -1,29 +0,0 @@
-# The vLLM Dockerfile is used to construct vLLM image that can be directly used
-# to run the OpenAI compatible server.
-
-FROM ubuntu:22.04 AS dev
-
-RUN apt-get update -y && \
-    apt-get install -y \
-        git python3-pip \
-        ffmpeg libsm6 libxext6 libgl1
-WORKDIR /workspace
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-RUN python3 -m pip install -U pip
-# install build requirements
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements-build.txt
-# build vLLM with OpenVINO backend
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace
-
-COPY examples/ /workspace/examples
-COPY benchmarks/ /workspace/benchmarks
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-CMD ["/bin/bash"]

Dockerfile.ppc64le

@@ -1,37 +0,0 @@
-FROM mambaorg/micromamba
-ARG MAMBA_DOCKERFILE_ACTIVATE=1
-USER root
-
-ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
-
-RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 
-
-# Some packages in requirements-cpu are installed here
-# IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
-# Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes
-
-COPY ./ /workspace/vllm
-
-WORKDIR /workspace/vllm
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-RUN --mount=type=cache,target=/root/.cache/pip  \
-    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
-        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        -r requirements-cpu.txt \
-        xformers uvloop==0.20.0
-
-RUN --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py install
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.xpu

@@ -1,69 +0,0 @@
-FROM intel/oneapi-basekit:2024.2.1-0-devel-ubuntu22.04 AS vllm-base
-
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
-    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
-    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
-    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
-    chmod 644 /usr/share/keyrings/intel-graphics.gpg
-
-RUN apt-get update -y && \
-    apt-get install -y --no-install-recommends --fix-missing \
-    curl \
-    ffmpeg \
-    git \
-    libsndfile1 \
-    libsm6 \
-    libxext6 \
-    libgl1 \
-    lsb-release \
-    numactl \
-    python3 \
-    python3-dev \
-    python3-pip \
-    # vim \
-    wget
-
-WORKDIR /workspace/vllm
-COPY requirements-xpu.txt /workspace/vllm/requirements-xpu.txt
-COPY requirements-common.txt /workspace/vllm/requirements-common.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install --no-cache-dir \
-    -r requirements-xpu.txt
-
-RUN git clone https://github.com/intel/pti-gpu && \
-    cd pti-gpu/sdk && \
-    git checkout 6c491f07a777ed872c2654ca9942f1d0dde0a082 && \
-    mkdir build && \
-    cd build && \
-    cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/icpx_toolchain.cmake -DBUILD_TESTING=OFF .. && \
-    make -j && \
-    cmake --install . --config Release --prefix "/usr/local"
-
-ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
-
-COPY . .
-ARG GIT_REPO_CHECK
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-ENV VLLM_TARGET_DEVICE=xpu
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,source=.git,target=.git \
-    python3 setup.py install
-
-CMD ["/bin/bash"]
-
-FROM vllm-base AS vllm-openai
-
-# install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer 'modelscope!=1.15.0'
-
-ENV VLLM_USAGE_SOURCE production-docker-image \
-    TRITON_XPU_PROFILE 1
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

MANIFEST.in

@@ -1,9 +1,9 @@
 include LICENSE
-include requirements-common.txt
-include requirements-cuda.txt
-include requirements-rocm.txt
-include requirements-neuron.txt
-include requirements-cpu.txt
+include requirements/common.txt
+include requirements/cuda.txt
+include requirements/rocm.txt
+include requirements/neuron.txt
+include requirements/cpu.txt
 include CMakeLists.txt
 
 recursive-include cmake *

README.md

@@ -10,22 +10,24 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
 
-We’re excited to invite you to the first **vLLM China Meetup** on **March 16** in **Beijing**!  
-
-Join us to connect with the **vLLM team** and explore how vLLM is leveraged in **post-training, fine-tuning, and deployment**, including [verl](https://github.com/volcengine/verl), [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory), and [vllm-ascend](https://github.com/vllm-project/vllm-ascend).
-
-👉 **[Register Now](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)** to be part of the discussion!  
-
----
-
 *Latest News* 🔥
-
+- [2025/05] We hosted [NYC vLLM Meetup](https://lu.ma/c1rqyf1f)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1_q_aW_ioMJWUImf1s1YM-ZhjXz8cUeL0IJvaquOYBeA/edit?usp=sharing).
+- [2025/05] vLLM is now a hosted project under PyTorch Foundation! Please find the announcement [here](https://pytorch.org/blog/pytorch-foundation-welcomes-vllm/).
+- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
+
+<details>
+<summary>Previous News</summary>
+
+- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
+- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
+- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
+- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
 - [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
@@ -41,8 +43,9 @@ Join us to connect with the **vLLM team** and explore how vLLM is leveraged in *
 - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
 
----
+</details>
 
+---
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.
@@ -90,14 +93,14 @@ pip install vllm
 ```
 
 Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
-- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation.html)
 - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
 - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
 
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
 
 ## Sponsors
 
@@ -120,6 +123,7 @@ Compute Resources:
 - Databricks
 - DeepInfra
 - Google Cloud
+- Intel
 - Lambda Lab
 - Nebius
 - Novita AI
@@ -150,10 +154,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-- For technical questions and feature requests, please use Github issues or discussions.
-- For discussing with fellow users and coordinating contributions and development, please use Slack.
-- For security disclosures, please use Github's security advisory feature.
-- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
+- For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
+- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
+- For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
 
 ## Media Kit
 

benchmarks/README.md

@@ -1,29 +1,343 @@
 # Benchmarking vLLM
 
-## Downloading the ShareGPT dataset
+This README guides you through running benchmark tests with the extensive
+datasets supported on vLLM. It’s a living document, updated as new features and datasets
+become available.
 
-You can download the dataset by running:
+## Dataset Overview
+
+<table style="width:100%; border-collapse: collapse;">
+  <thead>
+    <tr>
+      <th style="width:15%; text-align: left;">Dataset</th>
+      <th style="width:10%; text-align: center;">Online</th>
+      <th style="width:10%; text-align: center;">Offline</th>
+      <th style="width:65%; text-align: left;">Data Path</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td><strong>ShareGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json</code></td>
+    </tr>
+    <tr>
+      <td><strong>BurstGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
+    </tr>
+    <tr>
+      <td><strong>Sonnet</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>benchmarks/sonnet.txt</code></td>
+    </tr>
+    <tr>
+      <td><strong>Random</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
+    </tr>
+  </tbody>
+</table>
+
+✅: supported
+
+🟡: Partial support
+
+🚧: to be supported
+
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
+
+---
+## Example - Online Benchmark
+
+First start serving your model
 
 ```bash
-wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
 ```
 
-## Downloading the ShareGPT4V dataset
-
-The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts
-will ignore a datapoint if the referred image is missing.
+Then run the benchmarking script
 
 ```bash
-wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json
-mkdir coco -p
-wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip
-unzip coco/train2017.zip -d coco/
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
 ```
 
-# Downloading the BurstGPT dataset
+If successful, you will see the following output
 
-You can download the BurstGPT v1.1 dataset by running:
+```
+============ Serving Benchmark Result ============
+Successful requests:                     10        
+Benchmark duration (s):                  5.78      
+Total input tokens:                      1369      
+Total generated tokens:                  2212      
+Request throughput (req/s):              1.73      
+Output token throughput (tok/s):         382.89    
+Total Token throughput (tok/s):          619.85    
+---------------Time to First Token----------------
+Mean TTFT (ms):                          71.54     
+Median TTFT (ms):                        73.88     
+P99 TTFT (ms):                           79.49     
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          7.91      
+Median TPOT (ms):                        7.96      
+P99 TPOT (ms):                           8.03      
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           7.74      
+Median ITL (ms):                         7.70      
+P99 ITL (ms):                            8.39      
+==================================================
+```
+
+### VisionArena Benchmark for Vision Language Models
 
 ```bash
-wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
+# need a model with vision capability here
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-model "[ngram]" \
+    --ngram_prompt_lookup_min 2 \
+    --ngram-prompt-lookup-max 5 \
+    --num_speculative_tokens 5
+```
+
+``` bash
+python3 benchmarks/benchmark_serving.py \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
+
+```bash
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+```
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
+```
+
+### Running With Sampling Parameters
+
+When using OpenAI-compatible backends such as `vllm`, optional sampling
+parameters can be specified. Example client command:
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
+```
+
+---
+## Example - Offline Throughput Benchmark
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```
+Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
+Total num prompt tokens:  5014
+Total num output tokens:  1500
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+``` bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
+```
+
+The `num prompt tokens` now includes image token counts
+
+```
+Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
+Total num prompt tokens:  14527
+Total num output tokens:  1280
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+python3 vllm/benchmarks/benchmark_throughput.py \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-model="[ngram]" \
+    --ngram_prompt_lookup_min=2 \
+    --ngram-prompt-lookup-max=5 \
+    --num_speculative_tokens=5
+```
+
+```
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+```bash
+python3 benchmarks/benchmark_throughput.py \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
+### Benchmark with LoRA Adapters
+
+``` bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
+  ```

benchmarks/auto_tune.sh

@@ -0,0 +1,212 @@
+#!/bin/bash
+
+# This script aims to tune the best server parameter combinations to maximize throughput for given requirement. 
+# The current server parameter combination is  max_num_seqs and max_num_batched_tokens
+# It also supports additional requirement: e2e latency and prefix cache. 
+
+# Pre-requisite:
+# 1. Checkout to your branch, install/ update the correct running env. For TPU, activate conda env and install the corresponding torch, xla version. 
+# 2. If the model is customized, replace the MODEL's config with the customized config.
+# 3. Set variables (ALL REQUIRED)
+#   BASE: your directory for vllm repo
+#   MODEL: the model served by vllm
+#   DOWNLOAD_DIR: directory to download and load model weights.
+#   INPUT_LEN: request input len
+#   OUTPUT_LEN: request output len
+#   MIN_CACHE_HIT_PCT: prefix cache rate
+#   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
+# 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
+# 5. The final result will be saved in RESULT file. 
+
+
+# Example use cases 
+# 1. Given input_len=1800, output_len=20, what's the best max_num_seqs and max_num_batched_tokens to get highest throughput?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=100000000000
+# 2. If we have latency requirement to be lower than 500ms, what's the best server parameter?
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=0, MAX_LATENCY_ALLOWED_MS=500
+# 3. If we want to reach 60% prefix cache, what's the best server parameter? 
+# Use INPUT_LEN=1800,  OUTPUT_LEN=20, MIN_CACHE_HIT_PCT=60, MAX_LATENCY_ALLOWED_MS=500
+
+TAG=$(date +"%Y_%m_%d_%H_%M")
+BASE=""
+MODEL="meta-llama/Llama-3.1-8B-Instruct"
+DOWNLOAD_DIR=""
+INPUT_LEN=4000
+OUTPUT_LEN=16
+MIN_CACHE_HIT_PCT_PCT=0
+MAX_LATENCY_ALLOWED_MS=100000000000
+
+LOG_FOLDER="$BASE/auto-benchmark/$TAG"
+RESULT="$LOG_FOLDER/result.txt"
+
+echo "result file$ $RESULT"
+echo "model: $MODEL"
+echo
+
+rm -rf $LOG_FOLDER
+mkdir -p $LOG_FOLDER
+
+cd "$BASE/vllm"
+# create sonnet-4x.txt so that we can sample 2048 tokens for input
+echo "" > benchmarks/sonnet_4x.txt
+for _ in {1..4}
+do
+cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
+done
+
+pip install datasets
+
+current_hash=$(git rev-parse HEAD)
+echo "hash:$current_hash" >> "$RESULT"
+echo "current_hash: $current_hash"
+
+best_throughput=0
+best_max_num_seqs=0
+best_num_batched_tokens=0
+best_goodput=0
+run_benchmark() {
+    local max_num_seqs=$1
+    local max_num_batched_tokens=$2
+    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
+    echo "vllm_log: $vllm_log"
+    echo
+    rm -f $vllm_log
+
+    # start the server
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 vllm serve $MODEL \
+        --disable-log-requests \
+        --port 8004 \
+        --gpu-memory-utilization 0.98 \
+        --max-num-seqs $max_num_seqs \
+        --max-num-batched-tokens $max_num_batched_tokens \
+        --tensor-parallel-size 1 \
+        --enable-prefix-caching \
+        --load-format dummy \
+        --download-dir $DOWNLOAD_DIR \
+        --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
+    echo "wait for 10 minutes.."
+    echo
+    # wait for 10 minutes...
+    server_started=0
+    for i in {1..60}; do        
+        if grep -Fq "Application startup complete" "$vllm_log"; then
+            echo "Application started"
+            server_started=1
+            break
+        else
+            # echo "wait for 10 seconds..."
+            sleep 10
+        fi
+    done
+ 
+    if (( ! server_started )); then
+        echo "server did not start within 10 minutes, terminate the benchmarking. Please check server log at $vllm_log"
+        echo "pkill -f vllm"
+        echo
+        pkill vllm
+        sleep 10
+        return 1
+    fi
+    
+    echo "run benchmark test..."
+    echo
+    meet_latency_requirement=0
+    # get a basic qps by using request-rate inf
+    bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
+    prefix_len=$(( INPUT_LEN * MIN_CACHE_HIT_PCT / 100 ))
+    python benchmarks/benchmark_serving.py \
+        --backend vllm \
+        --model $MODEL  \
+        --dataset-name sonnet \
+        --dataset-path benchmarks/sonnet_4x.txt \
+        --sonnet-input-len $INPUT_LEN \
+        --sonnet-output-len $OUTPUT_LEN \
+        --ignore-eos \
+        --disable-tqdm \
+        --request-rate inf \
+        --percentile-metrics ttft,tpot,itl,e2el \
+        --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+        --num-prompts 100 \
+        --sonnet-prefix-len $prefix_len \
+        --port 8004 > "$bm_log"
+    through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+    e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+    goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+
+    if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+        meet_latency_requirement=1
+    fi
+
+    if (( ! meet_latency_requirement )); then
+    # start from request-rate as int(through_put) + 1
+        request_rate=$((${through_put%.*} + 1))
+        while ((request_rate > 0)); do
+            # clear prefix cache
+            curl -X POST http://0.0.0.0:8004/reset_prefix_cache
+            sleep 5
+            bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_${request_rate}.txt"
+            python benchmarks/benchmark_serving.py \
+                --backend vllm \
+                --model $MODEL  \
+                --dataset-name sonnet \
+                --dataset-path benchmarks/sonnet_4x.txt \
+                --sonnet-input-len $INPUT_LEN \
+                --sonnet-output-len $OUTPUT_LEN \
+                --ignore_eos \
+                --disable-tqdm \
+                --request-rate $request_rate \
+                --percentile-metrics ttft,tpot,itl,e2el \
+                --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
+                --num-prompts 100 \
+                --sonnet-prefix-len $prefix_len \
+                --port 8004 > "$bm_log"
+            through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
+            goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+            if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
+                meet_latency_requirement=1
+                break
+            fi
+            request_rate=$((request_rate-1))
+        done
+    fi
+    # write the results and update the best result.
+    if ((meet_latency_requirement)); then
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$through_put > $best_throughput" | bc -l) )); then
+            best_throughput=$through_put
+            best_max_num_seqs=$max_num_seqs
+            best_num_batched_tokens=$max_num_batched_tokens
+            best_goodput=$goodput
+        fi
+    else
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}" >> "$RESULT"
+    fi
+
+    echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+
+    echo "pkill -f vllm"
+    echo
+    pkill vllm
+    sleep 10
+    rm -f $vllm_log
+    printf '=%.0s' $(seq 1 20)
+    return 0
+}
+
+
+num_seqs_list="128 256"
+num_batched_tokens_list="512 1024 2048 4096"
+for num_seqs in $num_seqs_list; do
+    for num_batched_tokens in $num_batched_tokens_list; do
+        run_benchmark $num_seqs $num_batched_tokens
+        exit 0
+    done
+done
+echo "finish permutations"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" >> "$RESULT"
+

benchmarks/backend_request_func.py

@@ -1,5 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 
+import io
 import json
 import os
 import sys
@@ -14,7 +15,8 @@ from tqdm.asyncio import tqdm
 from transformers import (AutoTokenizer, PreTrainedTokenizer,
                           PreTrainedTokenizerFast)
 
-from vllm.model_executor.model_loader.weight_utils import get_lock
+# NOTE(simon): do not import vLLM here so the benchmark script
+# can run without vLLM installed.
 
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
@@ -27,11 +29,11 @@ class RequestFuncInput:
     output_len: int
     model: str
     model_name: Optional[str] = None
-    best_of: int = 1
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
     multi_modal_content: Optional[dict] = None
     ignore_eos: bool = False
+    language: Optional[str] = None
 
 
 @dataclass
@@ -58,13 +60,12 @@ async def async_request_tgi(
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
         params = {
-            "best_of": request_func_input.best_of,
             "max_new_tokens": request_func_input.output_len,
             "do_sample": True,
             "temperature": 0.01,  # TGI does not accept 0.0 temperature.
             "top_p": 0.99,  # TGI does not accept 1.0 top_p.
             "truncate": request_func_input.prompt_len,
-            # TGI does not accept ignore_eos flag.
+            "ignore_eos_token": request_func_input.ignore_eos,
         }
         payload = {
             "inputs": request_func_input.prompt,
@@ -72,6 +73,10 @@ async def async_request_tgi(
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
+        if request_func_input.ignore_eos:
+            output.output_tokens = request_func_input.output_len
+        else:
+            output.output_tokens = None
 
         ttft = 0.0
         st = time.perf_counter()
@@ -130,7 +135,6 @@ async def async_request_trt_llm(
 
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
         payload = {
             "accumulate_tokens": True,
             "text_input": request_func_input.prompt,
@@ -195,9 +199,9 @@ async def async_request_deepspeed_mii(
 ) -> RequestFuncOutput:
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
 
         payload = {
+            "model": request_func_input.model,
             "prompt": request_func_input.prompt,
             "max_tokens": request_func_input.output_len,
             "temperature": 0.01,  # deepspeed-mii does not accept 0.0 temp.
@@ -218,7 +222,15 @@ async def async_request_deepspeed_mii(
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
-                    output.generated_text = parsed_resp["text"][0]
+                    if "choices" in parsed_resp:
+                        output.generated_text = parsed_resp["choices"][0][
+                            "text"]
+                    elif "text" in parsed_resp:
+                        output.generated_text = parsed_resp["text"][0]
+                    else:
+                        output.error = ("Unexpected response format: "
+                                        "neither 'choices' nor 'text' found")
+                        output.success = False
                     output.success = True
                 else:
                     output.error = response.reason or ""
@@ -249,7 +261,7 @@ async def async_request_openai_completions(
                 if request_func_input.model_name else request_func_input.model,
             "prompt": request_func_input.prompt,
             "temperature": 0.0,
-            "best_of": request_func_input.best_of,
+            "repetition_penalty": 1.0,
             "max_tokens": request_func_input.output_len,
             "logprobs": request_func_input.logprobs,
             "stream": True,
@@ -338,7 +350,7 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "chat/completions"
+        ("chat/completions", "profile")
     ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
     async with aiohttp.ClientSession(trust_env=True,
@@ -428,10 +440,116 @@ async def async_request_openai_chat_completions(
     return output
 
 
+async def async_request_openai_audio(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    # Lazy import without PlaceholderModule to avoid vllm dep.
+    import soundfile
+    api_url = request_func_input.api_url
+    assert api_url.endswith(
+        ("transcriptions", "translations"
+         )), "OpenAI Chat Completions API URL must end with 'transcriptions' "
+    "or `translations`."
+
+    async with aiohttp.ClientSession(trust_env=True,
+                                     timeout=AIOHTTP_TIMEOUT) as session:
+        content = [{"type": "text", "text": request_func_input.prompt}]
+        payload = {
+            "model": request_func_input.model_name \
+                if request_func_input.model_name else request_func_input.model,
+            "temperature": 0.0,
+            "max_completion_tokens": request_func_input.output_len,
+            "stream": True,
+            "language": "en",
+            # Flattened due to multipart/form-data
+            "stream_include_usage": True,
+            "stream_continuous_usage_stats": True
+        }
+        if request_func_input.extra_body:
+            payload.update(request_func_input.extra_body)
+        headers = {
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
+        }
+
+        # Send audio file
+        def to_bytes(y, sr):
+            buffer = io.BytesIO()
+            soundfile.write(buffer, y, sr, format="WAV")
+            buffer.seek(0)
+            return buffer
+
+        with to_bytes(*request_func_input.multi_modal_content['audio']) as f:
+            form = aiohttp.FormData()
+            form.add_field('file', f, content_type='audio/wav')
+            for key, value in payload.items():
+                form.add_field(key, str(value))
+
+            output = RequestFuncOutput()
+            output.prompt_len = request_func_input.prompt_len
+
+            generated_text = ""
+            ttft = 0.0
+            st = time.perf_counter()
+            most_recent_timestamp = st
+            try:
+                async with session.post(url=api_url,
+                                        data=form,
+                                        headers=headers) as response:
+                    if response.status == 200:
+                        async for chunk_bytes in response.content:
+                            chunk_bytes = chunk_bytes.strip()
+                            if not chunk_bytes:
+                                continue
+
+                            chunk = chunk_bytes.decode("utf-8").removeprefix(
+                                "data: ")
+                            if chunk != "[DONE]":
+                                timestamp = time.perf_counter()
+                                data = json.loads(chunk)
+
+                                if choices := data.get("choices"):
+                                    content = choices[0]["delta"].get(
+                                        "content")
+                                    # First token
+                                    if ttft == 0.0:
+                                        ttft = timestamp - st
+                                        output.ttft = ttft
+
+                                    # Decoding phase
+                                    else:
+                                        output.itl.append(
+                                            timestamp - most_recent_timestamp)
+
+                                    generated_text += content or ""
+                                elif usage := data.get("usage"):
+                                    output.output_tokens = usage.get(
+                                        "completion_tokens")
+
+                                most_recent_timestamp = timestamp
+
+                        output.generated_text = generated_text
+                        output.success = True
+                        output.latency = most_recent_timestamp - st
+                    else:
+                        output.error = response.reason or ""
+                        output.success = False
+            except Exception:
+                output.success = False
+                exc_info = sys.exc_info()
+                output.error = "".join(traceback.format_exception(*exc_info))
+
+        if pbar:
+            pbar.update(1)
+        return output
+
+
 def get_model(pretrained_model_name_or_path: str) -> str:
     if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
         from modelscope import snapshot_download
 
+        from vllm.model_executor.model_loader.weight_utils import get_lock
+
         # Use file lock to prevent multiple processes from
         # downloading the same model weights at the same time.
         with get_lock(pretrained_model_name_or_path):
@@ -483,7 +601,14 @@ ASYNC_REQUEST_FUNCS = {
     "deepspeed-mii": async_request_deepspeed_mii,
     "openai": async_request_openai_completions,
     "openai-chat": async_request_openai_chat_completions,
+    "openai-audio": async_request_openai_audio,
     "tensorrt-llm": async_request_trt_llm,
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
 }
+
+OPENAI_COMPATIBLE_BACKENDS = [
+    k for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions,
+             async_request_openai_chat_completions)
+]

benchmarks/benchmark_guided.py

@@ -1,507 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Benchmark guided decoding throughput."""
