[V1] Fix profiling.py

Signed-off-by: Alexander Matveev <alexm@neuralmagic.com>

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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,5 +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-3-8B-Instruct -b 32 -l 250 -f 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
 model_name: "meta-llama/Meta-Llama-3-8B-Instruct"
 tasks:
 - name: "gsm8k"

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

@@ -1,4 +1,3 @@
-# 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-INT8-compressed-tensors.yaml

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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:

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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,5 +1,4 @@
-# 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
+# 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
 model_name: "mistralai/Mixtral-8x7B-Instruct-v0.1"
 tasks:
 - name: "gsm8k"

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

@@ -1,12 +1,11 @@
-# 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
+    value: 0.31
   - name: "exact_match,flexible-extract"
-    value: 0.465
+    value: 0.47
 limit: 1319
 num_fewshot: 5

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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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

@@ -1,4 +1,3 @@
-# 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/SparseLlama3.1_2of4_fp8_compressed.yaml

@@ -1,4 +1,3 @@
-# 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/test_lm_eval_correctness.py

@@ -16,7 +16,7 @@ import numpy
 import pytest
 import yaml
 
-RTOL = 0.08
+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")

.buildkite/release-pipeline.yaml

@@ -86,18 +86,3 @@ steps:
       - "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/scripts/hardware_ci/run-amd-test.sh

@@ -98,13 +98,6 @@ if [[ $commands == *" kernels "* ]]; then
   --ignore=kernels/test_machete_mm.py \
   --ignore=kernels/test_mha_attn.py \
   --ignore=kernels/test_block_fp8.py \
-  --ignore=kernels/test_cutlass_moe.py \
-  --ignore=kernels/test_mamba_ssm_ssd.py \
-  --ignore=kernels/test_attention.py \
-  --ignore=kernels/test_block_int8.py \
-  --ignore=kernels/test_fused_quant_layernorm.py \
-  --ignore=kernels/test_int8_kernel.py \
-  --ignore=kernels/test_triton_moe_ptpc_fp8.py \
   --ignore=kernels/test_permute_cols.py"
 fi
 

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

@@ -5,41 +5,10 @@
 set -ex
 
 # Setup cleanup
-remove_docker_container() {
-  if [[ -n "$container_id" ]]; then
-      podman rm -f "$container_id" || true
-  fi
-  podman system prune -f
-}
+remove_docker_container() { docker rm -f cpu-test || true; docker 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
+docker build -t cpu-test -f docker/Dockerfile.ppc64le .
 

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

@@ -1,13 +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
-
-# Setup cleanup
-remove_docker_container() { docker rm -f cpu-test || true; docker system prune -f; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Try building the docker image
-docker build -t cpu-test -f docker/Dockerfile.s390x .

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

@@ -17,13 +17,10 @@ source /etc/environment
 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 pytest \
     && 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 \
     && pytest -v -s /workspace/vllm/tests/v1/tpu/test_perf.py \
     && echo TEST_1 \
@@ -43,11 +40,7 @@ docker run --privileged --net host --shm-size=16G -it \
     && echo TEST_8 \
     && pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py \
     && echo TEST_9 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py \
-    && echo TEST_10 \
-    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py \
-    && echo TEST_11 \
-    && pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py" \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py" \
 
 
 # TODO: This test fails because it uses RANDOM_SEED sampling

.buildkite/test-pipeline.yaml

@@ -8,7 +8,6 @@
 # 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.
@@ -71,7 +70,6 @@ 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
@@ -106,7 +104,6 @@ steps:
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
   fast_check: true
-  torch_nightly: true
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
@@ -121,7 +118,7 @@ steps:
   - 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
   - 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/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/
   - pytest -v -s entrypoints/test_chat_utils.py
   - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
@@ -208,8 +205,6 @@ steps:
     - 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
@@ -299,7 +294,6 @@ steps:
   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
   source_file_dependencies:
@@ -318,46 +312,15 @@ steps:
   commands:
   - pytest -v -s compile/test_full_graph.py
 
-- label: Kernels Core Operation Test
+- label: Kernels Test %N # 1h each
+  # mirror_hardwares: [amd]
   source_file_dependencies:
   - csrc/
-  - tests/kernels/core
-  commands:
-    - pytest -v -s kernels/core
-
-- label: Kernels Attention Test %N
-  source_file_dependencies:
-  - csrc/attention/
   - vllm/attention
-  - vllm/v1/attention
-  - tests/kernels/attention
+  - tests/kernels
   commands:
-    - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 2
-
-- label: Kernels Quantization Test %N
-  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
-  source_file_dependencies:
-  - csrc/moe/
-  - tests/kernels/moe
-  - vllm/model_executor/layers/fused_moe/
-  commands:
-    - pytest -v -s kernels/moe
-
-- label: Kernels Mamba Test
-  source_file_dependencies:
-  - csrc/mamba/
-  - tests/kernels/mamba
-  commands:
-    - pytest -v -s kernels/mamba
+    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 4
 
 - label: Tensorizer Test # 11min
   # mirror_hardwares: [amd]
@@ -378,13 +341,6 @@ steps:
   commands:
   - bash scripts/run-benchmarks.sh
 
-- label: Benchmarks CLI Test # 10min
-  source_file_dependencies:
-  - vllm/
-  - tests/benchmarks/
-  commands:
-  - pytest -v -s benchmarks/
-
 - label: Quantization Test # 33min
   source_file_dependencies:
   - csrc/
@@ -422,10 +378,8 @@ steps:
   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  #####
 
@@ -437,9 +391,8 @@ steps:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.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 'not llama4'
     - 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
   #mirror_hardwares: [amd]
@@ -449,8 +402,6 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
-    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/language -m core_model
 
@@ -462,8 +413,6 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
-    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
-    - pip install causal-conv1d
     - 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'
 
@@ -584,14 +533,11 @@ steps:
   - 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)'
-  # 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
   - 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"

.github/CODEOWNERS

@@ -12,7 +12,6 @@
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth
 /vllm/model_executor/guided_decoding @mgoin @russellb
 /vllm/multimodal @DarkLight1337 @ywang96
-/vllm/vllm_flash_attn @LucasWilkinson
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1

.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/vllm/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/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/vllm/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/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

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/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/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/vllm/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/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/mergify.yml

@@ -55,19 +55,11 @@ pull_request_rules:
   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~=^tests/v1/structured_output/
-      - files=tests/v1/entrypoints/llm/test_guided_generate.py
-      - files~=^vllm/v1/structured_output/
+      - files=benchmarks/benchmark_serving_guided.py
+      - files=benchmarks/benchmark_guided.py
   actions:
     label:
       add:
@@ -126,28 +118,6 @@ pull_request_rules:
       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

.gitignore

@@ -3,6 +3,7 @@
 
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
+!vllm/vllm_flash_attn/fa_utils.py
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -202,6 +203,3 @@ benchmarks/**/*.json
 # Linting
 actionlint
 shellcheck*/
-
-# Ingore moe/marlin_moe gen code
-csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -11,6 +11,7 @@ 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
   hooks:

CMakeLists.txt

@@ -251,7 +251,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # 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 "v3.9.0" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v3.8.0" 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})
@@ -269,7 +269,7 @@ 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 v3.9.0
+        GIT_TAG v3.8.0
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -290,8 +290,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "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/attention/mla/cutlass_mla_entry.cu")
+    "csrc/cutlass_extensions/common.cpp")
 
   set_gencode_flags_for_srcs(
     SRCS "${VLLM_EXT_SRC}"
@@ -464,26 +463,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(FP4_ARCHS)
   endif()
 
-  # 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 "${SRCS}"
-      CUDA_ARCHS "${MLA_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MLA=1")
-    # Add MLA-specific include directories only to MLA source files
-    set_source_files_properties(${SRCS}
-      PROPERTIES INCLUDE_DIRECTORIES "${CUTLASS_DIR}/examples/77_blackwell_fmha;${CUTLASS_DIR}/examples/common")
-    message(STATUS "Building CUTLASS MLA for archs: ${MLA_ARCHS}")
-  else()
-    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
@@ -629,51 +609,21 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   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 ${CMAKE_COMMAND} -E env
-        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$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 ${MOE_MARLIN_GEN_SCRIPT_HASH}
-            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
-        message(STATUS "Marlin MOE generation completed successfully.")
-      endif()
-    else()
-      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
-    endif()
-
-    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MOE_WNAA16_MARLIN_SRC}"
+      SRCS "${MARLIN_MOE_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
 
-    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
-
+    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_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"
@@ -698,7 +648,6 @@ 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(

README.md

@@ -10,7 +10,7 @@ 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://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> |
+| <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://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---

benchmarks/backend_request_func.py

@@ -1,6 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 
-import io
 import json
 import os
 import sys
@@ -33,7 +32,6 @@ class RequestFuncInput:
     extra_body: Optional[dict] = None
     multi_modal_content: Optional[dict] = None
     ignore_eos: bool = False
-    language: Optional[str] = None
 
 
 @dataclass
@@ -438,110 +436,6 @@ 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
@@ -599,7 +493,6 @@ 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,

benchmarks/benchmark_dataset.py

@@ -64,7 +64,6 @@ class SampleRequest:
 
 class BenchmarkDataset(ABC):
     DEFAULT_SEED = 0
-    IS_MULTIMODAL = False
 
     def __init__(
         self,
@@ -622,7 +621,6 @@ class ConversationDataset(HuggingFaceDataset):
     SUPPORTED_DATASET_PATHS = {
         'lmms-lab/LLaVA-OneVision-Data', 'Aeala/ShareGPT_Vicuna_unfiltered'
     }
-    IS_MULTIMODAL = True
 
     def sample(self,
                tokenizer: PreTrainedTokenizerBase,
@@ -687,7 +685,6 @@ class VisionArenaDataset(HuggingFaceDataset):
         "lmarena-ai/vision-arena-bench-v0.1":
         lambda x: x["turns"][0][0]["content"]
     }
-    IS_MULTIMODAL = True
 
     def sample(
         self,
@@ -818,80 +815,3 @@ class AIMODataset(HuggingFaceDataset):
                 ))
         self.maybe_oversample_requests(sampled_requests, num_requests)
         return sampled_requests
-
-
-# -----------------------------------------------------------------------------
-# ASR Dataset Implementation
-# -----------------------------------------------------------------------------
-
-
-class ASRDataset(HuggingFaceDataset):
-    """
-    Dataset class for processing a ASR dataset for transcription.
-    Tested on the following set:
-
-    +----------------+----------------------------------------+--------------------------+-----------------------------+
-    | Dataset        | Domain                                 | Speaking Style           | hf-subset                   |
-    +----------------+----------------------------------------+--------------------------+-----------------------------+
-    | TED-LIUM       | TED talks                              | Oratory                  | release1, release2, release3|
-    |                |                                        |                          | release3-speaker-adaptation |
-    | VoxPopuli      | European Parliament                    | Oratory                  | en, de, it, fr,  ...        |
-    | LibriSpeech    | Audiobook                              | Narrated                 | "LIUM/tedlium"              |
-    | GigaSpeech     | Audiobook, podcast, YouTube            | Narrated, spontaneous    | xs, s, m, l, xl, dev, test  |
-    | SPGISpeech     | Financial meetings                     | Oratory, spontaneous     | S, M, L, dev, test          |
-    | AMI            | Meetings                               | Spontaneous              | ihm, sdm                    |
-    +----------------+----------------------------------------+--------------------------+-----------------------------+
-
-    """ # noqa: E501
-    SUPPORTED_DATASET_PATHS = {
-        "openslr/librispeech_asr", "facebook/voxpopuli", "LIUM/tedlium",
-        "edinburghcstr/ami", "speechcolab/gigaspeech", "kensho/spgispeech"
-    }
-
-    DEFAULT_OUTPUT_LEN = 128
-    IS_MULTIMODAL = True
-
-    # TODO Whisper-specific. Abstract interface when more models are supported.
-    TRANSCRIPTION_PREAMBLE = "<|startoftranscript|><|en|><|transcribe|>"\
-                              "<|notimestamps|>"
-    skip_long_audios: bool = True
-
-    def sample(
-        self,
-        tokenizer: PreTrainedTokenizerBase,
-        num_requests: int,
-        output_len: Optional[int] = None,
-        **kwargs,
-    ) -> list:
-        import librosa
-        output_len = (output_len
-                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
-        prompt = ASRDataset.TRANSCRIPTION_PREAMBLE
-        prompt_len = len(tokenizer(prompt).input_ids)
-        sampled_requests = []
-        skipped = 0
-        for item in self.data:
-            if len(sampled_requests) >= num_requests:
-                break
-            audio = item["audio"]
-            y, sr = audio["array"], audio["sampling_rate"]
-            duration_s = librosa.get_duration(y=y, sr=sr)
-            # Whisper max supported duration
-            if self.skip_long_audios and duration_s > 30:
-                skipped += 1
-                continue
-
-            mm_content = {"audio": (y, sr)}
-            sampled_requests.append(
-                SampleRequest(
-                    prompt=prompt,
-                    prompt_len=prompt_len,
-                    expected_output_len=output_len,
-                    multi_modal_data=mm_content,
-                ))
-        if skipped:
-            logger.warning("%d samples discarded from dataset due to" \
-                           " their length being greater than" \
-                           " what Whisper supports.", skipped)
-        self.maybe_oversample_requests(sampled_requests, num_requests)
-        return sampled_requests

benchmarks/benchmark_prefix_caching.py

@@ -63,16 +63,14 @@ class Request:
     output_len: int
 
 
-def sample_tokens(tokenizer: PreTrainedTokenizerBase,
-                  length: int) -> list[int]:
+def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> str:
     vocab = tokenizer.get_vocab()
-    all_special_ids = set(tokenizer.all_special_ids)
-
     # Remove the special tokens.
-    return random.choices(
-        [v for k, v in vocab.items() if k not in all_special_ids],
-        k=length,
-    )
+    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)
 
 
 def sample_requests_from_dataset(

benchmarks/benchmark_serving.py

@@ -50,7 +50,7 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
-from benchmark_dataset import (AIMODataset, ASRDataset, BurstGPTDataset,
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
                                ConversationDataset, HuggingFaceDataset,
                                InstructCoderDataset, RandomDataset,
                                SampleRequest, ShareGPTDataset, SonnetDataset,
@@ -274,6 +274,10 @@ async def benchmark(
         input_requests[0].expected_output_len, \
             input_requests[0].multi_modal_data
 