-import argparse
-import dataclasses
-import json
-import os
-import random
-import time
-
-import datasets
-import pandas as pd
-import uvloop
-from transformers import AutoTokenizer, PreTrainedTokenizerBase
-
-from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
-from vllm.entrypoints.openai.api_server import (
-    build_async_engine_client_from_engine_args)
-from vllm.sampling_params import GuidedDecodingParams
-from vllm.utils import FlexibleArgumentParser, merge_async_iterators
-
-
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    Attributes:
-        prompt: The input text prompt for the model.
-        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
-            images).
-        prompt_len: The length of the prompt in tokens.
-        expected_output_len: The expected length of the output in tokens.
-    """
-    prompt: str
-    prompt_len: int
-    expected_output_len: int
-    schema: dict
-    structure_type: str = 'json'
-    completion: str = None
-
-
-def run_vllm(requests: list[SampleRequest],
-             engine_args: EngineArgs,
-             n: int,
-             guided_decoding_rate: float = 1.0,
-             warmup: bool = False) -> float:
-    from vllm import LLM, SamplingParams
-    llm = LLM(**vars(engine_args))
-    assert all(
-        llm.llm_engine.model_config.max_model_len >= (
-            request.prompt_len + request.expected_output_len)
-        for request in requests), (
-            "Please ensure that max_model_len is greater than the sum of"
-            " prompt_len and expected_output_len for all requests.")
-
-    # Add the requests to the engine.
-    prompts: list[str] = []
-    sampling_params: list[SamplingParams] = []
-    # create a list containing random selected true or false
-    guided_decoding_req_idx = random.sample(
-        range(len(requests)), int(len(requests) * guided_decoding_rate))
-
-    if warmup:
-        print(">>>>> Running warmup prompt, for the first 5")
-        # We setup the first 5 requests to warmup FSM
-        # if using xgrammar dataset, we will skip warmup
-        warmup_requests = requests[:5]
-        for i, request in enumerate(warmup_requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if guided_decoding_rate > 0 else None,
-                ))
-        llm.generate(prompts, sampling_params, use_tqdm=False)
-
-    print(">>>>> Benchmark started...")
-    prompts = []
-    sampling_params = []
-    for i, request in enumerate(requests):
-        prompts.append(request.prompt)
-        sampling_params.append(
-            SamplingParams(
-                n=n,
-                temperature=1.0,
-                top_p=1.0,
-                ignore_eos=True,
-                max_tokens=request.expected_output_len,
-                guided_decoding=GuidedDecodingParams(
-                    **{request.structure_type: request.schema})
-                if i in guided_decoding_req_idx else None,
-            ))
-
-    start = time.perf_counter()
-    outputs = llm.generate(prompts, sampling_params, use_tqdm=False)
-    ret = []
-    for output, request in zip(outputs, requests):
-        generated_text = output.outputs[0].text
-        ret.append({
-            "generated": generated_text,
-            "expected": request.completion
-        })
-    end = time.perf_counter()
-    return end - start, ret
-
-
-async def run_vllm_async(
-        requests: list[SampleRequest],
-        engine_args: AsyncEngineArgs,
-        n: int,
-        guided_decoding_rate: float = 1.0,
-        warmup: bool = False,
-        disable_frontend_multiprocessing: bool = False) -> float:
-    from vllm import SamplingParams
-
-    async with build_async_engine_client_from_engine_args(
-            engine_args, disable_frontend_multiprocessing) as llm:
-
-        assert all(
-            llm.model_config.max_model_len >= (request.prompt_len +
-                                               request.expected_output_len)
-            for request in requests), (
-                "Please ensure that max_model_len is greater than the sum of"
-                " prompt_len and expected_output_len for all requests.")
-
-        # Add the requests to the engine.
-        prompts: list[str] = []
-        sampling_params: list[SamplingParams] = []
-        guided_decoding_req_idx = random.sample(
-            range(len(requests)), int(len(requests) * guided_decoding_rate))
-
-        if warmup:
-            print(">>>>>> Running warmup prompt, for the first 5")
-            # We setup the first 5 requests to warmup FSM
-            # if using xgrammar dataset, we will skip warmup
-            warmup_requests = requests[:5]
-            for i, request in enumerate(warmup_requests):
-                prompts.append(request.prompt)
-                sampling_params.append(
-                    SamplingParams(
-                        n=n,
-                        temperature=1.0,
-                        top_p=1.0,
-                        ignore_eos=True,
-                        max_tokens=request.expected_output_len,
-                        guided_decoding=GuidedDecodingParams(
-                            json=request.schema)
-                        if guided_decoding_rate > 0 else None,
-                    ))
-            generators = []
-            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-                generator = llm.generate(prompt, sp, request_id=f"test{i}")
-                generators.append(generator)
-            all_gens = merge_async_iterators(*generators)
-            async for i, res in all_gens:
-                pass
-
-        print(">>>>> Benchmark started...")
-        prompts = []
-        sampling_params = []
-        for i, request in enumerate(requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if i in guided_decoding_req_idx else None,
-                ))
-
-        generators = []
-        start_time = []
-        latencies = []
-        start = time.perf_counter()
-        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-            generator = llm.generate(prompt, sp, request_id=f"test{i}")
-            generators.append(generator)
-            start_time.append(time.perf_counter())
-            latencies.append([])
-        all_gens = merge_async_iterators(*generators)
-        generated_texts = [''] * len(requests)
-        async for i, res in all_gens:
-            generated_texts[i] = res.outputs[0].text
-            lat = time.perf_counter() - start_time[i]
-            latencies[i].append(lat)
-        ret = [{
-            'generated': gt,
-            'expected': req.completion
-        } for gt, req in zip(generated_texts, requests)]
-        end = time.perf_counter()
-        first_latency = pd.Series([lat[0] * 1000 for lat in latencies])
-        next_latency = pd.Series([(lat[-1] - lat[0]) / len(lat[1:]) * 1000
-                                  for lat in latencies])
-        return end - start, ret, (first_latency, next_latency)
-
-
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> list[SampleRequest]:
-    if args.dataset == 'json':
-        if args.json_schema_path is None:
-            dir_path = os.path.dirname(os.path.realpath(__file__))
-            args.json_schema_path = os.path.join(dir_path,
-                                                 "structured_schemas",
-                                                 "structured_schema_1.json")
-        with open(args.json_schema_path) as f:
-            schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "grammar":
-        schema = """
-            ?start: select_statement
-
-            ?select_statement: "SELECT " column_list " FROM " table_name
-
-            ?column_list: column_name ("," column_name)*
-
-            ?table_name: identifier
-
-            ?column_name: identifier
-
-            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
-        """
-        prompt = "Generate an SQL query to show the 'username' \
-            and 'email' from the 'users' table."
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "regex":
-        regex = r"\w+@\w+\.com\n"
-        args.regex = regex
-        prompt = "Generate an email address for Alan Turing, \
-            who works in Enigma. End in .com and new line. \
-                Example result: alan.turing@enigma.com\n"
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=regex,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "choice":
-        choice = ["Positive", "Negative"]
-        args.choice = choice
-        prompt = "Classify this sentiment: vLLM is wonderful!"
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=choice,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "xgrammar_bench":
-        args.warmup = False
-        requests: list[SampleRequest] = []
-        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
-                                        split="train")
-        print(f"dataset has {len(dataset)} entries")
-        len_dataset = len(dataset)
-        for data_point_idx in range(args.num_prompts):
-            idx = data_point_idx
-            while idx >= len_dataset:
-                idx -= len_dataset
-            schema = dataset["schema"][idx]
-            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
-                                                   tokenize=False)
-            input_len = len(tokenizer(prompt).input_ids)
-            completion = dataset["completion"][idx]
-
-            requests.append(
-                SampleRequest(prompt=prompt,
-                              prompt_len=input_len,
-                              expected_output_len=args.output_len,
-                              schema=schema,
-                              completion=completion))
-
-    return requests
-
-
-def evaluate(ret, args):
-
-    def _eval_correctness_json(expected, actual):
-        # extract json string from string using regex
-        import re
-        actual = actual.replace('\n', '').replace(' ', '').strip()
-        try:
-            actual = re.search(r'\{.*\}', actual).group()
-            actual = json.loads(actual)
-        except Exception:
-            return False
-
-        return True
-
-    def _eval_correctness_choice(expected, actual):
-        return actual in args.choice
-
-    def _eval_correctness_regex(expected, actual):
-        import re
-        return re.match(args.regex, actual) is not None
-
-    def _eval_correctness(expected, actual):
-        if args.structure_type == 'json':
-            return _eval_correctness_json(expected, actual)
-        elif args.structure_type == 'regex':
-            return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == 'choice':
-            return _eval_correctness_choice(expected, actual)
-        else:
-            return None
-
-    scores = []
-    for res in ret:
-        score = _eval_correctness(res['expected'], res['generated'])
-        res['correctness'] = score
-        scores.append(score)
-
-    not_none_scores = [score for score in scores if score is not None]
-
-    return (sum(not_none_scores) / len(not_none_scores) *
-            100) if len(not_none_scores) > 0 else None
-
-
-def main(args: argparse.Namespace):
-    print(args)
-    random.seed(args.seed)
-
-    # async engine is working for 'regex', 'choice' and 'grammar'
-    if args.dataset == 'grammar':
-        args.structure_type = 'grammar'
-        args.async_engine = False
-    elif args.dataset == 'regex':
-        args.structure_type = 'regex'
-        args.async_engine = False
-    elif args.dataset == 'choice':
-        args.structure_type = 'choice'
-        args.async_engine = False
-    else:
-        args.structure_type = 'json'
-
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
-    if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
-        result_file_name += f"_{args.model.split('/')[-1]}"
-        result_file_name += f"_{args.dataset}"
-        result_file_name += f"_{args.num_prompts}"
-        result_file_name += f"_out{args.output_len}"
-        result_file_name += f"_async{args.async_engine}"
-        result_file_name += f"_warmup{args.warmup}"
-        result_file_name += f"_chunkedprefill{args.enable_chunked_prefill}"
-        result_file_name += ".txt"
-    else:
-        result_file_name = None
-
-    # Synthesize a prompt with the given input length.
-    tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code)
-    requests = sample_requests(tokenizer, args)
-
-    if args.async_engine:
-        engine_args = AsyncEngineArgs.from_cli_args(args)
-        elapsed_time, ret, (first_latency, next_latency) = uvloop.run(
-            run_vllm_async(requests, engine_args, args.n,
-                           args.guided_decoding_ratio, args.warmup,
-                           args.disable_frontend_multiprocessing))
-    else:
-        engine_args = EngineArgs.from_cli_args(args)
-        elapsed_time, ret = run_vllm(requests, engine_args, args.n,
-                                     args.guided_decoding_ratio, args.warmup)
-        first_latency, next_latency = None, None
-
-    score = evaluate(ret, args)
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if first_latency is not None:
-        latency_breakdown = "\nFirst token latency(msecs):\n"
-        latency_breakdown += f"{first_latency.describe()}"
-        latency_breakdown += "\nNext token latency(msecs):\n"
-        latency_breakdown += f"{next_latency.describe()}"
-    print(
-        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
-        f"{total_output_tokens / elapsed_time:.2f} output tokens/s",
-        f"Correct rate is {score} %",
-        f"{latency_breakdown if first_latency is not None else ''}")
-
-    # Output JSON results if specified
-    if args.output_json or result_file_name:
-        results = {
-            "elapsed_time": elapsed_time,
-            "num_requests": len(requests),
-            "total_num_tokens": total_num_tokens,
-            "total_output_tokens": total_output_tokens,
-            "requests_per_second": len(requests) / elapsed_time,
-            "tokens_per_second": f"{total_num_tokens / elapsed_time:.2f}",
-            "output_tokens_per_second":
-            f"{total_output_tokens / elapsed_time:.2f}",
-            "correct_rate(%)": score
-        }
-        results = {"outputs": ret, **results}
-        if first_latency is not None:
-            results["first_token_latency(msecs)"] = first_latency.describe(
-            ).to_dict()
-            results["next_token_latency(msecs)"] = next_latency.describe(
-            ).to_dict()
-        if args.output_json:
-            with open(args.output_json, "w") as f:
-                json.dump(results, f, indent=4)
-        elif result_file_name:
-            with open(result_file_name, "w") as f:
-                json.dump(results, f, indent=4)
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
-    parser = AsyncEngineArgs.add_cli_args(parser)
-
-    parser.add_argument("--output-len",
-                        type=int,
-                        default=512,
-                        help="Output length for each request. Overrides the "
-                        "output length from the dataset.")
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
-    parser.add_argument("--json_schema_path",
-                        type=str,
-                        default=None,
-                        help="Path to json schema.")
-    parser.add_argument("--n",
-                        type=int,
-                        default=1,
-                        help="Number of generated sequences per prompt.")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=10,
-                        help="Number of prompts to process.")
-    parser.add_argument(
-        '--output-json',
-        type=str,
-        default=None,
-        help='Path to save the throughput results in JSON format.')
-    parser.add_argument("--async-engine",
-                        action='store_true',
-                        default=False,
-                        help="Use vLLM async engine rather than LLM class.")
-    parser.add_argument("--no-guided-decoding",
-                        action='store_true',
-                        default=False,
-                        help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
-                        type=float,
-                        default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--disable-frontend-multiprocessing",
-                        action='store_true',
-                        default=False,
-                        help="Disable decoupled async engine frontend.")
-    parser.add_argument("--warmup",
-                        action="store_true",
-                        default=False,
-                        help="Run warmup prompts before benchmark.")
-    parser.add_argument("--save-results",
-                        action="store_true",
-                        default=False,
-                        help="save output results.")
-    args = parser.parse_args()
-    if args.tokenizer is None:
-        args.tokenizer = args.model
-    main(args)

benchmarks/benchmark_latency.py

@@ -52,6 +52,7 @@ def main(args: argparse.Namespace):
         top_p=1.0,
         ignore_eos=True,
         max_tokens=args.output_len,
+        detokenize=not args.disable_detokenize,
     )
     print(sampling_params)
     dummy_prompt_token_ids = np.random.randint(10000,
@@ -173,6 +174,12 @@ if __name__ == "__main__":
         default=None,
         help="Path to save the latency results in JSON format.",
     )
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prefix_caching.py

@@ -63,14 +63,16 @@ class Request:
     output_len: int
 
 
-def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> str:
+def sample_tokens(tokenizer: PreTrainedTokenizerBase,
+                  length: int) -> list[int]:
     vocab = tokenizer.get_vocab()
+    all_special_ids = set(tokenizer.all_special_ids)
+
     # Remove the special tokens.
-    vocab = {
-        k: v
-        for k, v in vocab.items() if k not in tokenizer.all_special_ids
-    }
-    return random.choices(list(vocab.values()), k=length)
+    return random.choices(
+        [v for k, v in vocab.items() if k not in all_special_ids],
+        k=length,
+    )
 
 
 def sample_requests_from_dataset(
@@ -194,7 +196,9 @@ def main(args):
 
     llm = LLM(**dataclasses.asdict(engine_args))
 
-    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+    sampling_params = SamplingParams(temperature=0,
+                                     max_tokens=args.output_len,
+                                     detokenize=not args.disable_detokenize)
 
     print("Testing filtered requests")
     prompts = repeat_and_sort_requests(filtered_requests,
@@ -243,6 +247,12 @@ if __name__ == "__main__":
         "subtract this length when filtering prompts. Only used "
         "when dataset-path is not provided.",
     )
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prioritization.py

@@ -23,7 +23,7 @@ def sample_requests(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     fixed_output_len: Optional[int],
-) -> list[tuple[str, int, int]]:
+) -> list[tuple[str, int, int, int]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
 
@@ -71,6 +71,7 @@ def run_vllm(
     requests: list[tuple[str, int, int]],
     n: int,
     engine_args: EngineArgs,
+    disable_detokenize: bool = False,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
@@ -95,6 +96,7 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=output_len,
+                detokenize=not disable_detokenize,
             ))
 
     start = time.perf_counter()
@@ -121,7 +123,8 @@ def main(args: argparse.Namespace):
 
     if args.backend == "vllm":
         elapsed_time = run_vllm(requests, args.n,
-                                EngineArgs.from_cli_args(args))
+                                EngineArgs.from_cli_args(args),
+                                args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
     total_num_tokens = sum(prompt_len + output_len
@@ -174,6 +177,12 @@ if __name__ == "__main__":
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_serving.py

@@ -7,9 +7,6 @@ On the server side, run one of the following commands:
         --swap-space 16 \
         --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
@@ -25,25 +22,21 @@ On the client side, run:
 """
 import argparse
 import asyncio
-import base64
 import gc
-import io
 import json
 import os
 import random
 import time
 import warnings
-from collections.abc import AsyncGenerator, Collection
+from collections.abc import AsyncGenerator, Iterable
 from dataclasses import dataclass
 from datetime import datetime
 from typing import Any, Optional
 
 import numpy as np
-import pandas as pd
-from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+from backend_request_func import (ASYNC_REQUEST_FUNCS,
+                                  OPENAI_COMPATIBLE_BACKENDS, RequestFuncInput,
                                   RequestFuncOutput)
-from datasets import load_dataset
-from PIL.Image import Image
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
@@ -57,6 +50,12 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from benchmark_dataset import (AIMODataset, ASRDataset, BurstGPTDataset,
+                               ConversationDataset, HuggingFaceDataset,
+                               InstructCoderDataset, MTBenchDataset,
+                               NextEditPredictionDataset, RandomDataset,
+                               SampleRequest, ShareGPTDataset, SonnetDataset,
+                               VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@@ -92,325 +91,18 @@ class BenchmarkMetrics:
     percentiles_e2el_ms: list[tuple[float, float]]
 
 
-def sample_sharegpt_requests(
-    dataset_path: str,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int] = None,
-) -> list[tuple[str, int, int, None]]:
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Only keep the first two turns of each conversation.
-    dataset = [(data["conversations"][0]["value"],
-                data["conversations"][1]["value"]) for data in dataset]
-
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: list[tuple[str, int, int]] = []
-    for i in range(len(dataset)):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = dataset[i][0]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = dataset[i][1]
-        completion_token_ids = tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if prompt_len < 4 or (fixed_output_len is None and output_len < 4):
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append((prompt, prompt_len, output_len, None))
-
-    return filtered_dataset
-
-
-def sample_burstgpt_requests(
-    dataset_path: str,
-    num_requests: int,
-    random_seed: int,
-    tokenizer: PreTrainedTokenizerBase,
-) -> list[tuple[str, int, int, None]]:
-    df = pd.read_csv(dataset_path)
-    gpt4_df = df[df["Model"] == "GPT-4"]
-    # Remove the failed requests (i.e., response length is 0)
-    gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
-    # Randomly sample num_requests from the dataset
-    if num_requests <= len(gpt4_df):
-        gpt4_df = gpt4_df.sample(n=num_requests, random_state=random_seed)
-    else:
-        gpt4_df = gpt4_df.sample(n=num_requests,
-                                 random_state=random_seed,
-                                 replace=True)
-    # Convert the dataframe to a list of tuples
-    dataset = gpt4_df.values.tolist()
-    input_requests = []
-    for i in range(num_requests):
-        input_len = int(dataset[i][2])
-        output_len = int(dataset[i][3])
-        prompt = tokenizer.decode([(i + j) % tokenizer.vocab_size
-                                   for j in range(input_len)])
-        input_requests.append((prompt, input_len, output_len, None))
-    return input_requests
-
-
-def sample_sonnet_requests(
-    dataset_path: str,
-    num_requests: int,
-    input_len: int,
-    output_len: int,
-    prefix_len: int,
-    tokenizer: PreTrainedTokenizerBase,
-) -> list[tuple[str, str, int, int, None]]:
-    assert (
-        input_len > prefix_len
-    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
-
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        poem_lines = f.readlines()
-
-    # Tokenize the poem lines.
-    poem_token_ids = tokenizer(poem_lines).input_ids
-    average_poem_len = sum(
-        len(token_ids) for token_ids in poem_token_ids) / len(poem_token_ids)
-
-    # Base prefix for all requests.
-    base_prompt = "Pick as many lines as you can from these poem lines:\n"
-    base_message = [{
-        "role": "user",
-        "content": base_prompt,
-    }]
-    base_prompt_formatted = tokenizer.apply_chat_template(
-        base_message, add_generation_prompt=True, tokenize=False)
-    base_prompt_offset = len(tokenizer(base_prompt_formatted).input_ids)
-
-    assert (
-        input_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-input-len' higher than {base_prompt_offset}."
-    num_input_lines = round(
-        (input_len - base_prompt_offset) / average_poem_len)
-
-    # First approximately `prefix_len` number of tokens in the
-    # prompt are fixed poem lines.
-    assert (
-        prefix_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-prefix-len' higher than {base_prompt_offset}."
-
-    num_prefix_lines = round(
-        (prefix_len - base_prompt_offset) / average_poem_len)
-    prefix_lines = poem_lines[:num_prefix_lines]
-
-    # Sample the rest of lines per request.
-    sampled_requests: list[tuple[str, int, int]] = []
-    for _ in range(num_requests):
-        num_lines_needed = num_input_lines - num_prefix_lines
-        sampled_lines = "".join(prefix_lines +
-                                random.choices(poem_lines, k=num_lines_needed))
-
-        prompt = f"{base_prompt}{sampled_lines}"
-        message = [
-            {
-                "role": "user",
-                "content": prompt,
-            },
-        ]
-        prompt_formatted = tokenizer.apply_chat_template(
-            message, add_generation_prompt=True, tokenize=False)
-        prompt_len = len(tokenizer(prompt_formatted).input_ids)
-        sampled_requests.append(
-            (prompt, prompt_formatted, prompt_len, output_len, None))
-
-    return sampled_requests
-
-
-def sample_vision_arena_requests(
-    dataset,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int] = None,
-) -> list[tuple[str, str, int, Optional[dict[str, Collection[str]]]]]:
-    sampled_requests: list[tuple[str, int, int, dict[str,
-                                                     Collection[str]]]] = []
-    for data in dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        prompt = data["turns"][0][0]['content']
-
-        prompt_token_ids = tokenizer(prompt).input_ids
-        if fixed_output_len is None:
-            # Default max output len is set to 128
-            print("--hf-output-len is not provided. Using default value 128.")
-            fixed_output_len = 128
-
-        prompt_len = len(prompt_token_ids)
-        output_len = fixed_output_len
-
-        assert isinstance(
-            data["images"][0],
-            Image), ("Input image format must be `PIL.Image.Image`, "
-                     f"given {type(data['image'])}.")
-        image: Image = data["images"][0]
-        image = image.convert("RGB")
-        image_data = io.BytesIO()
-        image.save(image_data, format='JPEG')
-        image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
-        mm_content = {
-            "type": "image_url",
-            "image_url": {
-                "url": f"data:image/jpeg;base64,{image_base64}"
-            },
-        }
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_hf_requests(
-    dataset_path: str,
-    dataset_subset: Optional[str],
-    dataset_split: str,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    random_seed: int,
-    fixed_output_len: Optional[int] = None,
-) -> list[tuple[str, str, int, Optional[dict[str, Collection[str]]]]]:
-
-    # Special case for vision_arena dataset
-    if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \
-        and dataset_subset is None:
-        assert dataset_split == "train"
-        dataset = load_dataset(dataset_path,
-                               name=dataset_subset,
-                               split=dataset_split,
-                               streaming=True)
-        dataset = dataset.shuffle(seed=random_seed)
-        return sample_vision_arena_requests(dataset, num_requests, tokenizer,
-                                            fixed_output_len)
-
-    dataset = load_dataset(dataset_path,
-                           name=dataset_subset,
-                           split=dataset_split,
-                           streaming=True)
-    assert "conversations" in dataset.features, (
-        "HF Dataset must have 'conversations' column.")
-    filter_func = lambda x: len(x["conversations"]) >= 2
-    filtered_dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
-    sampled_requests: list[tuple[str, int, int, dict[str,
-                                                     Collection[str]]]] = []
-    for data in filtered_dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = data["conversations"][0]["value"]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = data["conversations"][1]["value"]
-        completion_token_ids = tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if fixed_output_len is None and (prompt_len < 4 or output_len < 4):
-            # Prune too short sequences.
-            continue
-        if fixed_output_len is None and \
-            (prompt_len > 1024 or prompt_len + output_len > 2048):
-            # Prune too long sequences.
-            continue
-
-        if "image" in data and isinstance(data["image"], Image):
-            image: Image = data["image"]
-            image = image.convert("RGB")
-            image_data = io.BytesIO()
-            image.save(image_data, format='JPEG')
-            image_base64 = base64.b64encode(
-                image_data.getvalue()).decode("utf-8")
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": f"data:image/jpeg;base64,{image_base64}"
-                },
-            }
-        elif "image" in data and isinstance(data["image"], str):
-            if (data["image"].startswith("http://") or \
-                data["image"].startswith("file://")):
-                image_url = data["image"]
-            else:
-                image_url = f"file://{data['image']}"
-
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": image_url
-                },
-            }
-        else:
-            mm_content = None
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_random_requests(
-    prefix_len: int,
-    input_len: int,
-    output_len: int,
-    num_prompts: int,
-    range_ratio: float,
-    tokenizer: PreTrainedTokenizerBase,
-) -> list[tuple[str, int, int]]:
-    prefix_token_ids = np.random.randint(0,
-                                         tokenizer.vocab_size,
-                                         size=prefix_len).tolist()
-
-    input_lens = np.random.randint(
-        int(input_len * range_ratio),
-        input_len + 1,
-        size=num_prompts,
-    )
-    output_lens = np.random.randint(
-        int(output_len * range_ratio),
-        output_len + 1,
-        size=num_prompts,
-    )
-    offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
-    input_requests = []
-    for i in range(num_prompts):
-        prompt = tokenizer.decode(prefix_token_ids +
-                                  [(offsets[i] + i + j) % tokenizer.vocab_size
-                                   for j in range(input_lens[i])])
-
-        input_requests.append((prompt, int(prefix_len + input_lens[i]),
-                               int(output_lens[i]), None))
-
-    return input_requests
-
-
 async def get_request(
-    input_requests: list[tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[tuple[str, int, int], None]:
+) -> AsyncGenerator[SampleRequest, None]:
     """
     Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
 
     Args:
         input_requests:
-            A list of input requests, each represented as a tuple.
+            A list of input requests, each represented as a SampleRequest.
         request_rate:
             The rate at which requests are generated (requests/s).
         burstiness (optional):
@@ -422,7 +114,7 @@ async def get_request(
             in more bursty requests, while a higher burstiness value
             (burstiness > 1) results in a more uniform arrival of requests.
     """
-    input_requests = iter(input_requests)
+    input_requests: Iterable[SampleRequest] = iter(input_requests)
 
     # Calculate scale parameter theta to maintain the desired request_rate.
     assert burstiness > 0, (
@@ -444,7 +136,7 @@ async def get_request(
 
 
 def calculate_metrics(
-    input_requests: list[tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     outputs: list[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
@@ -465,7 +157,7 @@ def calculate_metrics(
         if outputs[i].success:
             output_len = outputs[i].output_tokens
 
-            if output_len is None:
+            if not output_len:
                 # We use the tokenizer to count the number of output tokens
                 # for some serving backends instead of looking at
                 # len(outputs[i].itl) since multiple output tokens may be
@@ -475,7 +167,7 @@ def calculate_metrics(
                     tokenizer(outputs[i].generated_text,
                               add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
-            total_input += input_requests[i][1]
+            total_input += input_requests[i].prompt_len
             tpot = 0
             if output_len > 1:
                 latency_minus_ttft = outputs[i].latency - outputs[i].ttft
@@ -558,19 +250,19 @@ async def benchmark(
     model_id: str,
     model_name: str,
     tokenizer: PreTrainedTokenizerBase,
-    input_requests: list[tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     logprobs: Optional[int],
-    best_of: int,
     request_rate: float,
     burstiness: float,
     disable_tqdm: bool,
     profile: bool,
     selected_percentile_metrics: list[str],
-    selected_percentiles: list[str],
+    selected_percentiles: list[float],
     ignore_eos: bool,
     goodput_config_dict: dict[str, float],
     max_concurrency: Optional[int],
-    lora_modules: Optional[list[str]],
+    lora_modules: Optional[Iterable[str]],
+    extra_body: Optional[dict],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -578,12 +270,12 @@ async def benchmark(
         raise ValueError(f"Unknown backend: {backend}")
 
     print("Starting initial single prompt test run...")
-    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
-        input_requests[0])
-    if backend != "openai-chat" and test_mm_content is not None:
-        # multi-modal benchmark is only available on OpenAI Chat backend.
-        raise ValueError(
-            "Multi-modal content is only supported on 'openai-chat' backend.")
+    test_prompt, test_prompt_len, test_output_len, test_mm_content = \
+        input_requests[0].prompt, input_requests[0].prompt_len, \
+        input_requests[0].expected_output_len, \
+            input_requests[0].multi_modal_data
+
+    assert test_mm_content is None or isinstance(test_mm_content, dict)
     test_input = RequestFuncInput(
         model=model_id,
         model_name=model_name,
@@ -592,9 +284,9 @@ async def benchmark(
         prompt_len=test_prompt_len,
         output_len=test_output_len,
         logprobs=logprobs,
-        best_of=best_of,
         multi_modal_content=test_mm_content,
         ignore_eos=ignore_eos,
+        extra_body=extra_body,
     )
 
     test_output = await request_func(request_func_input=test_input)
@@ -608,7 +300,8 @@ async def benchmark(
     if lora_modules:
         # For each input request, choose a LoRA module at random.
         lora_modules = iter(
-            [random.choice(lora_modules) for _ in range(len(input_requests))])
+            [random.choice(lora_modules) \
+                for _ in range(len(input_requests))])
 
     if profile:
         print("Starting profiler...")
@@ -619,9 +312,9 @@ async def benchmark(
                                          prompt_len=test_prompt_len,
                                          output_len=test_output_len,
                                          logprobs=logprobs,
-                                         best_of=best_of,
                                          multi_modal_content=test_mm_content,
-                                         ignore_eos=ignore_eos)
+                                         ignore_eos=ignore_eos,
+                                         extra_body=extra_body)
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
             print("Profiler started")
@@ -655,7 +348,9 @@ async def benchmark(
     benchmark_start_time = time.perf_counter()
     tasks: list[asyncio.Task] = []
     async for request in get_request(input_requests, request_rate, burstiness):
-        prompt, prompt_len, output_len, mm_content = request
+        prompt, prompt_len, output_len, mm_content = request.prompt, \
+            request.prompt_len, request.expected_output_len, \
+                request.multi_modal_data
         req_model_id, req_model_name = model_id, model_name
         if lora_modules:
             req_lora_module = next(lora_modules)
@@ -668,9 +363,9 @@ async def benchmark(
                                               prompt_len=prompt_len,
                                               output_len=output_len,
                                               logprobs=logprobs,
-                                              best_of=best_of,
                                               multi_modal_content=mm_content,
-                                              ignore_eos=ignore_eos)
+                                              ignore_eos=ignore_eos,
+                                              extra_body=extra_body)
         tasks.append(
             asyncio.create_task(
                 limited_request_func(request_func_input=request_func_input,
@@ -686,7 +381,6 @@ async def benchmark(
             prompt_len=test_prompt_len,
             output_len=test_output_len,
             logprobs=logprobs,
-            best_of=best_of,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -872,78 +566,130 @@ def main(args: argparse.Namespace):
             "Please specify '--dataset-name' and the corresponding "
             "'--dataset-path' if required.")
 
-    elif args.dataset_name == "sharegpt":
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
-
-    elif args.dataset_name == "burstgpt":
-        input_requests = sample_burstgpt_requests(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            random_seed=args.seed,
-            tokenizer=tokenizer,
-        )
-
-    elif args.dataset_name == "sonnet":
-        # Do not format the prompt, pass to message directly
+    if args.dataset_name == "sonnet":
+        dataset = SonnetDataset(dataset_path=args.dataset_path)
+        # For the "sonnet" dataset, formatting depends on the backend.
         if args.backend == "openai-chat":
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=False)
         else:
-            assert (
-                tokenizer.chat_template or tokenizer.default_chat_template
-            ), "Tokenizer/model must have chat template for sonnet dataset."
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt_formatted, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            assert tokenizer.chat_template or tokenizer.default_chat_template, (
+                "Tokenizer/model must have chat template for sonnet dataset.")
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=True)
 
     elif args.dataset_name == "hf":
-        input_requests = sample_hf_requests(
+        # all following datasets are implemented from the
+        # HuggingFaceDataset base class
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = VisionArenaDataset
+            args.hf_split = "train"
+            args.hf_subset = None
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = InstructCoderDataset
+            args.hf_split = "train"
+        elif args.dataset_path in MTBenchDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = MTBenchDataset
+            args.hf_split = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = ConversationDataset
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = AIMODataset
+            args.hf_split = "train"
+        elif args.dataset_path in NextEditPredictionDataset.SUPPORTED_DATASET_PATHS:  # noqa: E501
+            dataset_class = NextEditPredictionDataset
+            args.hf_split = "train"
+        elif args.dataset_path in ASRDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = ASRDataset
+            args.hf_split = "train"
+        else:
+            supported_datasets = set([
+                dataset_name for cls in HuggingFaceDataset.__subclasses__()
+                for dataset_name in cls.SUPPORTED_DATASET_PATHS
+            ])
+            raise ValueError(
+                f"Unsupported dataset path: {args.dataset_path}. "
+                "Huggingface dataset only supports dataset_path"
+                f" from one of following: {supported_datasets}. "
+                "Please consider contributing if you would "
+                "like to add support for additional dataset formats.")
+
+        if (dataset_class.IS_MULTIMODAL and backend not in \
+            ["openai-chat", "openai-audio"]):
+            # multi-modal benchmark is only available on OpenAI Chat backend.
+            raise ValueError(
+                "Multi-modal content is only supported on 'openai-chat' and " \
+                "'openai-audio' backend.")
+        input_requests = dataset_class(
             dataset_path=args.dataset_path,
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
+            random_seed=args.seed,
+        ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
-            random_seed=args.seed,
-            fixed_output_len=args.hf_output_len,
-        )
-
-    elif args.dataset_name == "random":
-        input_requests = sample_random_requests(
-            prefix_len=args.random_prefix_len,
-            input_len=args.random_input_len,
-            output_len=args.random_output_len,
-            num_prompts=args.num_prompts,
-            range_ratio=args.random_range_ratio,
-            tokenizer=tokenizer,
+            output_len=args.hf_output_len,
         )
 
     else:
-        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+        # For datasets that follow a similar structure, use a mapping.
+        dataset_mapping = {
+            "sharegpt":
+            lambda: ShareGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).sample(
+                                        tokenizer=tokenizer,
+                                        num_requests=args.num_prompts,
+                                        output_len=args.sharegpt_output_len,
+                                    ),
+            "burstgpt":
+            lambda: BurstGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).
+            sample(tokenizer=tokenizer, num_requests=args.num_prompts),
+            "random":
+            lambda: RandomDataset(dataset_path=args.dataset_path).sample(
+                tokenizer=tokenizer,
+                num_requests=args.num_prompts,
+                prefix_len=args.random_prefix_len,
+                input_len=args.random_input_len,
+                output_len=args.random_output_len,
+                range_ratio=args.random_range_ratio,
+            )
+        }
 
+        try:
+            input_requests = dataset_mapping[args.dataset_name]()
+        except KeyError as err:
+            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
     goodput_config_dict = check_goodput_args(args)
 
+    # Collect the sampling parameters.
+    sampling_params = {
+        k: v
+        for k, v in {
+            "top_p": args.top_p,
+            "top_k": args.top_k,
+            "min_p": args.min_p,
+            "temperature": args.temperature
+        }.items() if v is not None
+    }
+
+    # Sampling parameters are only supported by openai-compatible backend.
+    if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS:
+        raise ValueError(
+            "Sampling parameters are only supported by openai-compatible "
+            "backends.")
+
+    if "temperature" not in sampling_params:
+        sampling_params["temperature"] = 0.0  # Default to greedy decoding.
+
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -958,7 +704,6 @@ def main(args: argparse.Namespace):
             tokenizer=tokenizer,
             input_requests=input_requests,
             logprobs=args.logprobs,
-            best_of=args.best_of,
             request_rate=args.request_rate,
             burstiness=args.burstiness,
             disable_tqdm=args.disable_tqdm,
@@ -971,10 +716,11 @@ def main(args: argparse.Namespace):
             goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
+            extra_body=sampling_params,
         ))
 
     # Save config and results to json
-    if args.save_result:
+    if args.save_result or args.append_result:
         result_json: dict[str, Any] = {}
 
         # Setup
@@ -983,7 +729,6 @@ def main(args: argparse.Namespace):
         result_json["backend"] = backend
         result_json["model_id"] = model_id
         result_json["tokenizer_id"] = tokenizer_id
-        result_json["best_of"] = args.best_of
         result_json["num_prompts"] = args.num_prompts
 
         # Metadata
@@ -996,7 +741,6 @@ def main(args: argparse.Namespace):
                     raise ValueError(
                         "Invalid metadata format. Please use KEY=VALUE format."
                     )
-
         # Traffic
         result_json["request_rate"] = (args.request_rate if args.request_rate
                                        < float("inf") else "inf")
@@ -1006,6 +750,15 @@ def main(args: argparse.Namespace):
         # Merge with benchmark result
         result_json = {**result_json, **benchmark_result}
 
+        if not args.save_detailed:
+            # Remove fields with too many data points
+            for field in [
+                    "input_lens", "output_lens", "ttfts", "itls",
+                    "generated_texts", "errors"
+            ]:
+                if field in result_json:
+                    del result_json[field]
+
         # Save to file
         base_model_id = model_id.split("/")[-1]
         max_concurrency_str = (f"-concurrency{args.max_concurrency}"
@@ -1015,7 +768,12 @@ def main(args: argparse.Namespace):
             file_name = args.result_filename
         if args.result_dir:
             file_name = os.path.join(args.result_dir, file_name)
-        with open(file_name, "w", encoding='utf-8') as outfile:
+        with open(file_name,
+                  mode="a+" if args.append_result else "w",
+                  encoding='utf-8') as outfile:
+            # Append a newline.
+            if args.append_result and outfile.tell() != 0:
+                outfile.write("\n")
             json.dump(result_json, outfile)
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
@@ -1081,13 +839,6 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
-    parser.add_argument(
-        "--best-of",
-        type=int,
-        default=1,
-        help="Generates `best_of` sequences per prompt and "
-        "returns the best one.",
-    )
     parser.add_argument("--use-beam-search", action="store_true")
     parser.add_argument(
         "--num-prompts",
@@ -1148,6 +899,17 @@ if __name__ == "__main__":
         action="store_true",
         help="Specify to save benchmark results to a json file",
     )
+    parser.add_argument(
+        "--save-detailed",
+        action="store_true",
+        help="When saving the results, whether to include per request "
+        "information such as response, error, ttfs, tpots, etc.",
+    )
+    parser.add_argument(
+        "--append-result",
+        action="store_true",
+        help="Append the benchmark result to the existing json file.",
+    )
     parser.add_argument(
         "--metadata",
         metavar="KEY=VALUE",
@@ -1181,7 +943,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -1189,7 +951,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -1256,18 +1018,23 @@ if __name__ == "__main__":
     random_group.add_argument(
         "--random-range-ratio",
         type=float,
-        default=1.0,
-        help="Range of sampled ratio of input/output length, "
-        "used only for random sampling.",
+        default=0.0,
+        help="Range ratio for sampling input/output length, "
+        "used only for random sampling. Must be in the range [0, 1) to define "
+        "a symmetric sampling range"
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
     random_group.add_argument(
         "--random-prefix-len",
         type=int,
         default=0,
-        help="Number of fixed prefix tokens before random "
-        " context. The length range of context in a random "
-        " request is [random-prefix-len, "
-        " random-prefix-len + random-prefix-len * random-range-ratio).")
+        help=("Number of fixed prefix tokens before the random context "
+              "in a request. "
+              "The total input length is the sum of `random-prefix-len` and "
+              "a random "
+              "context length sampled from [input_len * (1 - range_ratio), "
+              "input_len * (1 + range_ratio)]."),
+    )
 
     hf_group = parser.add_argument_group("hf dataset options")
     hf_group.add_argument("--hf-subset",
@@ -1286,6 +1053,33 @@ if __name__ == "__main__":
         "from the sampled HF dataset.",
     )
 
+    sampling_group = parser.add_argument_group("sampling parameters")
+    sampling_group.add_argument(
+        "--top-p",
+        type=float,
+        default=None,
+        help="Top-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--top-k",
+        type=int,
+        default=None,
+        help="Top-k sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--min-p",
+        type=float,
+        default=None,
+        help="Min-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--temperature",
+        type=float,
+        default=None,
+        help="Temperature sampling parameter. Only has effect on "
+        "openai-compatible backends. If not specified, default to greedy "
+        "decoding (i.e. temperature==0.0).")
+
     parser.add_argument(
         '--tokenizer-mode',
         type=str,
@@ -1312,4 +1106,5 @@ if __name__ == "__main__":
                         "script chooses a LoRA module at random.")
 
     args = parser.parse_args()
+
     main(args)

benchmarks/benchmark_serving_structured_output.py

@@ -1,20 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
-r"""Benchmark online serving throughput with guided decoding.