+    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.")
     assert test_mm_content is None or isinstance(test_mm_content, dict)
     test_input = RequestFuncInput(
         model=model_id,
@@ -600,9 +604,6 @@ def main(args: argparse.Namespace):
         elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
             dataset_class = AIMODataset
             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__()
@@ -614,13 +615,6 @@ def main(args: argparse.Namespace):
                 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,
@@ -713,7 +707,7 @@ def main(args: argparse.Namespace):
         ))
 
     # Save config and results to json
-    if args.save_result or args.append_result:
+    if args.save_result:
         result_json: dict[str, Any] = {}
 
         # Setup
@@ -734,14 +728,6 @@ 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")
-        result_json["burstiness"] = args.burstiness
-        result_json["max_concurrency"] = args.max_concurrency
-
-        # Merge with benchmark result
-        result_json = {**result_json, **benchmark_result}
 
         if not args.save_detailed:
             # Remove fields with too many data points
@@ -752,6 +738,15 @@ def main(args: argparse.Namespace):
                 if field in result_json:
                     del result_json[field]
 
+        # Traffic
+        result_json["request_rate"] = (args.request_rate if args.request_rate
+                                       < float("inf") else "inf")
+        result_json["burstiness"] = args.burstiness
+        result_json["max_concurrency"] = args.max_concurrency
+
+        # Merge with benchmark result
+        result_json = {**result_json, **benchmark_result}
+
         # Save to file
         base_model_id = model_id.split("/")[-1]
         max_concurrency_str = (f"-concurrency{args.max_concurrency}"
@@ -761,12 +756,7 @@ 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,
-                  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")
+        with open(file_name, "w", encoding='utf-8') as outfile:
             json.dump(result_json, outfile)
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
@@ -898,11 +888,6 @@ if __name__ == "__main__":
         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",

benchmarks/benchmark_serving_structured_output.py

@@ -51,7 +51,7 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
-from vllm.v1.structured_output.backend_xgrammar import (
+from vllm.v1.structured_output.utils import (
     has_xgrammar_unsupported_json_features)
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
@@ -150,17 +150,17 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
 
     elif args.dataset == "grammar":
         schema = """
-        root ::= select_statement
+            ?start: select_statement
 
-        select_statement ::= "SELECT " column " from " table " where " condition
+            ?select_statement: "SELECT " column_list " FROM " table_name
 
-        column ::= "col_1 " | "col_2 "
+            ?column_list: column_name ("," column_name)*
 
-        table ::= "table_1 " | "table_2 "
+            ?table_name: identifier
 
-        condition ::= column "= " number
+            ?column_name: identifier
 
-        number ::= "1 " | "2 "
+            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
         """
         prompt = "Generate an SQL query to show the 'username' \
             and 'email' from the 'users' table."

benchmarks/benchmark_throughput.py

@@ -523,13 +523,6 @@ def validate_args(args):
         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.")
@@ -604,26 +597,18 @@ if __name__ == "__main__":
     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.")
+        default=0,
+        help="Number of fixed prefix tokens before the random "
+        "context in a request (default: 0).",
+    )
     # 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 "
+        default=0.0,
+        help="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)].",
     )
 

benchmarks/kernels/benchmark_bitblas.py

@@ -1,236 +0,0 @@
-# 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_lora.py

@@ -17,14 +17,8 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-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.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())

benchmarks/kernels/benchmark_moe.py

@@ -553,8 +553,9 @@ def main(args: argparse.Namespace):
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
-        # Support for llama4
-        config = config.get_text_config()
+        if not hasattr(config, "hidden_size"):
+            # Support for llama4
+            config = config.text_config
         # Default: Mixtral.
         E = config.num_local_experts
         topk = config.num_experts_per_tok

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 8798f27777fb57f447070301bf33a9f9c607f491
+          GIT_TAG dc9d410b3e2d6534a4c70724c2515f4def670a22
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

collect_env.py

@@ -282,20 +282,12 @@ def get_vllm_version():
 
     if __version__ == "dev":
         return "N/A (dev)"
-    version_str = __version_tuple__[-1]
-    if isinstance(version_str, str) and version_str.startswith('g'):
-        # it's a dev build
-        if '.' in version_str:
-            # it's a dev build containing local changes
-            git_sha = version_str.split('.')[0][1:]
-            date = version_str.split('.')[-1][1:]
-            return f"{__version__} (git sha: {git_sha}, date: {date})"
-        else:
-            # it's a dev build without local changes
-            git_sha = version_str[1:]  # type: ignore
-            return f"{__version__} (git sha: {git_sha})"
-    return __version__
 
+    if len(__version_tuple__) == 4: # dev build
+        git_sha = __version_tuple__[-1][1:] # type: ignore
+        return f"{__version__} (git sha: {git_sha}"
+
+    return __version__
 
 def summarize_vllm_build_flags():
     # This could be a static method if the flags are constant, or dynamic if you need to check environment variables, etc.
@@ -510,9 +502,7 @@ def get_pip_packages(run_lambda, patterns=None):
             print("uv is set")
             cmd = ["uv", "pip", "list", "--format=freeze"]
         else:
-            raise RuntimeError(
-                "Could not collect pip list output (pip or uv module not available)"
-            )
+            raise RuntimeError("Could not collect pip list output (pip or uv module not available)")
 
         out = run_and_read_all(run_lambda, cmd)
         return "\n".join(line for line in out.splitlines()
@@ -545,12 +535,13 @@ def is_xnnpack_available():
     else:
         return "N/A"
 
-
 def get_env_vars():
     env_vars = ''
-    secret_terms = ('secret', 'token', 'api', 'access', 'password')
-    report_prefix = ("TORCH", "NCCL", "PYTORCH", "CUDA", "CUBLAS", "CUDNN",
-                     "OMP_", "MKL_", "NVIDIA")
+    secret_terms=('secret', 'token', 'api', 'access', 'password')
+    report_prefix = ("TORCH", "NCCL", "PYTORCH",
+                     "CUDA", "CUBLAS", "CUDNN",
+                     "OMP_", "MKL_",
+                     "NVIDIA")
     for k, v in os.environ.items():
         if any(term in k.lower() for term in secret_terms):
             continue
@@ -561,7 +552,6 @@ def get_env_vars():
 
     return env_vars
 
-
 def get_env_info():
     run_lambda = run
     pip_version, pip_list_output = get_pip_packages(run_lambda)

csrc/attention/merge_attn_states.cu

@@ -107,14 +107,13 @@ __global__ void merge_attn_states_kernel(
 
 #define LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS)                     \
   {                                                                         \
-    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS>                   \
-        <<<grid, block, 0, stream>>>(                                       \
-            reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr, \
-            reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),          \
-            reinterpret_cast<float*>(prefix_lse.data_ptr()),                \
-            reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),          \
-            reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,    \
-            num_heads, head_size);                                          \
+    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS><<<grid, block>>>( \
+        reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr,     \
+        reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),              \
+        reinterpret_cast<float*>(prefix_lse.data_ptr()),                    \
+        reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),              \
+        reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,        \
+        num_heads, head_size);                                              \
   }
 