+r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
     (vLLM OpenAI API server)
     vllm serve <your_model> --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
-    python benchmarks/benchmark_serving_guided.py \
+    python benchmarks/benchmark_serving_structured_output.py \
         --backend <backend> \
         --model <your_model> \
         --dataset json \
-        --guided-decoding-ratio 1.0 \
-        --guided-decoding-backend xgrammar \
+        --structured-output-ratio 1.0 \
+        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -24,11 +21,13 @@ On the client side, run:
 """
 import argparse
 import asyncio
+import copy
 import dataclasses
 import json
 import os
 import random
 import time
+import uuid
 import warnings
 from collections.abc import AsyncGenerator
 from dataclasses import dataclass
@@ -52,6 +51,9 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from vllm.v1.structured_output.backend_xgrammar import (
+    has_xgrammar_unsupported_json_features)
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
 
@@ -106,39 +108,61 @@ class SampleRequest:
 
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
                     args: argparse.Namespace) -> list[SampleRequest]:
-    if args.dataset == 'json':
+    if args.dataset == 'json' or args.dataset == 'json-unique':
         if args.json_schema_path is None:
             dir_path = os.path.dirname(os.path.realpath(__file__))
             args.json_schema_path = os.path.join(dir_path,
                                                  "structured_schemas",
                                                  "structured_schema_1.json")
+        json_schemas = []
         with open(args.json_schema_path) as f:
             schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
+
+        if args.dataset == 'json-unique':
+            json_schemas = [
+                copy.deepcopy(schema) for _ in range(args.num_prompts)
+            ]
+            for i in range(len(json_schemas)):
+                if "properties" not in json_schemas[i]:
+                    json_schemas[i]["properties"] = {}
+                json_schemas[i]["properties"][
+                    f"__optional_field_{uuid.uuid4()}"] = {
+                        "type":
+                        "string",
+                        "description":
+                        "An unique optional field to avoid cached schemas"
+                    }
+        else:
+            json_schemas = [schema] * args.num_prompts
+
+        def gen_prompt(index: int):
+            return f"Generate an example of a brief user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
+
+        def get_schema(index: int):
+            return json_schemas[index % len(json_schemas)]
+
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
+            SampleRequest(prompt=gen_prompt(i),
+                          prompt_len=len(tokenizer(gen_prompt(i)).input_ids),
                           expected_output_len=args.output_len,
-                          schema=schema,
+                          schema=get_schema(i),
                           structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
+            for i in range(args.num_prompts)
         ]
 
     elif args.dataset == "grammar":
         schema = """
-            ?start: select_statement
+        root ::= select_statement
 
-            ?select_statement: "SELECT " column_list " FROM " table_name
+        select_statement ::= "SELECT " column " from " table " where " condition
 
-            ?column_list: column_name ("," column_name)*
+        column ::= "col_1 " | "col_2 "
 
-            ?table_name: identifier
+        table ::= "table_1 " | "table_2 "
 
-            ?column_name: identifier
+        condition ::= column "= " number
 
-            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
+        number ::= "1 " | "2 "
         """
         prompt = "Generate an SQL query to show the 'username' \
             and 'email' from the 'users' table."
@@ -191,7 +215,17 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
         requests: list[SampleRequest] = []
         dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                         split="train")
-        print(f"dataset has {len(dataset)} entries")
+        full_dataset_len = len(dataset)
+
+        def _filter_func(item):
+            import json
+            schema = json.loads(item["schema"])
+            return not has_xgrammar_unsupported_json_features(schema)
+
+        dataset = dataset.filter(_filter_func)
+        num_filtered_out = full_dataset_len - len(dataset)
+        print(f"dataset has {len(dataset)} entries after filtering "
+              f"out {num_filtered_out} entries with unsupported features")
         len_dataset = len(dataset)
         for data_point_idx in range(args.num_prompts):
             idx = data_point_idx
@@ -199,7 +233,8 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
                 idx -= len_dataset
             schema = dataset["schema"][idx]
             prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
-                                                   tokenize=False)
+                                                   tokenize=False,
+                                                   add_generation_prompt=True)
             input_len = len(tokenizer(prompt).input_ids)
             completion = dataset["completion"][idx]
 
@@ -220,21 +255,21 @@ async def get_request(
     burstiness: float = 1.0,
 ) -> AsyncGenerator[tuple[int, SampleRequest], None]:
     """
-    Asynchronously generates requests at a specified rate 
+    Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
-    
+
     Args:
-        input_requests: 
+        input_requests:
             A list of input requests, each represented as a tuple.
-        request_rate: 
+        request_rate:
             The rate at which requests are generated (requests/s).
-        burstiness (optional): 
-            The burstiness factor of the request generation. 
+        burstiness (optional):
+            The burstiness factor of the request generation.
             Only takes effect when request_rate is not inf.
             Default value is 1, which follows a Poisson process.
             Otherwise, the request intervals follow a gamma distribution.
-            A lower burstiness value (0 < burstiness < 1) results 
-            in more bursty requests, while a higher burstiness value 
+            A lower burstiness value (0 < burstiness < 1) results
+            in more bursty requests, while a higher burstiness value
             (burstiness > 1) results in a more uniform arrival of requests.
     """
     input_requests = iter(input_requests)
@@ -378,8 +413,7 @@ async def benchmark(
     selected_percentiles: list[str],
     ignore_eos: bool,
     max_concurrency: Optional[int],
-    guided_decoding_ratio: float,
-    guided_decoding_backend: str,
+    structured_output_ratio: float,
     goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
@@ -391,16 +425,16 @@ async def benchmark(
         extra_body = {}
         # Add the schema to the extra_body
         extra_body[request.structure_type] = request.schema
-        # Add the specific guided_decoding_backend
-        extra_body["guided_decoding_backend"] = guided_decoding_backend
         return extra_body
 
     print("Starting initial single prompt test run...")
-    guided_decoding_req_idx = random.sample(
+    structured_output_req_idx = random.sample(
         range(len(input_requests)),
-        int(len(input_requests) * guided_decoding_ratio))
+        int(len(input_requests) * structured_output_ratio))
 
     test_request = input_requests[0]
+    test_req_extra_body = (prepare_extra_body(test_request)
+                           if 0 in structured_output_req_idx else None)
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_request.prompt,
@@ -408,7 +442,7 @@ async def benchmark(
         prompt_len=test_request.prompt_len,
         output_len=test_request.expected_output_len,
         ignore_eos=ignore_eos,
-        extra_body=prepare_extra_body(test_request),
+        extra_body=test_req_extra_body,
     )
     test_output = await request_func(request_func_input=test_input)
     if not test_output.success:
@@ -427,7 +461,7 @@ async def benchmark(
             prompt_len=test_request.prompt_len,
             output_len=test_request.expected_output_len,
             ignore_eos=ignore_eos,
-            extra_body=prepare_extra_body(test_request),
+            extra_body=test_req_extra_body,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -465,7 +499,7 @@ async def benchmark(
     async for i, request in get_request(input_requests, request_rate,
                                         burstiness):
         extra_body = prepare_extra_body(
-            request) if i in guided_decoding_req_idx else None
+            request) if i in structured_output_req_idx else None
         request_func_input = RequestFuncInput(
             model=model_id,
             prompt=request.prompt,
@@ -696,8 +730,11 @@ def main(args: argparse.Namespace):
         api_url = f"http://{args.host}:{args.port}{args.endpoint}"
         base_url = f"http://{args.host}:{args.port}"
 
-    tokenizer = get_tokenizer(tokenizer_id,
-                              trust_remote_code=args.trust_remote_code)
+    tokenizer = get_tokenizer(
+        tokenizer_id,
+        trust_remote_code=args.trust_remote_code,
+        tokenizer_mode=args.tokenizer_mode,
+    )
 
     if args.dataset == 'grammar':
         args.structure_type = 'guided_grammar'
@@ -708,10 +745,10 @@ def main(args: argparse.Namespace):
     else:
         args.structure_type = 'guided_json'
 
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
+    if args.no_structured_output:
+        args.structured_output_ratio = 0
     if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
+        result_file_name = f'{args.structured_output_ratio}guided'
         result_file_name += f"_{backend}"
         result_file_name += f"_{args.request_rate}qps"
         result_file_name += f"_{args.model.split('/')[-1]}"
@@ -744,8 +781,7 @@ def main(args: argparse.Namespace):
             ],
             ignore_eos=args.ignore_eos,
             max_concurrency=args.max_concurrency,
-            guided_decoding_ratio=args.guided_decoding_ratio,
-            guided_decoding_backend=args.guided_decoding_backend,
+            structured_output_ratio=args.structured_output_ratio,
             goodput_config_dict=goodput_config_dict,
         ))
 
@@ -806,11 +842,13 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
-    parser.add_argument("--json_schema_path",
+    parser.add_argument("--dataset",
+                        default='json',
+                        choices=[
+                            'json', 'json-unique', 'grammar', 'regex',
+                            'choice', 'xgrammar_bench'
+                        ])
+    parser.add_argument("--json-schema-path",
                         type=str,
                         default=None,
                         help="Path to json schema.")
@@ -838,6 +876,13 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
+    parser.add_argument(
+        "--tokenizer-mode",
+        type=str,
+        default="auto",
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+    )
     parser.add_argument(
         "--num-prompts",
         type=int,
@@ -918,7 +963,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -926,7 +971,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -943,19 +988,14 @@ if __name__ == "__main__":
         "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
         "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
 
-    parser.add_argument("--no-guided-decoding",
+    parser.add_argument("--no-structured-output",
                         action='store_true',
                         default=False,
                         help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
+    parser.add_argument("--structured-output-ratio",
                         type=float,
                         default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--guided-decoding-backend",
-                        type=str,
-                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
-                        default="xgrammar",
-                        help="Backend to use for guided decoding")
+                        help="Ratio of Structured Outputs requests")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -6,13 +6,16 @@ import json
 import os
 import random
 import time
-from functools import cache
-from typing import Any, Optional
+import warnings
+from typing import Any, Optional, Union
 
 import torch
 import uvloop
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, InstructCoderDataset,
+                               RandomDataset, SampleRequest, ShareGPTDataset,
+                               SonnetDataset, VisionArenaDataset)
 from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
-from PIL import Image
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
@@ -20,155 +23,19 @@ from transformers import (AutoModelForCausalLM, AutoTokenizer,
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
     build_async_engine_client_from_engine_args)
-from vllm.inputs import TextPrompt
+from vllm.inputs import TextPrompt, TokensPrompt
 from vllm.lora.request import LoRARequest
-from vllm.lora.utils import get_adapter_absolute_path
-from vllm.multimodal import MultiModalDataDict
+from vllm.outputs import RequestOutput
 from vllm.sampling_params import BeamSearchParams
-from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
 
 
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    Attributes:
-        prompt: The input text prompt for the model.
-        prompt_len: The length of the prompt in tokens.
-        expected_output_len: The expected length of the output in tokens.
-        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
-            images).
-        lora_request: Optional LoRARequest specifying the LoRA to use. 
-    """
-    prompt: str
-    prompt_len: int
-    expected_output_len: int
-    multi_modal_data: Optional[MultiModalDataDict] = None
-    lora_request: Optional[LoRARequest] = None
-
-
-def _get_prompt_for_image_model(question: str, *, model: str) -> str:
-    """Prepend and append special tokens around the question to form a prompt.
-
-    Args:
-        question: The input question text to wrap with special tokens
-        model: The name of the model being used, to determine which special
-            tokens to add
-
-    Returns:
-        The formatted prompt string with appropriate special tokens for the
-            model
-
-    Raises:
-        ValueError: If an unsupported model name is provided
-    """
-    model = model.lower()
-    if "pixtral" in model:
-        return f"<s>[INST]{question}\n[IMG][/INST]"
-    raise ValueError(f"Unsupported model {model}")
-
-
-@cache
-def lora_path_on_disk(lora_path: str) -> str:
-    return get_adapter_absolute_path(lora_path)
-
-
-lora_tokenizer_cache: dict[int, AnyTokenizer] = {}
-
-
-def get_random_lora_request(
-        args: argparse.Namespace
-) -> tuple[LoRARequest, Optional[AnyTokenizer]]:
-    global lora_tokenizer_cache
-    lora_id = random.randint(1, args.max_loras)
-    lora_request = LoRARequest(lora_name=str(lora_id),
-                               lora_int_id=lora_id,
-                               lora_path=lora_path_on_disk(args.lora_path))
-    if lora_id not in lora_tokenizer_cache:
-        lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
-    return lora_request, lora_tokenizer_cache[lora_id]
-
-
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> list[SampleRequest]:
-
-    dataset_path: str = args.dataset
-    num_requests: int = args.num_prompts
-    fixed_output_len: Optional[int] = args.output_len
-    model: str = args.model
-    if fixed_output_len is not None and fixed_output_len < 4:
-        raise ValueError("output_len too small")
-
-    # Load the dataset.
-    with open(dataset_path) as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: list[SampleRequest] = []
-    for data in tqdm(dataset,
-                     total=len(filtered_dataset),
-                     desc="sampling requests"):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Only keep the first two turns of each conversation.
-        prompt = data["conversations"][0]["value"]
-        completion = data["conversations"][1]["value"]
-
-        multi_modal_data: Optional[MultiModalDataDict] = None
-        if "image" in data:
-            multi_modal_data = multi_modal_data or {}
-            image_path = data["image"]
-            # TODO(vllm-project/vllm/issues/9778): Support multiple images.
-            assert isinstance(image_path,
-                              str), "Only support single image input"
-            try:
-                multi_modal_data["image"] = Image.open(image_path).convert(
-                    "RGB")
-            except FileNotFoundError:
-                # Ignore datapoint where asset is missing
-                continue
-            prompt = _get_prompt_for_image_model(question=prompt, model=model)
-
-        request_tokenizer = tokenizer
-        lora_request: Optional[LoRARequest] = None
-        if args.enable_lora:
-            lora_request, lora_tokenizer = get_random_lora_request(args)
-            if lora_tokenizer:
-                request_tokenizer = lora_tokenizer
-
-        # Tokenize the prompts and completions.
-        prompt_token_ids = request_tokenizer(prompt).input_ids
-        completion_token_ids = request_tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if prompt_len < 4 or output_len < 4:
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append(
-            SampleRequest(prompt=prompt,
-                          prompt_len=prompt_len,
-                          expected_output_len=output_len,
-                          multi_modal_data=multi_modal_data,
-                          lora_request=lora_request))
-
-    return filtered_dataset
-
-
 def run_vllm(
     requests: list[SampleRequest],
     n: int,
     engine_args: EngineArgs,
-) -> float:
+    disable_detokenize: bool = False,
+) -> tuple[float, Optional[list[RequestOutput]]]:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
     assert all(
@@ -178,10 +45,13 @@ def run_vllm(
             "Please ensure that max_model_len is greater than the sum of"
             " prompt_len and expected_output_len for all requests.")
     # Add the requests to the engine.
-    prompts: list[TextPrompt] = []
+    prompts: list[Union[TextPrompt, TokensPrompt]] = []
     sampling_params: list[SamplingParams] = []
     for request in requests:
         prompts.append(
+            TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                       multi_modal_data=request.multi_modal_data)
+            if "prompt_token_ids" in request.prompt else \
             TextPrompt(prompt=request.prompt,
                        multi_modal_data=request.multi_modal_data))
         sampling_params.append(
@@ -191,6 +61,7 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
             ))
     lora_requests: Optional[list[LoRARequest]] = None
     if engine_args.enable_lora:
@@ -198,12 +69,13 @@ def run_vllm(
 
     use_beam_search = False
 
+    outputs = None
     if not use_beam_search:
         start = time.perf_counter()
-        llm.generate(prompts,
-                     sampling_params,
-                     lora_request=lora_requests,
-                     use_tqdm=True)
+        outputs = llm.generate(prompts,
+                               sampling_params,
+                               lora_request=lora_requests,
+                               use_tqdm=True)
         end = time.perf_counter()
     else:
         assert lora_requests is None, "BeamSearch API does not support LoRA"
@@ -221,7 +93,46 @@ def run_vllm(
                 ignore_eos=True,
             ))
         end = time.perf_counter()
-    return end - start
+    return end - start, outputs
+
+
+def run_vllm_chat(
+        requests: list[SampleRequest],
+        n: int,
+        engine_args: EngineArgs,
+        disable_detokenize: bool = False) -> tuple[float, list[RequestOutput]]:
+    """
+    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
+    multimodal models as it properly handles multimodal inputs and chat
+    formatting. For non-multimodal models, use run_vllm() instead.
+    """
+    from vllm import LLM, SamplingParams
+    llm = LLM(**dataclasses.asdict(engine_args))
+
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of "
+            "prompt_len and expected_output_len for all requests.")
+
+    prompts = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            ))
+    start = time.perf_counter()
+    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start, outputs
 
 
 async def run_vllm_async(
@@ -229,6 +140,7 @@ async def run_vllm_async(
     n: int,
     engine_args: AsyncEngineArgs,
     disable_frontend_multiprocessing: bool = False,
+    disable_detokenize: bool = False,
 ) -> float:
     from vllm import SamplingParams
 
@@ -242,11 +154,14 @@ async def run_vllm_async(
                 " prompt_len and expected_output_len for all requests.")
 
         # Add the requests to the engine.
-        prompts: list[TextPrompt] = []
+        prompts: list[Union[TextPrompt, TokensPrompt]] = []
         sampling_params: list[SamplingParams] = []
         lora_requests: list[Optional[LoRARequest]] = []
         for request in requests:
             prompts.append(
+                TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                        multi_modal_data=request.multi_modal_data)
+                if "prompt_token_ids" in request.prompt else \
                 TextPrompt(prompt=request.prompt,
                            multi_modal_data=request.multi_modal_data))
             sampling_params.append(
@@ -256,6 +171,7 @@ async def run_vllm_async(
                     top_p=1.0,
                     ignore_eos=True,
                     max_tokens=request.expected_output_len,
+                    detokenize=not disable_detokenize,
                 ))
             lora_requests.append(request.lora_request)
 
@@ -282,6 +198,7 @@ def run_hf(
     n: int,
     max_batch_size: int,
     trust_remote_code: bool,
+    disable_detokenize: bool = False,
 ) -> float:
     llm = AutoModelForCausalLM.from_pretrained(
         model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
@@ -296,14 +213,17 @@ def run_hf(
     max_prompt_len = 0
     max_output_len = 0
     for i in range(len(requests)):
-        prompt, prompt_len, output_len = requests[i]
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
         # Add the prompt to the batch.
         batch.append(prompt)
         max_prompt_len = max(max_prompt_len, prompt_len)
         max_output_len = max(max_output_len, output_len)
         if len(batch) < max_batch_size and i != len(requests) - 1:
             # Check if we can add more requests to the batch.
-            _, next_prompt_len, next_output_len = requests[i + 1]
+            next_prompt_len = requests[i + 1].prompt_len
+            next_output_len = requests[i + 1].expected_output_len
             if (max(max_prompt_len, next_prompt_len) +
                     max(max_output_len, next_output_len)) <= 2048:
                 # We can add more requests to the batch.
@@ -321,8 +241,9 @@ def run_hf(
             use_cache=True,
             max_new_tokens=max_output_len,
         )
-        # Include the decoding time.
-        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        if not disable_detokenize:
+            # Include the decoding time.
+            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
         pbar.update(len(batch))
 
         # Clear the batch.
@@ -369,58 +290,74 @@ def save_to_pytorch_benchmark_format(args: argparse.Namespace,
         write_to_json(pt_file, pt_records)
 
 
+def get_requests(args, tokenizer):
+    # Common parameters for all dataset types.
+    common_kwargs = {
+        "dataset_path": args.dataset_path,
+        "random_seed": args.seed,
+    }
+    sample_kwargs = {
+        "tokenizer": tokenizer,
+        "lora_path": args.lora_path,
+        "max_loras": args.max_loras,
+        "num_requests": args.num_prompts,
+        "input_len": args.input_len,
+        "output_len": args.output_len,
+    }
+
+    if args.dataset_path is None or args.dataset_name == "random":
+        sample_kwargs["range_ratio"] = args.random_range_ratio
+        sample_kwargs["prefix_len"] = args.prefix_len
+        dataset_cls = RandomDataset
+    elif args.dataset_name == "sharegpt":
+        dataset_cls = ShareGPTDataset
+        if args.backend == "vllm-chat":
+            sample_kwargs["enable_multimodal_chat"] = True
+    elif args.dataset_name == "sonnet":
+        assert tokenizer.chat_template or tokenizer.default_chat_template, (
+            "Tokenizer/model must have chat template for sonnet dataset.")
+        dataset_cls = SonnetDataset
+        sample_kwargs["prefix_len"] = args.prefix_len
+        sample_kwargs["return_prompt_formatted"] = True
+    elif args.dataset_name == "burstgpt":
+        dataset_cls = BurstGPTDataset
+    elif args.dataset_name == "hf":
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = VisionArenaDataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            common_kwargs['dataset_split'] = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = ConversationDataset
+            common_kwargs['dataset_subset'] = args.hf_subset
+            common_kwargs['dataset_split'] = args.hf_split
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = AIMODataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+    else:
+        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
+    # Remove None values
+    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
+    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
+
+
 def main(args: argparse.Namespace):
+    if args.seed is None:
+        args.seed = 0
     print(args)
     random.seed(args.seed)
-
     # Sample the requests.
     tokenizer = AutoTokenizer.from_pretrained(
         args.tokenizer, trust_remote_code=args.trust_remote_code)
-    if args.dataset is None:
-        vocab_size = tokenizer.vocab_size
-        requests = []
-        for _ in range(args.num_prompts):
-
-            request_tokenizer = tokenizer
-            lora_request: Optional[LoRARequest] = None
-            if args.enable_lora:
-                lora_request, lora_tokenizer = get_random_lora_request(args)
-                if lora_tokenizer:
-                    request_tokenizer = lora_tokenizer
-
-            # Synthesize a prompt with the given input length.
-            candidate_ids = [
-                random.randint(0, vocab_size - 1)
-                for _ in range(args.input_len)
-            ]
-            # As tokenizer may add additional tokens like BOS, we need to try
-            # different lengths to get the desired input length.
-            for _ in range(5):  # Max attempts to correct
-                candidate_prompt = request_tokenizer.decode(candidate_ids)
-                tokenized_len = len(request_tokenizer.encode(candidate_prompt))
-
-                if tokenized_len == args.input_len:
-                    break
-
-                # Adjust length based on difference
-                diff = args.input_len - tokenized_len
-                if diff > 0:
-                    candidate_ids.extend([
-                        random.randint(100, vocab_size - 100)
-                        for _ in range(diff)
-                    ])
-                else:
-                    candidate_ids = candidate_ids[:diff]
-            requests.append(
-                SampleRequest(prompt=candidate_prompt,
-                              prompt_len=args.input_len,
-                              expected_output_len=args.output_len,
-                              lora_request=lora_request))
-    else:
-        requests = sample_requests(tokenizer, args)
-
+    requests = get_requests(args, tokenizer)
     is_multi_modal = any(request.multi_modal_data is not None
                          for request in requests)
+    request_outputs: Optional[list[RequestOutput]] = None
     if args.backend == "vllm":
         if args.async_engine:
             elapsed_time = uvloop.run(
@@ -429,31 +366,59 @@ def main(args: argparse.Namespace):
                     args.n,
                     AsyncEngineArgs.from_cli_args(args),
                     args.disable_frontend_multiprocessing,
+                    args.disable_detokenize,
                 ))
         else:
-            elapsed_time = run_vllm(requests, args.n,
-                                    EngineArgs.from_cli_args(args))
+            elapsed_time, request_outputs = run_vllm(
+                requests, args.n, EngineArgs.from_cli_args(args),
+                args.disable_detokenize)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
-                              args.hf_max_batch_size, args.trust_remote_code)
+                              args.hf_max_batch_size, args.trust_remote_code,
+                              args.disable_detokenize)
     elif args.backend == "mii":
         elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
                                args.output_len)
+    elif args.backend == "vllm-chat":
+        elapsed_time, request_outputs = run_vllm_chat(
+            requests, args.n, EngineArgs.from_cli_args(args),
+            args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if is_multi_modal:
-        print("\033[91mWARNING\033[0m: Multi-modal request detected. The "
+
+    if request_outputs:
+        # Note: with the vllm and vllm-chat backends,
+        # we have request_outputs, which we use to count tokens.
+        total_prompt_tokens = 0
+        total_output_tokens = 0
+        for ro in request_outputs:
+            if not isinstance(ro, RequestOutput):
+                continue
+            total_prompt_tokens += len(
+                ro.prompt_token_ids) if ro.prompt_token_ids else 0
+            total_output_tokens += sum(
+                len(o.token_ids) for o in ro.outputs if o)
+        total_num_tokens = total_prompt_tokens + total_output_tokens
+    else:
+        total_num_tokens = sum(r.prompt_len + r.expected_output_len
+                               for r in requests)
+        total_output_tokens = sum(r.expected_output_len for r in requests)
+        total_prompt_tokens = total_num_tokens - total_output_tokens
+
+    if is_multi_modal and args.backend != "vllm-chat":
+        print("\033[91mWARNING\033[0m: Multi-modal request with "
+              f"{args.backend} backend detected. The "
               "following metrics are not accurate because image tokens are not"
               " counted. See vllm-project/vllm/issues/9778 for details.")
-        # TODO(vllm-project/vllm/issues/9778): Count molti-modal token length.
+        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
+        # vllm-chat backend counts the image tokens now
+
     print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
           f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
           f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
+    print(f"Total num prompt tokens:  {total_prompt_tokens}")
+    print(f"Total num output tokens:  {total_output_tokens}")
 
     # Output JSON results if specified
     if args.output_json:
@@ -469,18 +434,127 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, results)
 
 
+def validate_args(args):
+    """
+    Validate command-line arguments.
+    """
+
+    # === Deprecation and Defaulting ===
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next release. "
+            "Please use '--dataset-name' and '--dataset-path' instead.",
+            stacklevel=2)
+        args.dataset_path = args.dataset
+
+    if not getattr(args, "tokenizer", None):
+        args.tokenizer = args.model
+
+    # === Backend Validation ===
+    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
+    if args.backend not in valid_backends:
+        raise ValueError(f"Unsupported backend: {args.backend}")
+
+    # === Dataset Configuration ===
+    if not args.dataset and not args.dataset_path:
+        print(
+            "When dataset path is not set, it will default to random dataset")
+        args.dataset_name = 'random'
+        if args.input_len is None:
+            raise ValueError("input_len must be provided for a random dataset")
+
+    # === Dataset Name Specific Checks ===
+    # --hf-subset and --hf-split: only used
+    # when dataset_name is 'hf'
+    if args.dataset_name != "hf" and (
+            getattr(args, "hf_subset", None) is not None
+            or getattr(args, "hf_split", None) is not None):
+        warnings.warn("--hf-subset and --hf-split will be ignored \
+                since --dataset-name is not 'hf'.",
+                      stacklevel=2)
+    elif args.dataset_name == "hf":
+        if args.dataset_path in (
+                VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+                | ConversationDataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm-chat", f"{args.dataset_path} needs to use vllm-chat as the backend."  #noqa: E501
+        elif args.dataset_path in (InstructCoderDataset.SUPPORTED_DATASET_PATHS
+                                   | AIMODataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm", f"{args.dataset_path} needs to use vllm as the backend."  #noqa: E501
+        else:
+            raise ValueError(
+                f"{args.dataset_path} is not supported by hf dataset.")
+
+    # --random-range-ratio: only used when dataset_name is 'random'
+    if args.dataset_name != 'random' and args.random_range_ratio is not None:
+        warnings.warn("--random-range-ratio will be ignored since \
+                --dataset-name is not 'random'.",
+                      stacklevel=2)
+
+    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
+    # set.
+    if args.dataset_name not in {"random", "sonnet", None
+                                 } and args.prefix_len is not None:
+        warnings.warn("--prefix-len will be ignored since --dataset-name\
+                 is not 'random', 'sonnet', or not set.",
+                      stacklevel=2)
+
+    # === LoRA Settings ===
+    if getattr(args, "enable_lora", False) and args.backend != "vllm":
+        raise ValueError(
+            "LoRA benchmarking is only supported for vLLM backend")
+    if getattr(args, "enable_lora", False) and args.lora_path is None:
+        raise ValueError("LoRA path must be provided when enable_lora is True")
+
+    # === Backend-specific Validations ===
+    if args.backend == "hf" and args.hf_max_batch_size is None:
+        raise ValueError("HF max batch size is required for HF backend")
+    if args.backend != "hf" and args.hf_max_batch_size is not None:
+        raise ValueError("HF max batch size is only for HF backend.")
+
+    if args.backend in {"hf", "mii"} and getattr(args, "quantization",
+                                                 None) is not None:
+        raise ValueError("Quantization is only for vLLM backend.")
+
+    if args.backend == "mii" and args.dtype != "auto":
+        raise ValueError("dtype must be auto for MII backend.")
+    if args.backend == "mii" and args.n != 1:
+        raise ValueError("n must be 1 for MII backend.")
+    if args.backend == "mii" and args.tokenizer != args.model:
+        raise ValueError(
+            "Tokenizer must be the same as the model for MII backend.")
+
+    # --data-parallel is not supported currently.
+    # https://github.com/vllm-project/vllm/issues/16222
+    if args.data_parallel_size > 1:
+        raise ValueError(
+            "Data parallel is not supported in offline benchmark, \
+            please use benchmark serving instead")
+
+
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument("--backend",
                         type=str,
-                        choices=["vllm", "hf", "mii"],
+                        choices=["vllm", "hf", "mii", "vllm-chat"],
                         default="vllm")
-    parser.add_argument("--dataset",
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
+        help="Name of the dataset to benchmark on.",
+        default="sharegpt")
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in\
+            the next release. The dataset is expected to "
+        "be a json in form of list[dict[..., conversations: "
+        "list[dict[..., value: <prompt_or_response>]]]]")
+    parser.add_argument("--dataset-path",
                         type=str,
                         default=None,
-                        help="Path to the dataset. The dataset is expected to "
-                        "be a json in form of list[dict[..., conversations: "
-                        "list[dict[..., value: <prompt_or_response>]]]]")
+                        help="Path to the dataset")
     parser.add_argument("--input-len",
                         type=int,
                         default=None,
@@ -515,6 +589,11 @@ if __name__ == "__main__":
                         action='store_true',
                         default=False,
                         help="Disable decoupled async engine frontend.")
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize the response (i.e. do not include "
+              "detokenization time in the measurement)"))
     # LoRA
     parser.add_argument(
         "--lora-path",
@@ -522,43 +601,45 @@ if __name__ == "__main__":
         default=None,
         help="Path to the lora adapters to use. This can be an absolute path, "
         "a relative path, or a Hugging Face model identifier.")
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
+        "controls how much of the input is fixed lines versus "
+        "random lines, but the total input length remains approximately "
+        "input_len tokens.")
+    # random dataset
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=None,
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
+    )
+
+    # hf dtaset
+    parser.add_argument("--hf-subset",
+                        type=str,
+                        default=None,
+                        help="Subset of the HF dataset.")
+    parser.add_argument("--hf-split",
+                        type=str,
+                        default=None,
+                        help="Split of the HF dataset.")
 
     parser = AsyncEngineArgs.add_cli_args(parser)
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model
-    if args.dataset is None:
-        assert args.input_len is not None
-        assert args.output_len is not None
-    else:
-        assert args.input_len is None
-    if args.enable_lora:
-        assert args.lora_path is not None
-
-    if args.backend == "vllm":
-        if args.hf_max_batch_size is not None:
-            raise ValueError("HF max batch size is only for HF backend.")
-    elif args.backend == "hf":
-        if args.hf_max_batch_size is None:
-            raise ValueError("HF max batch size is required for HF backend.")
-        if args.quantization is not None:
-            raise ValueError("Quantization is only for vLLM backend.")
-        if args.enable_lora is not None:
-            raise ValueError("LoRA benchmarking is only supported for vLLM"
-                             " backend")
-    elif args.backend == "mii":
-        if args.dtype != "auto":
-            raise ValueError("dtype must be auto for MII backend.")
-        if args.n != 1:
-            raise ValueError("n must be 1 for MII backend.")
-        if args.quantization is not None:
-            raise ValueError("Quantization is only for vLLM backend.")
-        if args.hf_max_batch_size is not None:
-            raise ValueError("HF max batch size is only for HF backend.")
-        if args.tokenizer != args.model:
-            raise ValueError("Tokenizer must be the same as the model for MII "
-                             "backend.")
-        if args.enable_lora is not None:
-            raise ValueError("LoRA benchmarking is only supported for vLLM"
-                             " backend")
+    validate_args(args)
     main(args)

benchmarks/kernels/benchmark_bitblas.py

@@ -0,0 +1,236 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+from vllm.model_executor.layers.quantization.utils.bitblas_utils import (
+    MINIMUM_BITBLAS_VERSION)
+
+try:
+    import bitblas
+    if bitblas.__version__ < MINIMUM_BITBLAS_VERSION:
+        raise ImportError("bitblas version is wrong. Please "
+                          f"install bitblas>={MINIMUM_BITBLAS_VERSION}")
+except ImportError as e:
+    bitblas_import_exception = e
+    raise ValueError("Trying to use the bitblas backend, but could not import"
+                     f"with the following error: {bitblas_import_exception}. "
+                     "Please install bitblas through the following command: "
+                     f"`pip install bitblas>={MINIMUM_BITBLAS_VERSION}`"
+                     ) from bitblas_import_exception
+
+from bitblas import Matmul, MatmulConfig, auto_detect_nvidia_target
+
+from vllm.utils import FlexibleArgumentParser
+
+parser = FlexibleArgumentParser(
+    description="Benchmark BitBLAS int4 on a specific target.")
+
+# Add arguments to the parser
+parser.add_argument(
+    "--target",
+    type=str,
+    default=auto_detect_nvidia_target(),
+    help="Specify the target device for benchmarking.",
+)
+parser.add_argument("--group_size",
+                    type=int,
+                    default=None,
+                    help="Group size for grouped quantization.")
+parser.add_argument(
+    "--A_dtype",
+    type=str,
+    default="float16",
+    choices=["float16", "float32", "float64", "int32", "int8"],
+    help="Data type of activation A.",
+)
+parser.add_argument(
+    "--W_dtype",
+    type=str,
+    default="int4",
+    choices=[
+        "float16",
+        "float32",
+        "float64",
+        "int32",
+        "int8",
+        "int4",
+        "int2",
+        "int1",
+        "nf4",
+        "fp4_e2m1",
+    ],
+    help="Data type of weight W.",
+)
+parser.add_argument(
+    "--accum_dtype",
+    type=str,
+    default="float16",
+    choices=["float16", "int32"],
+    help="Data type for accumulation.",
+)
+parser.add_argument(
+    "--out_dtype",
+    type=str,
+    default="float16",
+    choices=["float16", "float32", "int32", "int8"],
+    help="Data type for output.",
+)
+parser.add_argument(
+    "--layout",
+    type=str,
+    default="nt",
+    choices=["nt", "nn"],
+    help="Matrix layout, 'nt' for non-transpose A and transpose W.",
+)
+parser.add_argument("--with_bias",
+                    action="store_true",
+                    help="Include bias in the benchmark.")
+parser.add_argument(
+    "--with_scaling",
+    action="store_true",
+    help="Include scaling factor in the quantization.",
+)
+parser.add_argument("--with_zeros",
+                    action="store_true",
+                    help="Include zeros in the quantization.")
+parser.add_argument(
+    "--zeros_mode",
+    type=str,
+    default=None,
+    choices=["original", "rescale", "quantized"],
+    help="Specify the mode for calculating zeros.",
+)
+
+# Parse the arguments
+args = parser.parse_args()
+
+# Assign arguments to variables
+target = args.target
+A_dtype = args.A_dtype
+W_dtype = args.W_dtype
+accum_dtype = args.accum_dtype
+out_dtype = args.out_dtype
+layout = args.layout
+with_bias = args.with_bias
+group_size = args.group_size
+with_scaling = args.with_scaling
+with_zeros = args.with_zeros
+zeros_mode = args.zeros_mode
+
+# Define a list of shared arguments that repeat in every config
+shared_args = [
+    A_dtype,
+    W_dtype,
+    out_dtype,
+    accum_dtype,
+    layout,
+    with_bias,
+    group_size,
+    with_scaling,
+    with_zeros,
+    zeros_mode,
+]
+
+# Define just the (M, K, N) shapes in a more compact list
+shapes = [
+    # square test
+    (1, 16384, 16384),
+    # BLOOM-176B
+    (1, 43008, 14336),
+    (1, 14336, 14336),
+    (1, 57344, 14336),
+    (1, 14336, 57344),
+    # OPT-65B
+    (1, 9216, 9216),
+    (1, 36864, 9216),
+    (1, 9216, 36864),
+    (1, 22016, 8192),
+    # LLAMA-70B/65B
+    (1, 8192, 22016),
+    (1, 8192, 8192),
+    (1, 28672, 8192),
+    (1, 8192, 28672),
+    # square test
+    (16384, 16384, 16384),
+    # BLOOM-176B
+    (8192, 43008, 14336),
+    (8192, 14336, 14336),
+    (8192, 57344, 14336),
+    (8192, 14336, 57344),
+    # OPT-65B
+    (8192, 9216, 9216),
+    (8192, 36864, 9216),
+    (8192, 9216, 36864),
+    (8192, 22016, 8192),
+    # LLAMA-70B/65B
+    (8192, 8192, 22016),
+    (8192, 8192, 8192),
+    (8192, 28672, 8192),
+    (8192, 8192, 28672),
+]
+
+# Build test shapes with all the shared arguments
+test_shapes = [(MatmulConfig, Matmul, (*shape, *shared_args))
+               for shape in shapes]
+
+benchmark_sets = []
+benchmark_sets.extend(test_shapes)
+
+benchmark_results = {}
+for config_class, operator, input_args in benchmark_sets:
+    config = config_class(*input_args)
+    matmul = operator(config, target=target, enable_tuning=True)
+    kernel_latency = matmul.profile_latency()
+
+    print("Time cost is: {:.3f} ms".format(kernel_latency))
+
+    profile_config = {
+        f"{operator.__name__}-{'-'.join([str(i) for i in input_args])}": {
+            "BitBLAS_top20_latency": kernel_latency,
+        }
+    }
+
+    benchmark_results.update(profile_config)
+
+# Define headers for the table
+headers = [
+    "PrimFunc",
+    "Input Arguments",
+    "BitBLAS Top20 Latency",
+]
+
+# Calculate column widths for pretty printing
+col_widths = [0, 0, 0]
+for config_key, values in benchmark_results.items():
+    args_split = config_key.split("-")
+    func_name = args_split[0]
+    input_args_str = "-".join(args_split[1:])
+    col_widths[0] = max(col_widths[0], len(func_name) + 2, len(headers[0]) + 2)
+    col_widths[1] = max(col_widths[1],
+                        len(input_args_str) + 2,
+                        len(headers[1]) + 2)
+    col_widths[2] = max(col_widths[2],
+                        len(f"{values['BitBLAS_top20_latency']:.3f} ms") + 2,
+                        len(headers[2]) + 2)
+    # break only if you want to measure widths from a single example;
+    # otherwise, let it loop over all items.
+
+# Print header
+for i, header in enumerate(headers):
+    headers[i] = header.ljust(col_widths[i])
+print("".join(headers))
+print("-" * sum(col_widths))
+
+# Print rows
+for config_key, values in benchmark_results.items():
+    args_split = config_key.split("-")
+    func_name = args_split[0]
+    input_args_str = "-".join(args_split[1:])
+    row = [
+        func_name,
+        input_args_str,
+        f"{values['BitBLAS_top20_latency']:.3f} ms",
+    ]
+    row_str = "".join(
+        [str(cell).ljust(col_widths[idx]) for idx, cell in enumerate(row)])
+    print(row_str)

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -0,0 +1,341 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.utils.benchmark as benchmark
+from benchmark_shapes import WEIGHT_SHAPES_MOE
+
+from vllm import _custom_ops as ops
+from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
+from vllm.model_executor.layers.fused_moe.fused_moe import (cutlass_moe_fp8,
+                                                            fused_experts,
+                                                            fused_topk)
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1", "nm-testing/deepseekv2-lite",
+    "ibm-granite/granite-3.0-1b-a400m", "ibm-granite/granite-3.0-3b-a800m"
+]
+DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False]
+PER_OUT_CH_OPTS = [False]
+
+
+def to_fp8(tensor: torch.Tensor):
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def bench_run(results: list[benchmark.Measurement], model: str,
+              num_experts: int, topk: int, per_act_token: bool,
+              per_out_ch: bool, mkn: tuple[int, int, int]):
+    label = "Quant Matmul"
+
+    sub_label = (
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, "
+        "MKN=({})".format(model, num_experts, topk, per_act_token, per_out_ch,
+                          mkn))
+
+    print(f"Testing: {sub_label}")
+
+    (m, k, n) = mkn
+
+    dtype = torch.half
+
+    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
+    w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10
+
+    _, a_scale = ops.scaled_fp8_quant(a)
+
+    w1_q = torch.empty((num_experts, 2 * n, k),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w2_q = torch.empty((num_experts, k, n),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w1_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+    w2_scale = torch.empty((num_experts, 1, 1),
+                           device="cuda",
+                           dtype=torch.float32)
+
+    ab_strides1 = torch.full((num_experts, ),
+                             k,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides1 = torch.full((num_experts, ),
+                            2 * n,
+                            device="cuda",
+                            dtype=torch.int64)
+    ab_strides2 = torch.full((num_experts, ),
+                             n,
+                             device="cuda",
+                             dtype=torch.int64)
+    c_strides2 = torch.full((num_experts, ),
+                            k,
+                            device="cuda",
+                            dtype=torch.int64)
+
+    for expert in range(num_experts):
+        w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
+        w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
+    w1_q_notransp = w1_q.clone()
+    w2_q_notransp = w2_q.clone()
+    w1_q = w1_q.transpose(1, 2)
+    w2_q = w2_q.transpose(1, 2)
+
+    score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
+
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        a, score, topk, renormalize=False)
+
+    def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
+                       topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                       w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                       a_scale: torch.Tensor, num_repeats: int):
+        for _ in range(num_repeats):
+            fused_experts(a,
+                          w1,
+                          w2,
+                          topk_weights,
+                          topk_ids,
+                          use_fp8_w8a8=True,
+                          w1_scale=w1_scale,
+                          w2_scale=w2_scale,
+                          a1_scale=a_scale)
+
+    def run_cutlass_moe(a: torch.Tensor, a_scale: torch.Tensor,
+                        w1: torch.Tensor, w2: torch.Tensor,
+                        w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                        topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                        ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+                        ab_strides2: torch.Tensor, c_strides2: torch.Tensor,
+                        num_repeats: int):
+        for _ in range(num_repeats):
+            cutlass_moe_fp8(a,
+                            w1,
+                            w2,
+                            w1_scale,
+                            w2_scale,
+                            topk_weights,
+                            topk_ids,
+                            ab_strides1,
+                            c_strides1,
+                            ab_strides2,
+                            c_strides2,
+                            a1_scale=a_scale)
+
+    def run_cutlass_from_graph(
+            a: torch.Tensor, a_scale: torch.Tensor, w1_q: torch.Tensor,
+            w2_q: torch.Tensor, w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+            ab_strides1: torch.Tensor, c_strides1: torch.Tensor,
+            ab_strides2: torch.Tensor, c_strides2: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return cutlass_moe_fp8(a,
+                                   w1_q,
+                                   w2_q,
+                                   w1_scale,
+                                   w2_scale,
+                                   topk_weights,
+                                   topk_ids,
+                                   ab_strides1,
+                                   c_strides1,
+                                   ab_strides2,
+                                   c_strides2,
+                                   a1_scale=a_scale)
+
+    def run_triton_from_graph(a: torch.Tensor, w1: torch.Tensor,
+                              w2: torch.Tensor, topk_weights: torch.Tensor,
+                              topk_ids: torch.Tensor, w1_scale: torch.Tensor,
+                              w2_scale: torch.Tensor, a_scale: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return fused_experts(a,
+                                 w1,
+                                 w2,
+                                 topk_weights,
+                                 topk_ids,
+                                 use_fp8_w8a8=True,
+                                 w1_scale=w1_scale,
+                                 w2_scale=w2_scale,
+                                 a1_scale=a_scale)
+
+    def replay_graph(graph, num_repeats):
+        for _ in range(num_repeats):
+            graph.replay()
+        torch.cuda.synchronize()
+
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        run_cutlass_from_graph(a, a_scale, w1_q, w2_q, w1_scale, w2_scale,
+                               topk_weights, topk_ids, ab_strides1, c_strides1,
+                               ab_strides2, c_strides2)
+    torch.cuda.synchronize()
+
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        run_triton_from_graph(a, w1_q_notransp, w2_q_notransp, topk_weights,
+                              topk_ids, w1_scale, w2_scale, a_scale)
+    torch.cuda.synchronize()
+
+    min_run_time = 5
+    num_warmup = 5
+    num_runs = 25
+
+    globals = {
+        # Baseline params
+        "w1": w1,
+        "w2": w2,
+        "score": score,
+        "topk": topk,
+        "w1_q_notransp": w1_q_notransp,
+        "w2_q_notransp": w2_q_notransp,
+        # Cutlass params
+        "a_scale": a_scale,
+        "w1_q": w1_q,
+        "w2_q": w2_q,
+        "w1_scale": w1_scale,
+        "w2_scale": w2_scale,
+        "ab_strides1": ab_strides1,
+        "c_strides1": c_strides1,
+        "ab_strides2": ab_strides2,
+        "c_strides2": c_strides2,
+        # cuda graph params
+        "cutlass_graph": cutlass_graph,
+        "triton_graph": triton_graph,
+        # Gen params
+        "a": a,
+        "topk_weights": topk_weights,
+        "topk_ids": topk_ids,
+        "num_runs": num_runs,
+        # Kernels
+        "run_triton_moe": run_triton_moe,
+        "run_cutlass_moe": run_cutlass_moe,
+        "replay_graph": replay_graph,
+    }
+
+    # Warmup
+    run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids,
+                   w1_scale, w2_scale, a_scale, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(triton_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(triton_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights,
+                    topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2,
+                    num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(cutlass_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(cutlass_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+
+def main(args):
+    print("Benchmarking models:")
+    for i, model in enumerate(args.models):
+        print(f"[{i}]  {model}")
+
+    results: list[benchmark.Measurement] = []
+
+    for model in args.models:
+        for tp in args.tp_sizes:
+            for layer in WEIGHT_SHAPES_MOE[model]:
+                num_experts = layer[0]
+                topk = layer[1]
+                size_k = layer[2]
+                size_n = layer[3] // tp
+
+                if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                    continue
+
+                if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                    continue
+
+                for per_act_token in PER_ACT_TOKEN_OPTS:
+                    for per_out_ch in PER_OUT_CH_OPTS:
+                        for size_m in DEFAULT_BATCH_SIZES:
+                            mkn = (size_m, size_k, size_n)
+                            bench_run(results, model, num_experts, topk,
+                                      per_act_token, per_out_ch, mkn)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES_MOE.keys(),
+    )
+    parser.add_argument("--tp-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_TP_SIZES)
+    parser.add_argument("--batch-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-act-token",
+                        nargs="+",
+                        type=int,
+                        default=[])
+    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_layernorm.py

@@ -40,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_lora.py

@@ -17,12 +17,14 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand
-from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice
-from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink
-from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand
-from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink
-from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
+from vllm.triton_utils import HAS_TRITON
+
+if HAS_TRITON:
+    from vllm.lora.ops.triton_ops import (LoRAKernelMeta, lora_expand,
+                                          lora_shrink)
+    from vllm.lora.ops.triton_ops.utils import (_LORA_A_PTR_DICT,
+                                                _LORA_B_PTR_DICT)
+
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
@@ -153,7 +155,6 @@ def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor,
         result = torch.nn.functional.linear(x, w)
         result *= scaling
         out_list.append(result)
-    torch.cat(out_list, dim=0)
 
     cat_result = torch.cat(out_list, dim=0)
 
@@ -167,52 +168,25 @@ class OpType(Enum):
     """
     LoRA Ops to benchmark and its properties.