 /*@brief Merges the attention states from prefix and suffix
@@ -123,10 +122,10 @@ __global__ void merge_attn_states_kernel(
  * @param output [n,h,d] The output tensor to store the merged attention states.
  * @param output_lse [h,d] Optional tensor to store the log-sum-exp values.
  * @param prefix_output [n,h,d] The prefix attention states.
- * @param prefix_lse [h,n] The log-sum-exp values for the prefix attention
+ * @param prefix_lse [h,d] The log-sum-exp values for the prefix attention
  * states.
  * @param suffix_output [n,h,d] The suffix attention states.
- * @param suffix_lse [h,n] The log-sum-exp values for the suffix attention
+ * @param suffix_lse [h,d] The log-sum-exp values for the suffix attention
  * states.
  */
 template <typename scalar_t>
@@ -147,17 +146,13 @@ void merge_attn_states_launcher(torch::Tensor& output,
   if (output_lse.has_value()) {
     output_lse_ptr = output_lse.value().data_ptr<float>();
   }
-  // Process one pack elements per thread. for float, the
-  // pack_size is 4 for half/bf16, the pack_size is 8.
+  // process one pack elements per thread. float -> 4, half/bf16 -> 8
   const uint threads_per_head = head_size / pack_size;
   const uint total_threads = num_tokens * num_heads * threads_per_head;
 
   dim3 block(NUM_THREADS);
   dim3 grid((total_threads + NUM_THREADS - 1) / NUM_THREADS);
 
-  const c10::cuda::OptionalCUDAGuard device_guard(prefix_output.device());
-  auto stream = at::cuda::getCurrentCUDAStream();
-
   LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS);
 }
 

csrc/attention/mla/cutlass_mla_entry.cu

@@ -1,38 +0,0 @@
-/*
- * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <torch/all.h>
-
-#if defined ENABLE_CUTLASS_MLA && ENABLE_CUTLASS_MLA
-void cutlass_mla_decode_sm100a(torch::Tensor const& out,
-                               torch::Tensor const& q_nope,
-                               torch::Tensor const& q_pe,
-                               torch::Tensor const& kv_c_and_k_pe_cache,
-                               torch::Tensor const& seq_lens,
-                               torch::Tensor const& page_table, double scale);
-#endif
-
-void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
-                        torch::Tensor const& q_pe,
-                        torch::Tensor const& kv_c_and_k_pe_cache,
-                        torch::Tensor const& seq_lens,
-                        torch::Tensor const& page_table, double scale) {
-#if defined ENABLE_CUTLASS_MLA && ENABLE_CUTLASS_MLA
-  return cutlass_mla_decode_sm100a(out, q_nope, q_pe, kv_c_and_k_pe_cache,
-                                   seq_lens, page_table, scale);
-#endif
-  TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled cutlass MLA");
-}

csrc/attention/mla/cutlass_mla_kernels.cu

@@ -1,225 +0,0 @@
-/*
- * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <torch/all.h>
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include "cute/tensor.hpp"
-
-#include "cutlass/cutlass.h"
-#include "cutlass/kernel_hardware_info.h"
-
-#include "cutlass_extensions/common.hpp"
-
-#include "device/sm100_mla.hpp"
-#include "kernel/sm100_mla_tile_scheduler.hpp"
-
-using namespace cute;
-using namespace cutlass::fmha::kernel;
-
-template <typename T, bool PersistenceOption = true>
-struct MlaSm100 {
-  using Element = T;
-  using ElementAcc = float;
-  using ElementOut = T;
-
-  using TileShape = Shape<_128, _128, Shape<_512, _64>>;
-  using TileShapeH = cute::tuple_element_t<0, TileShape>;
-  using TileShapeD = cute::tuple_element_t<2, TileShape>;
-
-  // H K (D_latent D_rope) B
-  using ProblemShape = cute::tuple<TileShapeH, int, TileShapeD, int>;
-
-  using StrideQ = cute::tuple<int64_t, _1, int64_t>;  // H D B
-  using StrideK = cute::tuple<int64_t, _1, int64_t>;  // K D B
-  using StrideO = StrideK;                            // H D B
-  using StrideLSE = cute::tuple<_1, int>;             // H B
-
-  using TileScheduler =
-      std::conditional_t<PersistenceOption, Sm100MlaPersistentTileScheduler,
-                         Sm100MlaIndividualTileScheduler>;
-
-  using FmhaKernel =
-      cutlass::fmha::kernel::Sm100FmhaMlaKernelTmaWarpspecialized<
-          TileShape, Element, ElementAcc, ElementOut, ElementAcc, TileScheduler,
-          /*kIsCpAsync=*/true>;
-  using Fmha = cutlass::fmha::device::MLA<FmhaKernel>;
-};
-
-template <typename T>
-typename T::Fmha::Arguments args_from_options(
-    at::Tensor const& out, at::Tensor const& q_nope, at::Tensor const& q_pe,
-    at::Tensor const& kv_c_and_k_pe_cache, at::Tensor const& seq_lens,
-    at::Tensor const& page_table, double scale) {
-  cutlass::KernelHardwareInfo hw_info;
-  hw_info.device_id = q_nope.device().index();
-  hw_info.sm_count =
-      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
-          hw_info.device_id);
-
-  int batches = q_nope.sizes()[0];
-  int page_count_per_seq = page_table.sizes()[1];
-  int page_count_total = kv_c_and_k_pe_cache.sizes()[0];
-  int page_size = kv_c_and_k_pe_cache.sizes()[1];
-  int max_seq_len = page_size * page_count_per_seq;
-  using TileShapeH = typename T::TileShapeH;
-  using TileShapeD = typename T::TileShapeD;
-  auto problem_shape =
-      cute::make_tuple(TileShapeH{}, max_seq_len, TileShapeD{}, batches);
-
-  auto [H, K, D, B] = problem_shape;
-  auto [D_latent, D_rope] = D;
-
-  using StrideQ = typename T::StrideQ;
-  using StrideK = typename T::StrideK;
-  using StrideO = typename T::StrideO;
-  using StrideLSE = typename T::StrideLSE;
-
-  StrideQ stride_Q_latent = cute::make_tuple(
-      static_cast<int64_t>(D_latent), _1{}, static_cast<int64_t>(H * D_latent));
-  StrideQ stride_Q_rope = cute::make_tuple(static_cast<int64_t>(D_rope), _1{},
-                                           static_cast<int64_t>(H * D_rope));
-  StrideK stride_C =
-      cute::make_tuple(static_cast<int64_t>(D_latent + D_rope), _1{},
-                       static_cast<int64_t>(page_size * (D_latent + D_rope)));
-  StrideLSE stride_PT = cute::make_stride(_1{}, page_count_per_seq);
-  StrideLSE stride_LSE = cute::make_tuple(_1{}, static_cast<int>(H));
-  StrideO stride_O = cute::make_tuple(static_cast<int64_t>(D_latent), _1{},
-                                      static_cast<int64_t>(H * D_latent));
-
-  using Element = typename T::Element;
-  using ElementOut = typename T::ElementOut;
-  using ElementAcc = typename T::ElementAcc;
-  auto Q_latent_ptr = static_cast<Element*>(q_nope.data_ptr());
-  auto Q_rope_ptr = static_cast<Element*>(q_pe.data_ptr());
-  auto C_ptr = static_cast<Element*>(kv_c_and_k_pe_cache.data_ptr());
-  auto scale_f = static_cast<float>(scale);
-  typename T::Fmha::Arguments arguments{
-      problem_shape,
-      {scale_f, Q_latent_ptr, stride_Q_latent, Q_rope_ptr, stride_Q_rope, C_ptr,
-       stride_C, C_ptr + D_latent, stride_C,
-       static_cast<int*>(seq_lens.data_ptr()),
-       static_cast<int*>(page_table.data_ptr()), stride_PT, page_count_total,
-       page_size},
-      {static_cast<ElementOut*>(out.data_ptr()), stride_O,
-       static_cast<ElementAcc*>(nullptr), stride_LSE},
-      hw_info,
-      -1,       // split_kv
-      nullptr,  // is_var_split_kv
-  };
-  // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
-  // split_kv automatically based on batch size and sequence length to balance
-  // workload across available SMs. Consider using var_split_kv for manual
-  // control if needed.
-  T::Fmha::set_split_kv(arguments);
-  return arguments;
-}
-
-template <typename Element>
-void runMla(at::Tensor const& out, at::Tensor const& q_nope,
-            at::Tensor const& q_pe, at::Tensor const& kv_c_and_k_pe_cache,
-            at::Tensor const& seq_lens, at::Tensor const& page_table,
-            float scale, cudaStream_t stream) {
-  using MlaSm100Type = MlaSm100<Element>;
-  typename MlaSm100Type::Fmha fmha;
-  auto arguments = args_from_options<MlaSm100Type>(
-      out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, scale);
-  size_t workspace_size = MlaSm100Type::Fmha::get_workspace_size(arguments);
-  auto const workspace_options =
-      torch::TensorOptions().dtype(torch::kUInt8).device(q_nope.device());
-  auto workspace = torch::empty(workspace_size, workspace_options);
-
-  CUTLASS_CHECK(fmha.can_implement(arguments));
-
-  CUTLASS_CHECK(fmha.initialize(arguments, workspace.data_ptr(), stream));
-
-  CUTLASS_CHECK(fmha.run(arguments, workspace.data_ptr(), stream));
-}
-
-void cutlass_mla_decode_sm100a(torch::Tensor const& out,
-                               torch::Tensor const& q_nope,
-                               torch::Tensor const& q_pe,
-                               torch::Tensor const& kv_c_and_k_pe_cache,
-                               torch::Tensor const& seq_lens,
-                               torch::Tensor const& page_table, double scale) {
-  TORCH_CHECK(q_nope.device().is_cuda(), "q_nope must be on CUDA");
-  TORCH_CHECK(q_nope.dim() == 3, "q_nope must be a 3D tensor");
-  TORCH_CHECK(q_pe.dim() == 3, "q_pe must be a 3D tensor");
-  TORCH_CHECK(kv_c_and_k_pe_cache.dim() == 3,
-              "kv_c_and_k_pe_cache must be a 3D tensor");
-  TORCH_CHECK(seq_lens.dim() == 1, "seq_lens must be a 1D tensor");
-  TORCH_CHECK(page_table.dim() == 2, "page_table must be a 2D tensor");
-  TORCH_CHECK(out.dim() == 3, "out must be a 3D tensor");
-
-  auto B_q_nope = q_nope.size(0);
-  auto H_q_nope = q_nope.size(1);
-  auto D_q_nope = q_nope.size(2);
-  auto B_q_pe = q_pe.size(0);
-  auto H_q_pe = q_pe.size(1);
-  auto D_q_pe = q_pe.size(2);
-  auto B_pt = page_table.size(0);
-  auto PAGE_NUM = page_table.size(1);
-  auto PAGE_SIZE = kv_c_and_k_pe_cache.size(1);
-  auto D_ckv = kv_c_and_k_pe_cache.size(2);
-  auto B_o = out.size(0);
-  auto H_o = out.size(1);
-  auto D_o = out.size(2);
-
-  TORCH_CHECK(D_q_nope == 512, "D_q_nope must be equal to 512");
-  TORCH_CHECK(D_q_pe == 64, "D_q_pe must be equal to 64");
-  TORCH_CHECK(D_ckv == 576, "D_ckv must be equal to 576");
-  TORCH_CHECK(H_q_nope == H_q_pe && H_q_nope == H_o && H_o == 128,
-              "H_q_nope, H_q_pe, and H_o must be equal to 128");
-  TORCH_CHECK(PAGE_SIZE > 0 && (PAGE_SIZE & (PAGE_SIZE - 1)) == 0,
-              "PAGE_SIZE must be a power of 2");
-  TORCH_CHECK(
-      B_q_nope == B_q_pe && B_q_nope == B_pt && B_q_nope == B_o,
-      "Batch dims must be same for page_table, q_nope and q_pe, and out");
-  TORCH_CHECK(PAGE_NUM % (128 / PAGE_SIZE) == 0,
-              "PAGE_NUM must be divisible by 128 / PAGE_SIZE");
-  TORCH_CHECK(D_o == 512, "D_o must be equal to 512");
-
-  TORCH_CHECK(q_nope.dtype() == at::ScalarType::Half ||
-                  q_nope.dtype() == at::ScalarType::BFloat16 ||
-                  q_nope.dtype() == at::ScalarType::Float8_e4m3fn,
-              "q_nope must be a half, bfloat16, or float8_e4m3fn tensor");
-  TORCH_CHECK(kv_c_and_k_pe_cache.dtype() == q_nope.dtype() &&
-                  q_nope.dtype() == q_pe.dtype(),
-              "kv_c_and_k_pe_cache, q_nope, and q_pe must be the same type");
-  TORCH_CHECK(seq_lens.dtype() == torch::kInt32,
-              "seq_lens must be a 32-bit integer tensor");
-  TORCH_CHECK(page_table.dtype() == torch::kInt32,
-              "page_table must be a 32-bit integer tensor");
-
-  auto in_dtype = q_nope.dtype();
-  at::cuda::CUDAGuard device_guard{(char)q_nope.get_device()};
-  const cudaStream_t stream =
-      at::cuda::getCurrentCUDAStream(q_nope.get_device());
-  if (in_dtype == at::ScalarType::Half) {
-    runMla<cutlass::half_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens,
-                            page_table, scale, stream);
-  } else if (in_dtype == at::ScalarType::BFloat16) {
-    runMla<cutlass::bfloat16_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache,
-                                seq_lens, page_table, scale, stream);
-  } else if (in_dtype == at::ScalarType::Float8_e4m3fn) {
-    runMla<cutlass::float_e4m3_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache,
-                                  seq_lens, page_table, scale, stream);
-  } else {
-    TORCH_CHECK(false, "Unsupported input data type of MLA");
-  }
-}