     """
-    SGMV_SHRINK = auto()
-    BGMV_SHRINK = auto()
-    SGMV_EXPAND = auto()
-    BGMV_EXPAND = auto()
-    BGMV_EXPAND_SLICE = auto()
+    LORA_SHRINK = auto()
+    LORA_EXPAND = auto()
 
     @staticmethod
     def from_str(s: str) -> "OpType":
-        if s.lower() == 'sgmv_shrink':
-            return OpType.SGMV_SHRINK
-        if s.lower() == 'sgmv_expand':
-            return OpType.SGMV_EXPAND
-        if s.lower() == 'bgmv_shrink':
-            return OpType.BGMV_SHRINK
-        if s.lower() == 'bgmv_expand':
-            return OpType.BGMV_EXPAND
-        if s.lower() == "bgmv_expand_slice":
-            return OpType.BGMV_EXPAND_SLICE
+        if s.lower() == "lora_shrink":
+            return OpType.LORA_SHRINK
+        if s.lower() == "lora_expand":
+            return OpType.LORA_EXPAND
         raise ValueError(f"Unrecognized str {s} to convert to OpType")
 
     def is_shrink_fn(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.BGMV_SHRINK]
+        return self in [OpType.LORA_SHRINK]
 
     def is_expand_fn(self) -> bool:
-        return self in [OpType.SGMV_EXPAND, OpType.BGMV_EXPAND]
-
-    def is_prefill_op(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.SGMV_EXPAND]
-
-    def is_decode_op(self) -> bool:
-        return self in [
-            OpType.BGMV_SHRINK, OpType.BGMV_EXPAND, OpType.BGMV_EXPAND_SLICE
-        ]
-
-    def is_expand_slice_fn(self) -> bool:
-        return self in [OpType.BGMV_EXPAND_SLICE]
+        return self in [OpType.LORA_EXPAND]
 
     def num_slices(self) -> list[int]:
-        if self in [OpType.SGMV_EXPAND, OpType.SGMV_SHRINK]:
-            # SGMV kernels supports slices
-            return [1, 2, 3]
-        if self in [OpType.BGMV_SHRINK, OpType.BGMV_EXPAND]:
-            return [1]
-        if self in [OpType.BGMV_EXPAND_SLICE]:
-            return [2, 3]
-        raise ValueError(f"Unrecognized OpType {self}")
+        return [1, 2, 3]
 
     def mkn(self, batch_size: int, seq_length: int, hidden_size: int,
             lora_rank: int) -> tuple[int, int, int]:
@@ -222,7 +196,7 @@ class OpType(Enum):
             k = hidden_size
             n = lora_rank
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             m = num_tokens
             k = lora_rank
             n = hidden_size
@@ -237,7 +211,7 @@ class OpType(Enum):
         if self.is_shrink_fn():
             return op_dtype, op_dtype, torch.float32
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             return torch.float32, op_dtype, op_dtype
 
     def matmul_shapes(
@@ -251,56 +225,39 @@ class OpType(Enum):
         m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank)
 
         b_shape = (num_loras, n, k)  # col-major
-        if self == OpType.SGMV_SHRINK:
-            # SGMV shrink supports num_slices inherently in the kernel
+        if self in [OpType.LORA_SHRINK]:
+            # LoRA shrink kernels support num_slices inherently in the kernel.
             return ((m, k), b_shape, (num_slices, m, n))
-        if self == OpType.SGMV_EXPAND:
-            # SGMV expand supports num_slices inherently in the kernel
+        if self in [OpType.LORA_EXPAND]:
+            # LoRA expand kernels support num_slices inherently in the kernel
             return ((num_slices, m, k), b_shape, (m, n * num_slices))
-        if self == OpType.BGMV_SHRINK:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return ((num_slices, m, k), b_shape, (m, n * num_slices))
-
         raise ValueError(f"Unrecognized op_type {self}")
 
     def bench_fn(self) -> Callable:
+        if self == OpType.LORA_SHRINK:
+            return lora_shrink
+        if self == OpType.LORA_EXPAND:
+            return lora_expand
 
-        def emulate_bgmv_expand_slice(kwargs_list: list[dict[str, Any]]):
-            for x in kwargs_list:
-                bgmv_expand_slice(**x)
-
-        if self == OpType.SGMV_SHRINK:
-            return sgmv_shrink
-        if self == OpType.SGMV_EXPAND:
-            return sgmv_expand
-        if self == OpType.BGMV_SHRINK:
-            return bgmv_shrink
-        if self == OpType.BGMV_EXPAND:
-            return bgmv_expand
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return emulate_bgmv_expand_slice
         raise ValueError(f"Unrecognized optype {self}")
 
     def run_ref_group_gemm(self, output: torch.Tensor, input: torch.Tensor,
                            lora_weights: list[torch.Tensor],
                            **kwargs) -> Callable:
-        """Each benchmark operation expected the input, lora_weights and outputs
+        """Each benchmark operation expects the input, lora_weights and outputs
            in a slightly different format. Refer to self.matmul_shapes().
            run_ref_group_gemm accounts for those differences in executing a
            reference group gemm for correctness testing.
         """
         w_dtype = lora_weights[0].dtype
         num_slices = len(lora_weights)
-        if self == OpType.SGMV_SHRINK:
+        if self in [OpType.LORA_SHRINK]:
             for slice_idx in range(num_slices):
                 ref_group_gemm(ref_out=output[slice_idx, :],
                                input=input,
                                lora_weights=lora_weights[slice_idx],
                                **kwargs)
-        if self == OpType.SGMV_EXPAND:
+        elif self in [OpType.LORA_EXPAND]:
             hidden_size = lora_weights[0].shape[1]
             for slice_idx in range(num_slices):
                 slice_offset = slice_idx * hidden_size
@@ -309,28 +266,8 @@ class OpType(Enum):
                     input=input[slice_idx].clone().to(dtype=w_dtype),
                     lora_weights=lora_weights[slice_idx],
                     **kwargs)
-        if self == OpType.BGMV_SHRINK:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input,
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input.clone().to(dtype=w_dtype),
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND_SLICE:
-            hidden_size = lora_weights[0].shape[1]
-            for slice_idx in range(num_slices):
-                slice_offset = slice_idx * hidden_size
-                ref_group_gemm(
-                    ref_out=output[:, slice_offset:slice_offset + hidden_size],
-                    input=input[slice_idx].clone().to(dtype=w_dtype),
-                    lora_weights=lora_weights[slice_idx],
-                    **kwargs)
-        raise ValueError(f"Unrecognized optype {self}")
+        else:
+            raise ValueError(f"Unrecognized optype {self}")
 
 
 @dataclass
@@ -386,11 +323,11 @@ class BenchmarkTensors:
     input: torch.Tensor
     lora_weights_lst: list[torch.Tensor]
     output: torch.Tensor
-    # metadata tensors
+    # LoRA kernel metadata
+    lora_kernel_meta: LoRAKernelMeta
+    # Metadata tensors used in testing correctness
     seq_lens: torch.Tensor
-    seq_start_loc: torch.Tensor
     prompt_lora_mapping: torch.Tensor
-    token_lora_mapping: torch.Tensor
 
     def io_types(self) -> str:
         return (f"{dtype_to_str(self.input.dtype)}x"
@@ -417,26 +354,29 @@ class BenchmarkTensors:
         assert ctx.num_active_loras <= ctx.num_loras
         total_tokens = ctx.batch_size * ctx.seq_length
 
+        # Make metadata tensors involved in correctness testing.
         # Prepare seq lens tensor
         seq_len_tensor = torch.randint(ctx.seq_length, ctx.seq_length + 1,
                                        (ctx.batch_size, ))
-        # Prepare seq_start_loc tensor
-        seq_start_loc_tensor = torch.cumsum(torch.tensor(
-            [0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
-                                            dim=0)
         assert total_tokens == seq_len_tensor.sum()
         # Prepare prompt lora indices tensor
         prompt_lora_indices_tensor = make_prompt_lora_mapping(
             ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu")
-        # Prepare token lora indices tensor
+
+        # Make LoRAKernelMeta
         token_lora_indices_tensor = make_token_lora_mapping(
             total_tokens, ctx.batch_size, prompt_lora_indices_tensor,
             seq_len_tensor, "cpu")
+        lora_kernel_meta = LoRAKernelMeta.make(
+            max_loras=ctx.num_loras,
+            max_num_tokens=token_lora_indices_tensor.size(0),
+            device="cpu")
+        lora_kernel_meta.prepare_tensors(
+            token_lora_mapping=token_lora_indices_tensor)
 
         return BenchmarkTensors(input_tensor, lora_weights, output_tensor,
-                                seq_len_tensor, seq_start_loc_tensor,
-                                prompt_lora_indices_tensor,
-                                token_lora_indices_tensor)
+                                lora_kernel_meta, seq_len_tensor,
+                                prompt_lora_indices_tensor)
 
     def sanity_check(self) -> None:
         """
@@ -446,9 +386,9 @@ class BenchmarkTensors:
         # check metadata tensors
         assert torch.sum(self.seq_lens) == num_tokens
         num_seqs = self.seq_lens.shape[0]
-        assert self.seq_start_loc.shape[0] == num_seqs
+        #assert self.seq_start_loc.shape[0] == num_seqs
         assert self.prompt_lora_mapping.shape[0] == num_seqs
-        assert self.token_lora_mapping.shape[0] == num_tokens
+        assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens
 
     def to_device(self, device: str):
         """
@@ -463,54 +403,31 @@ class BenchmarkTensors:
         self.input = to_device(self.input)
         self.output = to_device(self.output)
         self.seq_lens = to_device(self.seq_lens)
-        self.seq_start_loc = to_device(self.seq_start_loc)
         self.prompt_lora_mapping = to_device(self.prompt_lora_mapping)
-        self.token_lora_mapping = to_device(self.token_lora_mapping)
         for i in range(len(self.lora_weights_lst)):
             self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i])
 
+        # LoRA meta
+        for field_name in LoRAKernelMeta.__dataclass_fields__:
+            field = getattr(self.lora_kernel_meta, field_name)
+            assert isinstance(field, torch.Tensor)
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
+
     def metadata(self) -> tuple[int, int, int]:
         """
         Return num_seqs, num_tokens and max_seq_len
         """
         num_seqs = self.seq_lens.shape[0]
-        num_tokens = self.token_lora_mapping.shape[0]
+        num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0]
         max_seq_len = torch.max(self.seq_lens).item()
         num_slices = len(self.lora_weights_lst)
         return num_seqs, num_tokens, max_seq_len, num_slices
 
-    def convert_to_sgmv_benchmark_tensors(self):
-        """
-        For sgmv punica kernels, when consecutive sequences have the
-        same LoRA ID, we just merge them together.
-        This happens in punica.py::compute_metadata
-        """
-
-        # Collapse seq_lens and seq_start_loc
-        _, seq_lens = torch.unique_consecutive(self.token_lora_mapping,
-                                               return_counts=True)
-        cum_result = torch.cumsum(seq_lens, dim=0)
-        seq_start_loc = torch.zeros_like(seq_lens)
-        seq_start_loc[1:].copy_(cum_result[:-1])
-
-        # Collapse prompt mapping
-        prompt_lora_mapping = torch.unique_consecutive(
-            self.prompt_lora_mapping)
-
-        assert torch.sum(seq_lens) == torch.sum(self.seq_lens), \
-         f"dont match - new {torch.sum(seq_lens)} vs {torch.sum(self.seq_lens)}"
-
-        self.prompt_lora_mapping = prompt_lora_mapping.to(
-            dtype=self.prompt_lora_mapping.dtype)
-        self.seq_lens = seq_lens.to(dtype=self.seq_lens.dtype)
-        self.seq_start_loc = seq_start_loc.to(dtype=self.seq_start_loc.dtype)
-
-    def as_sgmv_shrink_kwargs(self) -> dict[str, Any]:
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_shrink_kwargs(self) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -531,22 +448,20 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_a_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'scaling': 1.0,
         }
 
-    def as_sgmv_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
-
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -568,106 +483,16 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_b_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'offset_start': 0,
             'add_inputs': add_inputs,
         }
 
-    def as_bgmv_shrink_kwargs(self) -> dict[str, Any]:
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, hidden_size]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [num_loras, lora_rank, hidden_size]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == hidden_size
-        lora_rank = lw_shape[1]
-        # Expected output shape [num_tokens, lora_rank]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, lora_rank)
-
-        return {
-            'inputs': self.input,
-            'lora_a_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'scaling': 1.0
-        }
-
-    def as_bgmv_expand_kwargs(self, add_inputs: bool):
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, lora_rank]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        lora_rank = i_shape[1]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size)
-
-        return {
-            'inputs': self.input,
-            'lora_b_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'add_inputs': add_inputs
-        }
-
-    def as_bgmv_expand_slice_kwargs(self, add_inputs: bool) -> dict[str, Any]:
-
-        _, num_tokens, _, num_slices = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_slices, num_tokens, lora_rank]
-        assert len(i_shape) == 3
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[2]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size * num_slices]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size * num_slices)
-
-        self.to_device(self.input.device)
-
-        kwargs_list = []
-        for i in range(num_slices):
-            kwargs_list.append({
-                'inputs': self.input[i],
-                'lora_b_weights': self.lora_weights_lst[i],
-                'output_tensor': self.output,
-                'lora_indices_tensor': self.token_lora_mapping,
-                'slice_offset': i * hidden_size,
-                'slice_size': hidden_size,
-                'add_inputs': add_inputs,
-            })
-        return {'kwargs_list': kwargs_list}
-
     def bench_fn_kwargs(self,
                         op_type: OpType,
                         add_inputs: Optional[bool] = None) -> dict[str, Any]:
@@ -676,16 +501,10 @@ class BenchmarkTensors:
         else:
             assert add_inputs is not None
 
-        if op_type == OpType.SGMV_SHRINK:
-            return self.as_sgmv_shrink_kwargs()
-        if op_type == OpType.SGMV_EXPAND:
-            return self.as_sgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_SHRINK:
-            return self.as_bgmv_shrink_kwargs()
-        if op_type == OpType.BGMV_EXPAND:
-            return self.as_bgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_EXPAND_SLICE:
-            return self.as_bgmv_expand_slice_kwargs(add_inputs)
+        if op_type == OpType.LORA_SHRINK:
+            return self.as_lora_shrink_kwargs()
+        if op_type == OpType.LORA_EXPAND:
+            return self.as_lora_expand_kwargs(add_inputs)
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(self, op_type: OpType,
@@ -873,14 +692,7 @@ def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]):
     timers = []
     for bench_ctx in bench_ctxs:
         for seq_len in args.seq_lengths:
-            bench_ops: list[OpType] = []
-            if seq_len == 1:
-                # bench all decode ops
-                bench_ops = [op for op in args.op_types if op.is_decode_op()]
-            else:
-                # bench all prefill ops
-                bench_ops = [op for op in args.op_types if op.is_prefill_op()]
-
+            bench_ops: list[OpType] = args.op_types
             seq_len_timers = []
             for bench_op in bench_ops:
                 for num_slices in bench_op.num_slices():
@@ -1090,13 +902,13 @@ Benchmark LoRA kernels:
     {use_cuda_graph_recommendation()}
 
     list_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
+        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
 
     model_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
+        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
 
     range_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
+        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
             """,  # noqa: E501
         formatter_class=argparse.RawTextHelpFormatter)
 

benchmarks/kernels/benchmark_machete.py

@@ -45,7 +45,6 @@ def terse_type_name(dt):
         torch.float16: "fp16",
         torch.int8: "int8",
         torch.float8_e4m3fn: "fp8",
-        torch.bfloat16: "bf16",
         torch.float: "float",
         torch.int: "int",
     }[dt]
@@ -259,7 +258,7 @@ def machete_create_bench_fn(bt: BenchmarkTensors,
 
     return lambda: ops.machete_mm(
         a=bt.a,
-        b_q=bt.w_q,
+        b_q=w_q,
         b_type=bt.wtype,
         b_group_scales=bt.w_g_s,
         b_group_zeros=w_g_zp,

benchmarks/kernels/benchmark_moe.py

@@ -1,24 +1,25 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import argparse
+import json
 import time
 from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
+from types import SimpleNamespace
 from typing import Any, TypedDict
 
 import ray
 import torch
-import triton
 from ray.experimental.tqdm_ray import tqdm
-from transformers import AutoConfig
 
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
+from vllm.transformers_utils.config import get_config
+from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
-FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm(
-) else torch.float8_e4m3fn
+FP8_DTYPE = current_platform.fp8_dtype()
 
 
 class BenchmarkConfig(TypedDict):
@@ -30,19 +31,18 @@ class BenchmarkConfig(TypedDict):
     num_stages: int
 
 
-def benchmark_config(
-    config: BenchmarkConfig,
-    num_tokens: int,
-    num_experts: int,
-    shard_intermediate_size: int,
-    hidden_size: int,
-    topk: int,
-    dtype: torch.dtype,
-    use_fp8_w8a8: bool,
-    use_int8_w8a16: bool,
-    num_iters: int = 100,
-    block_quant_shape: List[int] = None,
-) -> float:
+def benchmark_config(config: BenchmarkConfig,
+                     num_tokens: int,
+                     num_experts: int,
+                     shard_intermediate_size: int,
+                     hidden_size: int,
+                     topk: int,
+                     dtype: torch.dtype,
+                     use_fp8_w8a8: bool,
+                     use_int8_w8a16: bool,
+                     num_iters: int = 100,
+                     block_quant_shape: List[int] = None,
+                     use_deep_gemm: bool = False) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_int8_w8a16:
@@ -115,22 +115,41 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
-            fused_moe(
-                x,
-                w1,
-                w2,
-                input_gating,
-                topk,
-                renormalize=True,
-                inplace=True,
-                use_fp8_w8a8=use_fp8_w8a8,
-                use_int8_w8a16=use_int8_w8a16,
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a1_scale,
-                a2_scale=a2_scale,
-                block_shape=block_quant_shape,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False)
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -365,6 +384,8 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
+        block_quant_shape: List[int] = None,
+        use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
@@ -385,10 +406,18 @@ class BenchmarkWorker:
         else:
             config = op_config[min(op_config.keys(),
                                    key=lambda x: abs(x - num_tokens))]
-        kernel_time = benchmark_config(config, num_tokens, num_experts,
-                                       shard_intermediate_size, hidden_size,
-                                       topk, dtype, use_fp8_w8a8,
-                                       use_int8_w8a16)
+        kernel_time = benchmark_config(config,
+                                       num_tokens,
+                                       num_experts,
+                                       shard_intermediate_size,
+                                       hidden_size,
+                                       topk,
+                                       dtype,
+                                       use_fp8_w8a8,
+                                       use_int8_w8a16,
+                                       num_iters=100,
+                                       block_quant_shape=block_quant_shape,
+                                       use_deep_gemm=use_deep_gemm)
         return config, kernel_time
 
     def tune(
@@ -403,6 +432,7 @@ class BenchmarkWorker:
         use_int8_w8a16: bool,
         search_space: list[dict[str, int]],
         block_quant_shape: list[int],
+        use_deep_gemm: bool,
     ) -> dict[str, int]:
         best_config = None
         best_time = float("inf")
@@ -413,8 +443,14 @@ class BenchmarkWorker:
                                                    hidden_size, search_space,
                                                    is_fp16, topk)
 
-        with torch.cuda.device(self.device_id) if current_platform.is_rocm(
-        ) else nullcontext():
+        need_device_guard = False
+        if current_platform.is_rocm():
+            visible_device = os.environ.get("ROCR_VISIBLE_DEVICES", None)
+            if visible_device != f"{self.device_id}":
+                need_device_guard = True
+
+        with torch.cuda.device(
+                self.device_id) if need_device_guard else nullcontext():
             for config in tqdm(search_space):
                 try:
                     kernel_time = benchmark_config(
@@ -428,7 +464,8 @@ class BenchmarkWorker:
                         use_fp8_w8a8,
                         use_int8_w8a16,
                         num_iters=20,
-                        block_quant_shape=block_quant_shape)
+                        block_quant_shape=block_quant_shape,
+                        use_deep_gemm=use_deep_gemm)
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -487,11 +524,23 @@ def save_configs(configs: dict[int, BenchmarkConfig], num_experts: int,
         f.write("\n")
 
 
+def get_weight_block_size_safety(config, default_value=None):
+
+    quantization_config = getattr(config, 'quantization_config', {})
+    if isinstance(quantization_config, dict):
+        return quantization_config.get('weight_block_size', default_value)
+    return default_value
+
+
 def main(args: argparse.Namespace):
     print(args)
-    block_quant_shape = None
-    config = AutoConfig.from_pretrained(
-        args.model, trust_remote_code=args.trust_remote_code)
+
+    config = get_config(model=args.model,
+                        trust_remote_code=args.trust_remote_code)
+    if args.model_prefix:
+        config = getattr(config, args.model_prefix)
+    config = SimpleNamespace(**config)
+
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
@@ -502,14 +551,21 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif (config.architectures[0] == "DeepseekV3ForCausalLM"
-          or config.architectures[0] == "DeepseekV2ForCausalLM"):
+    elif (config.architectures[0]
+          in ("DeepseekV3ForCausalLM", "DeepseekV2ForCausalLM")):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-        block_quant_shape = config.quantization_config['weight_block_size']
+    elif config.architectures[0] in ("Qwen2MoeForCausalLM",
+                                     "Qwen3MoeForCausalLM"):
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.moe_intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
+        # Support for llama4
+        config = config.get_text_config()