csrc/cache_kernels.cu

@@ -270,10 +270,9 @@ __global__ void reshape_and_cache_flash_kernel(
     cache_t* __restrict__ value_cache,   // [num_blocks, block_size, num_heads,
                                          // head_size]
     const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int64_t block_stride, const int64_t page_stride,
-    const int64_t head_stride, const int64_t key_stride,
-    const int64_t value_stride, const int num_heads, const int head_size,
-    const int block_size, const float* k_scale, const float* v_scale) {
+    const int block_stride, const int key_stride, const int value_stride,
+    const int num_heads, const int head_size, const int block_size,
+    const float* k_scale, const float* v_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   // NOTE: slot_idx can be -1 if the token is padded
@@ -289,8 +288,8 @@ __global__ void reshape_and_cache_flash_kernel(
     const int head_idx = i / head_size;
     const int head_offset = i % head_size;
     const int64_t tgt_key_value_idx = block_idx * block_stride +
-                                      block_offset * page_stride +
-                                      head_idx * head_stride + head_offset;
+                                      block_offset * num_heads * head_size +
+                                      head_idx * head_size + head_offset;
     scalar_t tgt_key = key[src_key_idx];
     scalar_t tgt_value = value[src_value_idx];
     if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
@@ -397,16 +396,16 @@ void reshape_and_cache(
 // KV_T is the data type of key and value tensors.
 // CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
-#define CALL_RESHAPE_AND_CACHE_FLASH(KV_T, CACHE_T, KV_DTYPE)             \
-  vllm::reshape_and_cache_flash_kernel<KV_T, CACHE_T, KV_DTYPE>           \
-      <<<grid, block, 0, stream>>>(                                       \
-          reinterpret_cast<KV_T*>(key.data_ptr()),                        \
-          reinterpret_cast<KV_T*>(value.data_ptr()),                      \
-          reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),               \
-          reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),             \
-          slot_mapping.data_ptr<int64_t>(), block_stride, page_stride,    \
-          head_stride, key_stride, value_stride, num_heads, head_size,    \
-          block_size, reinterpret_cast<const float*>(k_scale.data_ptr()), \
+#define CALL_RESHAPE_AND_CACHE_FLASH(KV_T, CACHE_T, KV_DTYPE)         \
+  vllm::reshape_and_cache_flash_kernel<KV_T, CACHE_T, KV_DTYPE>       \
+      <<<grid, block, 0, stream>>>(                                   \
+          reinterpret_cast<KV_T*>(key.data_ptr()),                    \
+          reinterpret_cast<KV_T*>(value.data_ptr()),                  \
+          reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
+          reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
+          slot_mapping.data_ptr<int64_t>(), block_stride, key_stride, \
+          value_stride, num_heads, head_size, block_size,             \
+          reinterpret_cast<const float*>(k_scale.data_ptr()),         \
           reinterpret_cast<const float*>(v_scale.data_ptr()));
 
 void reshape_and_cache_flash(
@@ -433,11 +432,9 @@ void reshape_and_cache_flash(
   int head_size = key.size(2);
   int block_size = key_cache.size(1);
 
-  int64_t key_stride = key.stride(0);
-  int64_t value_stride = value.stride(0);
-  int64_t block_stride = key_cache.stride(0);
-  int64_t page_stride = key_cache.stride(1);
-  int64_t head_stride = key_cache.stride(2);
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+  int block_stride = key_cache.stride(0);
   TORCH_CHECK(key_cache.stride(0) == value_cache.stride(0));
 
   dim3 grid(num_tokens);

csrc/custom_all_reduce.cuh

@@ -375,7 +375,7 @@ class CustomAllreduce {
   bool fully_connected_;
 
   RankSignals sg_;
-  // Stores a map from a pointer to its peer pointers from all ranks.
+  // Stores an map from a pointer to its peer pointers from all ranks.
   std::unordered_map<void*, RankData*> buffers_;
   Signal* self_sg_;
 

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -1,103 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-import glob
-import itertools
-import os
-import subprocess
-
-import jinja2
-
-FILE_HEAD = """
-// auto generated by generate.py
-// clang-format off
-
-#include "kernel.h"
-#include "marlin_template.h"
-
-namespace MARLIN_NAMESPACE_NAME {
-""".strip()
-
-TEMPLATE = ("template __global__ void Marlin<"
-            "{{scalar_t}}, "
-            "{{w_type_id}}, "
-            "{{threads}}, "
-            "{{thread_m_blocks}}, "
-            "{{thread_n_blocks}}, "
-            "{{thread_k_blocks}}, "
-            "{{'true' if m_block_size_8 else 'false'}}, "
-            "{{stages}}, "
-            "{{'true' if has_act_order else 'false'}}, "
-            "{{'true' if has_zp else 'false'}}, "
-            "{{group_blocks}}, "
-            "{{'true' if is_zp_float else 'false'}}>"
-            "( MARLIN_KERNEL_PARAMS );")
-
-# int8 with zero point case (vllm::kU8) is also supported,
-# we don't add it to reduce wheel size.
-SCALAR_TYPES = ["vllm::kU4", "vllm::kU4B8", "vllm::kU8B128"]
-THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128)]
-
-THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4]
-# group_blocks:
-#   = 0 : act order case
-#   = -1 : channelwise quantization
-#   > 0 : group_size=16*group_blocks
-GROUP_BLOCKS = [0, -1, 2, 4, 8]
-DTYPES = ["fp16", "bf16"]
-
-
-def remove_old_kernels():
-    for filename in glob.glob(os.path.dirname(__file__) + "/kernel_*.cu"):
-        subprocess.call(["rm", "-f", filename])
-
-
-def generate_new_kernels():
-    for scalar_type, dtype in itertools.product(SCALAR_TYPES, DTYPES):
-        has_zp = "B" not in scalar_type
-        all_template_str_list = []
-
-        for group_blocks, m_blocks, thread_configs in itertools.product(
-                GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS):
-
-            has_act_order = group_blocks == 0
-            if has_zp and has_act_order:
-                continue
-            if thread_configs[2] == 256:
-                if m_blocks <= 1 and thread_configs[0] != 128:
-                    continue
-                if m_blocks > 1 and thread_configs[0] != 64:
-                    continue
-
-            k_blocks = thread_configs[0] // 16
-            n_blocks = thread_configs[1] // 16
-            threads = thread_configs[2]
-
-            c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
-
-            template_str = jinja2.Template(TEMPLATE).render(
-                scalar_t=c_dtype,
-                w_type_id=scalar_type + ".id()",
-                threads=threads,
-                thread_m_blocks=max(m_blocks, 1),
-                thread_n_blocks=n_blocks,
-                thread_k_blocks=k_blocks,
-                m_block_size_8=m_blocks == 0.5,
-                stages="pipe_stages",
-                has_act_order=has_act_order,
-                has_zp=has_zp,
-                group_blocks=group_blocks,
-                is_zp_float=False,
-            )
-
-            all_template_str_list.append(template_str)
-
-        file_content = FILE_HEAD + "\n\n"
-        file_content += "\n\n".join(all_template_str_list) + "\n\n}\n"
-        filename = f"kernel_{dtype}_{scalar_type[6:].lower()}.cu"
-
-        with open(os.path.join(os.path.dirname(__file__), filename), "w") as f:
-            f.write(file_content)
-
-
-if __name__ == "__main__":
-    remove_old_kernels()
-    generate_new_kernels()

csrc/moe/marlin_moe_wna16/kernel.h

@@ -1,44 +0,0 @@
-
-#ifndef MARLIN_NAMESPACE_NAME
-  #define MARLIN_NAMESPACE_NAME marlin_moe_wna16
-#endif
-
-#include "quantization/gptq_marlin/marlin.cuh"
-#include "quantization/gptq_marlin/marlin_dtypes.cuh"
-#include "core/scalar_type.hpp"
-
-#define MARLIN_KERNEL_PARAMS                                                \
-  const int4 *__restrict__ A, const int4 *__restrict__ B,                   \
-      int4 *__restrict__ C, int4 *__restrict__ C_tmp,                       \
-      const int4 *__restrict__ scales_ptr, const int4 *__restrict__ zp_ptr, \
-      const int *__restrict__ g_idx,                                        \
-      const int32_t *__restrict__ sorted_token_ids_ptr,                     \
-      const int32_t *__restrict__ expert_ids_ptr,                           \
-      const int32_t *__restrict__ num_tokens_past_padded_ptr,               \
-      const float *__restrict__ topk_weights_ptr, int top_k,                \
-      bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,        \
-      int prob_n, int prob_k, int *locks, bool use_atomic_add,              \
-      bool use_fp32_reduce
-
-namespace MARLIN_NAMESPACE_NAME {
-template <typename scalar_t,  // compute dtype, half or nv_float16
-          const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
-          const int threads,          // number of threads in a threadblock
-          const int thread_m_blocks,  // number of 16x16 blocks in the m
-                                      // dimension (batchsize) of the
-                                      // threadblock
-          const int thread_n_blocks,  // same for n dimension (output)
-          const int thread_k_blocks,  // same for k dimension (reduction)
-          const bool m_block_size_8,  // whether m_block_size == 8
-                                      // only works when thread_m_blocks == 1
-          const int stages,  // number of stages for the async global->shared
-                             // fetch pipeline
-          const bool has_act_order,  // whether act_order is enabled
-          const bool has_zp,         // whether zero-points are enabled
-          const int group_blocks,    // number of consecutive 16x16 blocks
-                                     // with a separate quantization scale
-          const bool is_zp_float     // is zero point of float16 type?
-          >
-__global__ void Marlin(MARLIN_KERNEL_PARAMS);
-
-}