         # Default: Mixtral.
         E = config.num_local_experts
         topk = config.num_experts_per_tok
@@ -517,9 +573,11 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
+    dtype = torch.float16 if current_platform.is_rocm() else getattr(
+        torch, config.torch_dtype)
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
+    block_quant_shape = get_weight_block_size_safety(config)
 
     if args.batch_size is None:
         batch_sizes = [
@@ -529,6 +587,17 @@ def main(args: argparse.Namespace):
     else:
         batch_sizes = [args.batch_size]
 
+    use_deep_gemm = bool(args.use_deep_gemm)
+
+    if current_platform.is_rocm() and "HIP_VISIBLE_DEVICES" in os.environ:
+        # Ray will set ROCR_VISIBLE_DEVICES for device visibility
+        logger.warning(
+            "Ray uses ROCR_VISIBLE_DEVICES to control device accessibility."
+            "Replacing HIP_VISIBLE_DEVICES with ROCR_VISIBLE_DEVICES.")
+        val = os.environ["HIP_VISIBLE_DEVICES"]
+        os.environ["ROCR_VISIBLE_DEVICES"] = val
+        del os.environ["HIP_VISIBLE_DEVICES"]
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
@@ -551,10 +620,10 @@ def main(args: argparse.Namespace):
 
         start = time.time()
         configs = _distribute(
-            "tune",
-            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, search_space, block_quant_shape)
-             for batch_size in batch_sizes])
+            "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
+                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space,
+                      block_quant_shape, use_deep_gemm)
+                     for batch_size in batch_sizes])
         best_configs = {
             M: sort_config(config)
             for M, config in zip(batch_sizes, configs)
@@ -568,7 +637,7 @@ def main(args: argparse.Namespace):
         outputs = _distribute(
             "benchmark",
             [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
-              use_fp8_w8a8, use_int8_w8a16, block_quant_shape)
+              use_fp8_w8a8, use_int8_w8a16, block_quant_shape, use_deep_gemm)
              for batch_size in batch_sizes])
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
@@ -590,10 +659,12 @@ if __name__ == "__main__":
                         type=str,
                         choices=["auto", "fp8_w8a8", "int8_w8a16"],
                         default="auto")
+    parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--tune", action="store_true")
     parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument("--model-prefix", type=str, required=False)
     args = parser.parse_args()
 
     main(args)

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -0,0 +1,349 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+from typing import Any, TypedDict
+
+import ray
+import torch
+from transformers import AutoConfig
+
+from vllm.model_executor.layers.fused_moe.deep_gemm_moe import (
+    _moe_permute, _moe_unpermute_and_reduce)
+from vllm.model_executor.layers.fused_moe.fused_moe import *
+from vllm.model_executor.layers.fused_moe.moe_permute_unpermute import *
+from vllm.model_executor.layers.fused_moe.utils import _fp8_quantize
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+FP8_DTYPE = current_platform.fp8_dtype()
+
+
+class BenchmarkConfig(TypedDict):
+    BLOCK_SIZE_M: int
+    BLOCK_SIZE_N: int
+    BLOCK_SIZE_K: int
+    GROUP_SIZE_M: int
+    num_warps: int
+    num_stages: int
+
+
+def benchmark_permute(num_tokens: int,
+                      num_experts: int,
+                      hidden_size: int,
+                      topk: int,
+                      dtype: torch.dtype,
+                      use_fp8_w8a8: bool,
+                      use_int8_w8a16: bool,
+                      num_iters: int = 100,
+                      use_customized_permute: bool = False) -> float:
+    # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
+    hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    # output_hidden_states = torch.empty_like(hidden_states)
+    if use_fp8_w8a8:
+        align_block_size = 128  # deepgemm needs 128 m aligned block
+        qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
+    else:
+        align_block_size = None
+        qhidden_states = hidden_states
+
+    gating_output = torch.randn(num_iters,
+                                num_tokens,
+                                num_experts,
+                                dtype=torch.float32)
+
+    input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        qhidden_states, input_gating, topk, False)
+
+    def prepare(i: int):
+        input_gating.copy_(gating_output[i])
+
+    def run():
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx,
+             m_indices) = moe_permute(
+                 qhidden_states,
+                 topk_weights=topk_weights,
+                 topk_ids=topk_ids,
+                 token_expert_indices=token_expert_indices,
+                 topk=topk,
+                 n_expert=num_experts,
+                 n_local_expert=num_experts,
+                 expert_map=None,
+                 align_block_size=align_block_size,
+             )
+        else:
+            (permuted_hidden_states, a1q_scale, sorted_token_ids, expert_ids,
+             inv_perm) = _moe_permute(qhidden_states, None, topk_ids,
+                                      num_experts, None, align_block_size)
+
+    # JIT compilation & warmup
+    run()
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run()
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        prepare(i)
+        torch.cuda.synchronize()
+
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+def benchmark_unpermute(num_tokens: int,
+                        num_experts: int,
+                        hidden_size: int,
+                        topk: int,
+                        dtype: torch.dtype,
+                        use_fp8_w8a8: bool,
+                        use_int8_w8a16: bool,
+                        num_iters: int = 100,
+                        use_customized_permute: bool = False) -> float:
+    # init_dtype = torch.float16 if use_fp8_w8a8 else dtype
+    hidden_states = torch.randn(num_tokens, hidden_size, dtype=dtype)
+    output_hidden_states = torch.empty_like(hidden_states)
+    if use_fp8_w8a8:
+        align_block_size = 128  # deepgemm needs 128 m aligned block
+        qhidden_states, scale = _fp8_quantize(hidden_states, None, None)
+    else:
+        align_block_size = None
+        qhidden_states = hidden_states
+
+    input_gating = torch.randn(num_tokens, num_experts, dtype=torch.float32)
+
+    topk_weights, topk_ids, token_expert_indices = fused_topk(
+        qhidden_states, input_gating, topk, False)
+
+    def prepare():
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx,
+             m_indices) = moe_permute(
+                 qhidden_states,
+                 topk_weights=topk_weights,
+                 topk_ids=topk_ids,
+                 token_expert_indices=token_expert_indices,
+                 topk=topk,
+                 n_expert=num_experts,
+                 n_local_expert=num_experts,
+                 expert_map=None,
+                 align_block_size=align_block_size,
+             )
+            # convert to fp16/bf16 as gemm output
+            return (permuted_hidden_states.to(dtype), first_token_off,
+                    inv_perm_idx, m_indices)
+        else:
+            (permuted_qhidden_states, a1q_scale, sorted_token_ids, expert_ids,
+             inv_perm) = _moe_permute(qhidden_states, None, topk_ids,
+                                      num_experts, None, align_block_size)
+            # convert to fp16/bf16 as gemm output
+            return (permuted_qhidden_states.to(dtype), a1q_scale,
+                    sorted_token_ids, expert_ids, inv_perm)
+
+    def run(input: tuple):
+        if use_customized_permute:
+            (permuted_hidden_states, first_token_off, inv_perm_idx,
+             m_indices) = input
+            moe_unpermute(permuted_hidden_states, topk_weights, topk_ids,
+                          inv_perm_idx, first_token_off, topk, num_experts,
+                          num_experts)
+        else:
+            (permuted_hidden_states, a1q_scale, sorted_token_ids, expert_ids,
+             inv_perm) = input
+            _moe_unpermute_and_reduce(output_hidden_states,
+                                      permuted_hidden_states, inv_perm,
+                                      topk_weights)
+
+    # JIT compilation & warmup
+    input = prepare()
+    run(input)
+    torch.cuda.synchronize()
+
+    # Capture 10 invocations with CUDA graph
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        for _ in range(10):
+            run(input)
+    torch.cuda.synchronize()
+
+    # Warmup
+    for _ in range(5):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    graph.reset()
+    return avg
+
+
+@ray.remote(num_gpus=1)
+class BenchmarkWorker:
+
+    def __init__(self, seed: int) -> None:
+        torch.set_default_device("cuda")
+        current_platform.seed_everything(seed)
+        self.seed = seed
+        # Get the device ID to allocate tensors and kernels
+        # on the respective GPU. This is required for Ray to work
+        # correctly with multi-GPU tuning on the ROCm platform.
+        self.device_id = int(ray.get_gpu_ids()[0])
+
+    def benchmark(
+        self,
+        num_tokens: int,
+        num_experts: int,
+        hidden_size: int,
+        topk: int,
+        dtype: torch.dtype,
+        use_fp8_w8a8: bool,
+        use_int8_w8a16: bool,
+        use_customized_permute: bool = False,
+    ) -> tuple[dict[str, int], float]:
+        current_platform.seed_everything(self.seed)
+
+        permute_time = benchmark_permute(
+            num_tokens,
+            num_experts,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            num_iters=100,
+            use_customized_permute=use_customized_permute)
+        unpermute_time = benchmark_unpermute(
+            num_tokens,
+            num_experts,
+            hidden_size,
+            topk,
+            dtype,
+            use_fp8_w8a8,
+            use_int8_w8a16,
+            num_iters=100,
+            use_customized_permute=use_customized_permute)
+        return permute_time, unpermute_time
+
+
+def get_weight_block_size_safety(config, default_value=None):
+
+    quantization_config = getattr(config, 'quantization_config', {})
+    if isinstance(quantization_config, dict):
+        return quantization_config.get('weight_block_size', default_value)
+    return default_value
+
+
+def main(args: argparse.Namespace):
+    print(args)
+
+    config = AutoConfig.from_pretrained(
+        args.model, trust_remote_code=args.trust_remote_code)
+    if config.architectures[0] == "DbrxForCausalLM":
+        E = config.ffn_config.moe_num_experts
+        topk = config.ffn_config.moe_top_k
+    elif config.architectures[0] == "JambaForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+    elif (config.architectures[0] == "DeepseekV3ForCausalLM"
+          or config.architectures[0] == "DeepseekV2ForCausalLM"):
+        E = config.n_routed_experts
+        topk = config.num_experts_per_tok
+    elif config.architectures[0] in [
+            "Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"
+    ]:
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+
+    else:
+        # Support for llama4
+        config = config.get_text_config()
+        # Default: Mixtral.
+        E = config.num_local_experts
+        topk = config.num_experts_per_tok
+
+    hidden_size = config.hidden_size
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
+    use_fp8_w8a8 = args.dtype == "fp8_w8a8"
+    use_int8_w8a16 = args.dtype == "int8_w8a16"
+    use_customized_permute = args.use_customized_permute
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
+            2048, 3072, 4096
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+
+    ray.init()
+    num_gpus = int(ray.available_resources()["GPU"])
+    workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
+
+    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
+        outputs = []
+        worker_idx = 0
+        for input_args in inputs:
+            worker = workers[worker_idx]
+            worker_method = getattr(worker, method)
+            output = worker_method.remote(*input_args)
+            outputs.append(output)
+            worker_idx = (worker_idx + 1) % num_gpus
+        return ray.get(outputs)
+
+    outputs = _distribute(
+        "benchmark", [(batch_size, E, hidden_size, topk, dtype, use_fp8_w8a8,
+                       use_int8_w8a16, use_customized_permute)
+                      for batch_size in batch_sizes])
+
+    for batch_size, (permute, unpermute) in zip(batch_sizes, outputs):
+        print(f"Batch size: {batch_size}")
+        print(f"Permute time: {permute:.2f} us")
+        print(f"Unpermute time: {unpermute:.2f} us")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument("--model",
+                        type=str,
+                        default="mistralai/Mixtral-8x7B-Instruct-v0.1")
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["auto", "fp8_w8a8", "int8_w8a16"],
+                        default="auto")
+    parser.add_argument("--use-customized-permute", action="store_true")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/benchmark_paged_attention.py

@@ -7,10 +7,13 @@ from typing import Optional
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
 
+logger = init_logger(__name__)
+
 NUM_BLOCKS = 128 * 1024
 PARTITION_SIZE = 512
 PARTITION_SIZE_ROCM = 256
@@ -176,7 +179,7 @@ def main(
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.
@@ -193,6 +196,9 @@ def main(
 
 
 if __name__ == '__main__':
+    logger.warning("This script benchmarks the paged attention kernel. "
+                   "By default this is no longer used in vLLM inference.")
+
     parser = FlexibleArgumentParser(
         description="Benchmark the paged attention kernel.")
     parser.add_argument("--version",

benchmarks/kernels/benchmark_quant.py

@@ -40,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_rmsnorm.py

@@ -4,11 +4,11 @@ import itertools
 from typing import Optional, Union
 
 import torch
-import triton
 from flashinfer.norm import fused_add_rmsnorm, rmsnorm
 from torch import nn
 
 from vllm import _custom_ops as vllm_ops
+from vllm.triton_utils import triton
 
 
 class HuggingFaceRMSNorm(nn.Module):
@@ -139,7 +139,7 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
 
     print(f"Naive output={output_naive}")
     print(f"FlashInfer output={output_flashinfer}")
-    print(f"VLLM output={output_vllm}")
+    print(f"vLLM output={output_vllm}")
 
     if torch.allclose(output_naive, output_flashinfer, atol=1e-2,
                       rtol=1e-2) and torch.allclose(

benchmarks/kernels/benchmark_shapes.py

@@ -75,3 +75,19 @@ WEIGHT_SHAPES = {
         [7168, 8192],
     ],
 }
+
+WEIGHT_SHAPES_MOE = {
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
+        [8, 2, 4096, 28672],
+        [8, 2, 14336, 4096],
+    ],
+    "nm-testing/deepseekv2-lite": [
+        [64, 6, 2048, 1408],
+    ],
+    "ibm-granite/granite-3.0-1b-a400m": [
+        [32, 8, 1024, 1024],
+    ],
+    "ibm-granite/granite-3.0-3b-a800m": [
+        [40, 8, 1024, 1536],
+    ],
+}

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -0,0 +1,420 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from sglang quantization/tuning_block_wise_kernel.py
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import torch
+import tqdm
+import triton
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    _w8a8_block_fp8_matmul)
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+mp.set_start_method("spawn", force=True)
+
+assert current_platform.is_cuda(
+), "Only support tune w8a8 block fp8 kernel on CUDA device."
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    config: dict[str, Any],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with 
+    block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization. 
+                    It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    N, K = B.shape
+    assert triton.cdiv(N, block_n) == Bs.shape[0]
+    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    C_shape = A.shape[:-1] + (N, )
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    def grid(META):
+        return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
+                triton.cdiv(N, META["BLOCK_SIZE_N"]), )
+
+    if A.dtype == torch.float8_e4m3fn:
+        kernel = _w8a8_block_fp8_matmul
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    kernel[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        **config,
+    )
+
+    return C
+
+
+def get_configs_compute_bound():
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_m,
+                                "BLOCK_SIZE_N": block_n,
+                                "BLOCK_SIZE_K": block_k,
+                                "GROUP_SIZE_M": group_size,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
+    return configs
+
+
+def get_weight_shapes(tp_size):
+    # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3.
+    # Modify them, if you tune for another different model.
+    # cannot TP
+    total = [
+        (512 + 64, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (7168, 16384),
+        (7168, 18432),
+    ]
+    # N can TP
+    n_tp = [
+        (18432 * 2, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (24576, 1536),
+        (12288, 7168),
+        (4096, 7168),
+    ]
+    # K can TP
+    k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)]
+
+    weight_shapes = []
+    for t in total:
+        weight_shapes.append(t)
+    for n_t in n_tp:
+        new_t = (n_t[0] // tp_size, n_t[1])
+        weight_shapes.append(new_t)
+    for k_t in k_tp:
+        new_t = (k_t[0], k_t[1] // tp_size)
+        weight_shapes.append(new_t)
+    return weight_shapes
+
+
+def benchmark_config(A,
+                     B,
+                     As,
+                     Bs,
+                     block_size,
+                     config,
+                     out_dtype=torch.float16,
+                     num_iters=10):
+
+    def run():
+        w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
+
+    torch.cuda.synchronize()
+    # JIT complication & warmup
+    for _ in range(5):
+        run()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        run()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    return avg
+
+
+def tune(M, N, K, block_size, out_dtype, search_space, input_type):
+    factor_for_scale = 1e-2
+
+    if input_type == "fp8":
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (
+            (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (
+            (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    As = torch.rand(M, k_tiles, dtype=torch.float32,
+                    device="cuda") * factor_for_scale
+    Bs = (torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") *
+          factor_for_scale)
+
+    best_config = None
+    best_time = float("inf")
+    for config in tqdm(search_space):
+        try:
+            kernel_time = benchmark_config(
+                A,
+                B,
+                As,
+                Bs,
+                block_size,
+                config,
+                out_dtype,
+                num_iters=10,
+            )
+        except triton.runtime.autotuner.OutOfResources:
+            # Some configurations may be invalid and fail to compile.
+            continue
+
+        if kernel_time < best_time:
+            best_time = kernel_time
+            best_config = config
+    now = datetime.now()
+    print(f"{now.ctime()}] Completed tuning for batch_size={M}")
+    assert best_config is not None
+    return best_config
+
+
+def save_configs(
+    N,
+    K,
+    block_n,
+    block_k,
+    configs,
+    save_path,
+    input_type="fp8",
+) -> None:
+    os.makedirs(save_path, exist_ok=True)
+    device_name = current_platform.get_device_name().replace(" ", "_")
+    json_file_name = (
+        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
+        f"block_shape=[{block_n},{block_k}].json")
+
+    config_file_path = os.path.join(save_path, json_file_name)
+    print(f"Writing best config to {config_file_path}...")
+
+    with open(config_file_path, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def tune_on_gpu(args_dict):
+    """Run tuning on a specific GPU."""
+    gpu_id = args_dict["gpu_id"]
+    batch_sizes = args_dict["batch_sizes"]
+    weight_shapes = args_dict["weight_shapes"]
+    args = args_dict["args"]
+
+    torch.cuda.set_device(gpu_id)
+    print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
+
+    block_n = args.block_n
+    block_k = args.block_k
+    out_dtype = DTYPE_MAP[args.out_dtype]
+    save_path = args.save_path
+    input_type = args.input_type
+
+    search_space = get_configs_compute_bound()
+    search_space = [
+        config for config in search_space
+        if block_k % config["BLOCK_SIZE_K"] == 0
+    ]
+
+    start = time.time()
+    for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"):
+        N, K = shape[0], shape[1]
+        print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`")
+        benchmark_results = [
+            tune(
+                batch_size,
+                N,
+                K,
+                [block_n, block_k],
+                out_dtype,
+                search_space,
+                input_type,
+            ) for batch_size in tqdm(batch_sizes,
+                                     desc=f"GPU {gpu_id} - Batch sizes")
+        ]
+        best_configs = {
+            M: config
+            for M, config in zip(batch_sizes, benchmark_results)
+        }
+        save_configs(N, K, block_n, block_k, best_configs, save_path,
+                     input_type)
+
+    end = time.time()
+    print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds")
+
+
+def distribute_batch_sizes(batch_sizes, num_gpus):
+    """Distribute batch sizes across available GPUs."""
+    batches_per_gpu = []
+    for i in range(num_gpus):
+        start_idx = i * len(batch_sizes) // num_gpus
+        end_idx = (i + 1) * len(batch_sizes) // num_gpus
+        batches_per_gpu.append(batch_sizes[start_idx:end_idx])
+    return batches_per_gpu
+
+
+def main(args):
+    print(args)
+    num_gpus = torch.cuda.device_count()
+    if num_gpus == 0:
+        raise RuntimeError("No GPU available for tuning")
+    print(f"Found {num_gpus} GPUs for parallel tuning")
+
+    torch.cuda.init()
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+        num_gpus = 1  # If only one batch size, use only one GPU
+
+    weight_shapes = get_weight_shapes(args.tp_size)
+
+    batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus)
+
+    process_args = []
+    for gpu_id in range(num_gpus):
+        process_args.append({
+            "gpu_id": gpu_id,
+            "batch_sizes": batches_per_gpu[gpu_id],
+            "weight_shapes":
+            weight_shapes,  # Each GPU processes all weight shapes
+            "args": args,
+        })
+
+    ctx = mp.get_context("spawn")
+    with ctx.Pool(num_gpus) as pool:
+        pool.map(tune_on_gpu, process_args)
+
+    print("Multi-GPU tuning completed")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""
+Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1:
+    python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8
+Then copy to model_executor/layers/quantization/utils/configs
+        """,
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--tp-size", "-tp", type=int, default=8)
+    parser.add_argument("--input-type",
+                        type=str,
+                        choices=["fp8"],
+                        default="fp8")
+    parser.add_argument(
+        "--out-dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16", "half"],
+        default="float16",
+    )
+    parser.add_argument("--block-n", type=int, default=128)
+    parser.add_argument("--block-k", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--save-path", type=str, default="./")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/deepgemm/README.md

@@ -0,0 +1,129 @@
+# DeepSeek DeepGEMM Kernels Benchmark
+
+This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
+
+Currently this just includes dense GEMMs and only works on Hopper GPUs.
+
+## Setup
+
+You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
+
+```
+git clone --recursive https://github.com/deepseek-ai/DeepGEMM
+cd DeepGEMM
+python setup.py install
+uv pip install -e .
+```
+
+## Usage
+
+```
+python benchmark_fp8_block_dense_gemm.py
+INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
+===== STARTING FP8 GEMM BENCHMARK =====
+PyTorch version: 2.5.1+cu124
+CUDA version: 12.4
+Triton version: 3.1.0
+Using device: NVIDIA H100 80GB HBM3
+WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+
+===== PERFORMANCE COMPARISON =====
+
+DeepGEMM Implementation:
++------+-------+-------+-----------+--------+--------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   |
++------+-------+-------+-----------+--------+--------+
+|    8 |  4096 |  7168 | 102.9     | 4.6    | 286.4  |
+|    8 |  7168 | 18432 | 70.8      | 29.8   | 1868.8 |
+|    8 | 18432 |  7168 | 69.3      | 30.5   | 1911.8 |
+|   64 |  4096 |  7168 | 69.1      | 54.4   | 439.0  |
+|   64 |  7168 | 18432 | 69.4      | 243.6  | 1933.6 |
+|   64 | 18432 |  7168 | 70.4      | 240.3  | 1917.2 |
+|   64 | 24576 |  1536 | 70.1      | 68.9   | 584.6  |
+|   64 | 32768 |   512 | 68.4      | 31.4   | 307.1  |
+|   64 |  7168 | 16384 | 69.5      | 216.3  | 1718.5 |
+|  128 |  4096 |  7168 | 141.1     | 53.3   | 222.1  |
+|  128 |  7168 | 18432 | 71.9      | 470.5  | 1896.1 |
+|  128 | 18432 |  7168 | 69.3      | 488.2  | 1988.2 |
+| 1024 |  4096 |  7168 | 89.7      | 670.1  | 502.5  |
+| 1024 | 18432 |  7168 | 279.0     | 969.8  | 635.2  |
+| 2048 |  4096 |  7168 | 175.1     | 687.0  | 347.4  |
+| 4096 |  4096 |  7168 | 335.4     | 717.0  | 275.1  |
++------+-------+-------+-----------+--------+--------+
+
+vLLM Triton Implementation:
++------+-------+-------+-----------+--------+--------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  |
++------+-------+-------+-----------+--------+--------+--------------+
+|    8 |  4096 |  7168 | 74.0      | 6.3    | 398.2  | 1.39x faster |
+|    8 |  7168 | 18432 | 89.6      | 23.6   | 1478.1 | 0.79x slower |
+|    8 | 18432 |  7168 | 113.2     | 18.7   | 1170.4 | 0.61x slower |
+|   64 |  4096 |  7168 | 79.4      | 47.3   | 382.2  | 0.87x slower |
+|   64 |  7168 | 18432 | 98.5      | 171.7  | 1363.0 | 0.70x slower |
+|   64 | 18432 |  7168 | 119.5     | 141.5  | 1129.4 | 0.59x slower |
+|   64 | 24576 |  1536 | 37.6      | 128.4  | 1089.7 | 1.86x faster |
+|   64 | 32768 |   512 | 38.7      | 55.5   | 542.6  | 1.77x faster |
+|   64 |  7168 | 16384 | 86.1      | 174.5  | 1386.4 | 0.81x slower |
+|  128 |  4096 |  7168 | 90.7      | 82.9   | 345.4  | 1.56x faster |
+|  128 |  7168 | 18432 | 144.0     | 234.9  | 946.9  | 0.50x slower |
+|  128 | 18432 |  7168 | 229.5     | 147.4  | 600.1  | 0.30x slower |
+| 1024 |  4096 |  7168 | 242.3     | 248.2  | 186.1  | 0.37x slower |
+| 1024 | 18432 |  7168 | 897.8     | 301.4  | 197.4  | 0.31x slower |
+| 2048 |  4096 |  7168 | 463.0     | 259.7  | 131.4  | 0.38x slower |
+| 4096 |  4096 |  7168 | 901.8     | 266.7  | 102.3  | 0.37x slower |
++------+-------+-------+-----------+--------+--------+--------------+
+
+vLLM CUTLASS Implementation:
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  | vs Triton    |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+|    8 |  4096 |  7168 | 34.6      | 13.6   | 852.3  | 2.98x faster | 2.14x faster |
+|    8 |  7168 | 18432 | 78.9      | 26.8   | 1677.3 | 0.90x slower | 1.13x faster |
+|    8 | 18432 |  7168 | 81.2      | 26.0   | 1631.1 | 0.85x slower | 1.39x faster |
+|   64 |  4096 |  7168 | 36.9      | 101.9  | 822.9  | 1.87x faster | 2.15x faster |
+|   64 |  7168 | 18432 | 87.4      | 193.4  | 1535.2 | 0.79x slower | 1.13x faster |
+|   64 | 18432 |  7168 | 85.0      | 199.0  | 1587.6 | 0.83x slower | 1.41x faster |
+|   64 | 24576 |  1536 | 28.0      | 172.8  | 1465.8 | 2.51x faster | 1.35x faster |
+|   64 | 32768 |   512 | 28.8      | 74.5   | 728.5  | 2.37x faster | 1.34x faster |
+|   64 |  7168 | 16384 | 77.9      | 193.0  | 1532.8 | 0.89x slower | 1.11x faster |
+|  128 |  4096 |  7168 | 39.1      | 192.4  | 802.0  | 3.61x faster | 2.32x faster |
+|  128 |  7168 | 18432 | 93.7      | 360.8  | 1454.2 | 0.77x slower | 1.54x faster |
+|  128 | 18432 |  7168 | 85.7      | 394.8  | 1608.0 | 0.81x slower | 2.68x faster |
+| 1024 |  4096 |  7168 | 99.7      | 603.1  | 452.2  | 0.90x slower | 2.43x faster |
+| 1024 | 18432 |  7168 | 331.3     | 816.7  | 534.9  | 0.84x slower | 2.71x faster |
+| 2048 |  4096 |  7168 | 198.3     | 606.6  | 306.7  | 0.88x slower | 2.34x faster |
+| 4096 |  4096 |  7168 | 392.2     | 613.2  | 235.3  | 0.86x slower | 2.30x faster |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+
+===== AVERAGE PERFORMANCE =====
++----------------+------------+----------+---------------+
+| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) |
++----------------+------------+----------+---------------+
+| DeepGEMM       | 310.98     | 1052.10  | 0.11          |
+| vLLM Triton    | 144.30     | 715.60   | 0.23          |
+| vLLM CUTLASS   | 286.78     | 1076.67  | 0.11          |
++----------------+------------+----------+---------------+
+
+===== AVERAGE SPEEDUPS =====
++-----------------------------+--------------+
+| Comparison                  | Speedup      |
++-----------------------------+--------------+
+| DeepGEMM vs vLLM Triton     | 1.71x faster |
+| DeepGEMM vs vLLM CUTLASS    | 0.94x slower |
+| vLLM CUTLASS vs vLLM Triton | 1.84x faster |
++-----------------------------+--------------+
+
+===== ACCURACY COMPARISON =====
++----------------+-----------------------+
+| Implementation | Avg Diff vs Reference |
++----------------+-----------------------+
+| DeepGEMM       | 0.000684              |
+| vLLM Triton    | 0.000684              |
+| vLLM CUTLASS   | 0.000684              |
++----------------+-----------------------+
+```

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -0,0 +1,464 @@
+# SPDX-License-Identifier: Apache-2.0
+# fmt: off
+# ruff: noqa: E501
+import time
+
+# Import DeepGEMM functions
+import deep_gemm
+import torch
+from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
+
+# Import vLLM functions
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    per_token_group_quant_fp8, w8a8_block_fp8_matmul)
+from vllm.triton_utils import triton
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9
+def per_token_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-token scaling."""
+    assert x.dim() == 2 and x.size(1) % 128 == 0
+    m, n = x.shape
+    x_view = x.view(m, -1, 128)
+    x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
+    return (x_view * (448.0 / x_amax.unsqueeze(2))).to(
+        torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17
+def per_block_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-block scaling."""
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128),
+                           dtype=x.dtype,
+                           device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
+    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
+        x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+
+
+def benchmark_shape(m: int,
+                    n: int,
+                    k: int,
+                    warmup: int = 100,
+                    repeat: int = 10000,
+                    verbose: bool = False) -> dict:
+    """Benchmark all implementations for a specific (m, n, k) shape."""
+    if verbose:
+        print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===")
+
+    # Create test tensors
+    A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
+    B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
+
+    # Reference result in BF16
+    torch.cuda.synchronize()
+    C_ref = A @ B.t()
+
+    # Pre-quantize B for all implementations
+    # (weights can be pre-quantized offline)
+    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B)
+    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B)
+
+    # Block size configuration
+    block_size = [128, 128]
+
+    # Pre-quantize A for all implementations
+    A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+    A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
+    C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
+    A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
+    A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
+        A, block_size[1], column_major_scales=True)
+
+    # === DeepGEMM Implementation ===
+    def deepgemm_gemm():
+        # A quantization is inside the loop as it depends on activations
+        # A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+        # A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(
+        #     A, block_size[1])
+        # A_scale_aligned = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
+        # C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
+        deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
+                                       (B_deepgemm, B_scale_deepgemm),
+                                       C_deepgemm)
+        return C_deepgemm
+
+    # === vLLM Triton Implementation ===
+    def vllm_triton_gemm():
+        # A quantization is inside the loop as it depends on activations
+        # A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
+        return w8a8_block_fp8_matmul(A_vllm,
+                                     B_vllm,
+                                     A_scale_vllm,
+                                     B_scale_vllm,
+                                     block_size,
+                                     output_dtype=torch.bfloat16)
+
+    # === vLLM CUTLASS Implementation ===
+    def vllm_cutlass_gemm():
+        # A quantization is inside the loop as it depends on activations
+        # A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
+        #     A, block_size[1], column_major_scales=True)
+        return ops.cutlass_scaled_mm(A_vllm_cutlass,
+                                     B_vllm.T,
+                                     scale_a=A_scale_vllm_cutlass,
+                                     scale_b=B_scale_vllm.T,
+                                     out_dtype=torch.bfloat16)
+
+    # Run correctness check first
+    if verbose:
+        print("Running correctness check...")
+    C_deepgemm = deepgemm_gemm()
+    C_vllm_triton = vllm_triton_gemm()
+    C_vllm_cutlass = vllm_cutlass_gemm()
+
+    deepgemm_diff = calc_diff(C_deepgemm, C_ref)
+    vllm_triton_diff = calc_diff(C_vllm_triton, C_ref)
+    vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref)
+
+    if verbose:
+        print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}")
+        print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}")
+        print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}")
+        print("vLLM Triton vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}")
+        print("vLLM CUTLASS vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}")
+
+    # Benchmark implementations
+    implementations = {
+        "DeepGEMM": deepgemm_gemm,
+        "vLLM Triton": vllm_triton_gemm,
+        "vLLM CUTLASS": vllm_cutlass_gemm
+    }
+
+    benchmark_results = {
+        "shape": {
+            "m": m,
+            "n": n,
+            "k": k
+        },
+        "implementations": {}
+    }
+
+    for name, func in implementations.items():
+        # Warmup
+        for _ in range(warmup):
+            func()
+            torch.cuda.synchronize()
+
+        # Timing loop
+        torch.cuda.synchronize()
+        start = time.time()
+        for _ in range(repeat):
+            func()
+        torch.cuda.synchronize()
+        end = time.time()
+
+        # Calculate timing and TFLOPS
+        avg_time_ms = (end - start) / repeat * 1000
+        avg_time_us = avg_time_ms * 1000
+        tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12
+        gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3)
+
+        benchmark_results["implementations"][name] = {
+            "time_ms": avg_time_ms,
+            "time_us": avg_time_us,
+            "tflops": tflops,
+            "gb_s": gb_s,
+            "diff": {
+                "DeepGEMM":
+                0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm),
+                "Reference":
+                deepgemm_diff if name == "DeepGEMM" else
+                (vllm_triton_diff
+                 if name == "vLLM Triton" else vllm_cutlass_diff)
+            }
+        }
+
+        if verbose:
+            print(
+                f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s"
+            )
+
+    # Calculate speedups
+    baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"]
+    for name, data in benchmark_results["implementations"].items():
+        if name != "DeepGEMM":
+            speedup = baseline / data["time_ms"]
+            benchmark_results["implementations"][name][
+                "speedup_vs_deepgemm"] = speedup
+            if verbose:
+                print(f"DeepGEMM is {1/speedup:.2f}x "
+                      f"{'faster' if 1/speedup > 1 else 'slower'} than {name}")
+
+    vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][
+        "time_ms"]
+    vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][
+        "time_ms"]
+    cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time
+    benchmark_results["implementations"]["vLLM CUTLASS"][
+        "speedup_vs_triton"] = cutlass_vs_triton
+    if verbose:
+        print(
+            f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x "
+            f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton"
+        )
+
+    return benchmark_results
+
+
+def format_table_row(values, widths):
+    """Format a row with specified column widths."""
+    return "| " + " | ".join(f"{val:{w}}"
+                             for val, w in zip(values, widths)) + " |"
+
+
+def print_table(headers, rows, title=None):
+    """Print a table with headers and rows."""
+    if title:
+        print(f"\n{title}")
+
+    # Calculate column widths based on headers and data
+    widths = [
+        max(len(str(h)), max(len(str(row[i])) for row in rows))
+        for i, h in enumerate(headers)
+    ]
+
+    # Create separator line
+    separator = "+-" + "-+-".join("-" * w for w in widths) + "-+"
+
+    # Print table
+    print(separator)
+    print(format_table_row(headers, widths))
+    print(separator)
+    for row in rows:
+        print(format_table_row(row, widths))
+    print(separator)
+
+
+def format_speedup(value):
+    """Format speedup value with indicator if it's faster or slower."""
+    return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}"
+
+
+def run_benchmarks(verbose: bool = False):
+    """Run benchmarks for a set of common shapes."""
+    print("===== STARTING FP8 GEMM BENCHMARK =====")
+
+    # Make sure we're using the GPU
+    if not torch.cuda.is_available():
+        print("CUDA not available! Tests require GPU.")
+        return
+
+    # Print system information
+    print(f"PyTorch version: {torch.__version__}")
+    print(f"CUDA version: {torch.version.cuda}")
+    print(f"Triton version: {triton.__version__}")
+    print(f"Using device: {torch.cuda.get_device_name()}")
+
+    # Enable TF32 for better performance
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+    # Set seeds for reproducibility
+    torch.manual_seed(42)
+    torch.cuda.manual_seed(42)
+
+    # Define benchmark shapes (m, n, k)
+    shapes = [
+        (8, 4096, 7168),
+        (8, 7168, 18432),
+        (8, 18432, 7168),
+        (64, 4096, 7168),
+        (64, 7168, 18432),
+        (64, 18432, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 18432),
+        (128, 18432, 7168),
+        (1024, 4096, 7168),
+        (1024, 18432, 7168),
+        (2048, 4096, 7168),
+        (4096, 4096, 7168),
+    ]
+    shapes = [
+        # (64, 2112, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (64, 4096, 7168),
+        (64, 7168, 2048),
+        # (128, 2112, 7168),
+        (128, 24576, 1536),
+        (128, 32768, 512),
+        (128, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 2048),
+        # (4096, 2112, 7168),
+        (4096, 24576, 1536),
+        (4096, 32768, 512),
+        (4096, 7168, 16384),
+        (4096, 4096, 7168),
+        (4096, 7168, 2048),
+    ]
+
+    all_results = []
+    for m, n, k in shapes:
+        result = benchmark_shape(m, n, k, verbose=verbose)
+        all_results.append(result)
+
+    # Print results in a nicely formatted table
+    print("\n===== PERFORMANCE COMPARISON =====")
+
+    # Print DeepGEMM table
+    deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"]
+    deepgemm_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["DeepGEMM"]
+        deepgemm_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}"
+        ])
+
+    print_table(deepgemm_headers,
+                deepgemm_rows,
+                title="DeepGEMM Implementation:")
+
+    # Print vLLM Triton table
+    triton_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"
+    ]
+    triton_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM Triton"]
+        speedup = impl_data.get("speedup_vs_deepgemm", 1.0)
+        triton_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(speedup)
+        ])
+
+    print_table(triton_headers,
+                triton_rows,
+                title="vLLM Triton Implementation:")
+
+    # Print vLLM CUTLASS table
+    cutlass_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM",
+        "vs Triton"
+    ]
+    cutlass_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM CUTLASS"]
+        vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0)
+        vs_triton = impl_data.get("speedup_vs_triton", 1.0)
+        cutlass_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(vs_deepgemm),
+            format_speedup(vs_triton)
+        ])
+
+    print_table(cutlass_headers,
+                cutlass_rows,
+                title="vLLM CUTLASS Implementation:")
+
+    # Calculate and print averages
+    print("\n===== AVERAGE PERFORMANCE =====")
+
+    implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"]
+    avg_metrics = {
+        impl: {
+            "tflops": 0,
+            "gb_s": 0,
+            "time_ms": 0
+        }
+        for impl in implementations
+    }
+
+    for result in all_results:
+        for impl in implementations:
+            impl_data = result["implementations"][impl]
+            avg_metrics[impl]["tflops"] += impl_data["tflops"]
+            avg_metrics[impl]["gb_s"] += impl_data["gb_s"]
+            avg_metrics[impl]["time_ms"] += impl_data["time_ms"]
+
+    num_shapes = len(all_results)
+    avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"]
+    avg_rows = []
+
+    for impl in implementations:
+        avg_tflops = avg_metrics[impl]["tflops"] / num_shapes
+        avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes
+        avg_time = avg_metrics[impl]["time_ms"] / num_shapes
+        avg_rows.append([
+            impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"
+        ])
+
+    print_table(avg_headers, avg_rows)
+
+    # Calculate average speedups
+    avg_speedups = {
+        "DeepGEMM vs vLLM Triton": 0,
+        "DeepGEMM vs vLLM CUTLASS": 0,
+        "vLLM CUTLASS vs vLLM Triton": 0
+    }
+
+    for result in all_results:
+        deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"]
+        vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"]
+        vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][
+            "time_ms"]
+
+        avg_speedups[
+            "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time
+        avg_speedups[
+            "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time
+        avg_speedups[
+            "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time
+
+    print("\n===== AVERAGE SPEEDUPS =====")
+    speedup_headers = ["Comparison", "Speedup"]
+    speedup_rows = []
+    for comparison, total in avg_speedups.items():
+        avg_speedup = total / num_shapes
+        status = "faster" if avg_speedup > 1 else "slower"
+        speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"])
+
+    print_table(speedup_headers, speedup_rows)
+
+    # Average accuracy comparison
+    print("\n===== ACCURACY COMPARISON =====")
+    avg_diff = {impl: 0 for impl in implementations}
+
+    for result in all_results:
+        for impl in implementations:
+            avg_diff[impl] += result["implementations"][impl]["diff"][
+                "Reference"]
+
+    diff_headers = ["Implementation", "Avg Diff vs Reference"]
+    diff_rows = []
+    for impl in implementations:
+        diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"])
+
+    print_table(diff_headers, diff_rows)
+
+
+if __name__ == "__main__":
+    run_benchmarks(verbose=False)