csrc/moe/marlin_moe_wna16/ops.cu

@@ -1,927 +0,0 @@
-/*
- * Modified by Neural Magic
- * Copyright (C) Marlin.2024 Elias Frantar
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *         http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-/*
- * Adapted from https://github.com/IST-DASLab/marlin
- */
-
-#ifndef MARLIN_NAMESPACE_NAME
-  #define MARLIN_NAMESPACE_NAME marlin_moe_wna16
-#endif
-
-#include "kernel.h"
-#include "core/registration.h"
-
-#define STATIC_ASSERT_SCALAR_TYPE_VALID(scalar_t)               \
-  static_assert(std::is_same<scalar_t, half>::value ||          \
-                    std::is_same<scalar_t, nv_bfloat16>::value, \
-                "only float16 and bfloat16 is supported");
-
-namespace MARLIN_NAMESPACE_NAME {
-
-__global__ void MarlinDefault(MARLIN_KERNEL_PARAMS){};
-
-using MarlinFuncPtr = void (*)(MARLIN_KERNEL_PARAMS);
-
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
-
-template <int moe_block_size>
-__global__ void permute_cols_kernel(
-    int4 const* __restrict__ a_int4_ptr, int const* __restrict__ perm_int_ptr,
-    int4* __restrict__ out_int4_ptr,
-    const int32_t* __restrict__ sorted_token_ids_ptr,
-    const int32_t* __restrict__ expert_ids_ptr,
-    const int32_t* __restrict__ num_tokens_past_padded_ptr, int size_m,
-    int size_k, int top_k) {};
-
-}  // namespace marlin
-
-torch::Tensor moe_wna16_marlin_gemm(
-    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
-    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
-    std::optional<torch::Tensor> const& b_zeros_or_none,
-    std::optional<torch::Tensor> const& g_idx_or_none,
-    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
-    torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
-    torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
-    int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
-    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
-    int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
-    bool is_zp_float) {
-  TORCH_CHECK_NOT_IMPLEMENTED(false,
-                              "marlin_gemm(..) requires CUDA_ARCH >= 8.0");
-  return torch::empty({1, 1});
-}
-
-#else
-
-// For a given "a" of size [M,K] performs a permutation of the K columns based
-// on the given "perm" indices.
-template <int moe_block_size>
-__global__ void permute_cols_kernel(
-    int4 const* __restrict__ a_int4_ptr, int const* __restrict__ perm_int_ptr,
-    int4* __restrict__ out_int4_ptr,
-    const int32_t* __restrict__ sorted_token_ids_ptr,
-    const int32_t* __restrict__ expert_ids_ptr,
-    const int32_t* __restrict__ num_tokens_past_padded_ptr, int size_m,
-    int size_k, int top_k) {
-  int num_tokens_past_padded = num_tokens_past_padded_ptr[0];
-  int num_moe_blocks = div_ceil(num_tokens_past_padded, moe_block_size);
-  int32_t block_sorted_ids[moe_block_size];
-  int block_num_valid_tokens = 0;
-  int64_t old_expert_id = 0;
-  int64_t expert_id = 0;
-  int row_stride = size_k * sizeof(half) / 16;
-
-  auto read_moe_block_data = [&](int block_id) {
-    block_num_valid_tokens = moe_block_size;
-    int4* tmp_block_sorted_ids = reinterpret_cast<int4*>(block_sorted_ids);
-    for (int i = 0; i < moe_block_size / 4; i++) {
-      tmp_block_sorted_ids[i] =
-          ((int4*)sorted_token_ids_ptr)[block_id * moe_block_size / 4 + i];
-    }
-    for (int i = 0; i < moe_block_size; i++) {
-      if (block_sorted_ids[i] >= size_m * top_k) {
-        block_num_valid_tokens = i;
-        break;
-      };
-    }
-  };
-
-  auto permute_row = [&](int row) {
-    int iters = size_k / default_threads;
-    int rest = size_k % default_threads;
-
-    int in_offset = (row / top_k) * row_stride;
-    int out_offset = row * row_stride;
-
-    half const* a_row_half =
-        reinterpret_cast<half const*>(a_int4_ptr + in_offset);
-    half* out_half = reinterpret_cast<half*>(out_int4_ptr + out_offset);
-
-    int base_k = 0;
-
-    for (int i = 0; i < iters; i++) {
-      int cur_k = base_k + threadIdx.x;
-      int src_pos = perm_int_ptr[cur_k];
-
-      out_half[cur_k] = a_row_half[src_pos];
-
-      base_k += default_threads;
-    }
-
-    if (rest) {
-      if (threadIdx.x < rest) {
-        int cur_k = base_k + threadIdx.x;
-        int src_pos = perm_int_ptr[cur_k];
-
-        out_half[cur_k] = a_row_half[src_pos];
-      }
-    }
-  };
-
-  for (int index = blockIdx.x; index < num_moe_blocks; index += gridDim.x) {
-    old_expert_id = expert_id;
-    int tmp_expert_id = expert_ids_ptr[index];
-    if (tmp_expert_id == -1) continue;
-    expert_id = tmp_expert_id;
-    perm_int_ptr += (expert_id - old_expert_id) * size_k;
-    read_moe_block_data(index);
-
-    for (int i = 0; i < block_num_valid_tokens; i++)
-      permute_row(block_sorted_ids[i]);
-  }
-}
-
-typedef struct {
-  int thread_k;
-  int thread_n;
-  int num_threads;
-} thread_config_t;
-
-thread_config_t small_batch_thread_configs[] = {
-    // Ordered by priority
-
-    // thread_k, thread_n, num_threads
-    {128, 128, 256},
-    {64, 128, 128}};
-
-thread_config_t large_batch_thread_configs[] = {
-    // Ordered by priority
-
-    // thread_k, thread_n, num_threads
-    {64, 256, 256},
-    {64, 128, 128}};
-
-typedef struct {
-  int blocks_per_sm;
-  thread_config_t tb_cfg;
-} exec_config_t;
-
-int get_scales_cache_size(thread_config_t const& th_config, int prob_m,
-                          int prob_n, int prob_k, int num_bits, int group_size,
-                          bool has_act_order, bool is_k_full) {
-  bool cache_scales_chunk = has_act_order && !is_k_full;
-
-  int tb_n = th_config.thread_n;
-  int tb_k = th_config.thread_k;
-
-  // Get max scale groups per thread-block
-  int tb_groups;
-  if (group_size == -1) {
-    tb_groups = 1;
-  } else if (group_size == 0) {
-    tb_groups = div_ceil(tb_k, 32);  // Worst case is 32 group size
-  } else {
-    tb_groups = div_ceil(tb_k, group_size);
-  }
-
-  if (cache_scales_chunk) {
-    int load_groups =
-        tb_groups * pipe_stages * 2;     // Chunk size is 2x pipeline over dim K
-    load_groups = max(load_groups, 32);  // We load at least 32 scale groups
-    return load_groups * tb_n * 2;
-
-  } else {
-    int tb_scales = tb_groups * tb_n * 2;
-
-    return tb_scales * pipe_stages;
-  }
-}
-
-int get_kernel_cache_size(thread_config_t const& th_config, int thread_m_blocks,
-                          int prob_m, int prob_n, int prob_k, int num_bits,
-                          int group_size, bool has_act_order, bool is_k_full,
-                          int has_zp, int is_zp_float) {
-  int pack_factor = 32 / num_bits;
-
-  // Get B size
-  int tb_k = th_config.thread_k;
-  int tb_n = th_config.thread_n;
-  int tb_m = thread_m_blocks * 16;
-
-  // shm size for block_sorted_ids/block_topk_weights
-  // both of them requires tb_m * 4 bytes (tb_m * int32 or tb_m * float32)
-  int sh_block_meta_size = tb_m * 4 * 2;
-  int sh_a_size = pipe_stages * (tb_m * tb_k) * 2;
-  int sh_b_size = pipe_stages * (tb_k * tb_n / pack_factor) * 4;
-  int sh_s_size =
-      get_scales_cache_size(th_config, prob_m, prob_n, prob_k, num_bits,
-                            group_size, has_act_order, is_k_full);
-  int sh_g_idx_size = has_act_order && !is_k_full ? pipe_stages * tb_k / 4 : 0;
-  int sh_zp_size = 0;
-  if (has_zp) {
-    if (is_zp_float)
-      sh_zp_size = sh_s_size;
-    else if (num_bits == 4)
-      sh_zp_size = sh_s_size / 4;
-    else if (num_bits == 8)
-      sh_zp_size = sh_s_size / 2;
-  }
-
-  int total_size = sh_a_size + sh_b_size + sh_s_size + sh_zp_size +
-                   sh_g_idx_size + sh_block_meta_size;
-
-  return total_size;
-}
-
-bool is_valid_config(thread_config_t const& th_config, int thread_m_blocks,
-                     int prob_m, int prob_n, int prob_k, int num_bits,
-                     int group_size, bool has_act_order, bool is_k_full,
-                     int has_zp, int is_zp_float, int max_shared_mem) {
-  // Sanity
-  if (th_config.thread_k == -1 || th_config.thread_n == -1 ||
-      th_config.num_threads == -1) {
-    return false;
-  }
-
-  // Verify K/N are divisible by thread K/N
-  if (prob_k % th_config.thread_k != 0 || prob_n % th_config.thread_n != 0) {
-    return false;
-  }
-
-  // Verify min for thread K/N
-  if (th_config.thread_n < min_thread_n || th_config.thread_k < min_thread_k) {
-    return false;
-  }
-
-  // num_threads must be at least 128 (= 4 warps)
-  if (th_config.num_threads < 128) {
-    return false;
-  }
-
-  // Check that pipeline fits into cache
-  int cache_size = get_kernel_cache_size(
-      th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits, group_size,
-      has_act_order, is_k_full, has_zp, is_zp_float);
-  return cache_size <= max_shared_mem;
-}
-
-  #define __GET_IF(W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS, \
-                   M_BLOCK_SIZE_8, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS,       \
-                   NUM_THREADS, IS_ZP_FLOAT)                                  \
-    else if (q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS &&        \
-             thread_n_blocks == THREAD_N_BLOCKS &&                            \
-             thread_k_blocks == THREAD_K_BLOCKS &&                            \
-             m_block_size_8 == M_BLOCK_SIZE_8 &&                              \
-             has_act_order == HAS_ACT_ORDER && has_zp == HAS_ZP &&            \
-             group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS &&    \
-             is_zp_float == IS_ZP_FLOAT) {                                    \
-      kernel = Marlin<scalar_t, W_TYPE.id(), NUM_THREADS, THREAD_M_BLOCKS,    \
-                      THREAD_N_BLOCKS, THREAD_K_BLOCKS, M_BLOCK_SIZE_8,       \
-                      pipe_stages, HAS_ACT_ORDER, HAS_ZP, GROUP_BLOCKS,       \
-                      IS_ZP_FLOAT>;                                           \
-    }
-
-  #define GPTQ_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)              \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, true, false, 0, NUM_THREADS, \
-             false)                                                            \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, true, false, 0,             \
-             NUM_THREADS, false)                                               \
-                                                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, -1,            \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 2,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 4,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, false, 8,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, -1,           \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 2,            \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 4,            \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, false, 8,            \
-             NUM_THREADS, false)
-
-  #define GPTQ_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)  \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, true, false, 0,   \
-             NUM_THREADS, false)                                     \
-                                                                     \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
-             NUM_THREADS, false)                                     \
-                                                                     \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
-             NUM_THREADS, false)                                     \
-                                                                     \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, -1, \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 2,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 4,  \
-             NUM_THREADS, false)                                     \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, false, 8,  \
-             NUM_THREADS, false)
-
-  #define AWQ_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, -1,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 2, NUM_THREADS, \
-             false)                                                            \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 4, NUM_THREADS, \
-             false)                                                            \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 8, NUM_THREADS, \
-             false)                                                            \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, -1,            \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 2,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
-             NUM_THREADS, false)                                               \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 8,             \
-             NUM_THREADS, false)
-
-  #define AWQ_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)  \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
-             NUM_THREADS, false)                                    \
-                                                                    \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
-             NUM_THREADS, false)                                    \
-                                                                    \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, -1, \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 2,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 4,  \
-             NUM_THREADS, false)                                    \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 8,  \
-             NUM_THREADS, false)
-
-  // We currently have 4-bit models only with group_blocks == 4
-  #define HQQ_GET_IF(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)                  \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, false, true, 4, NUM_THREADS, \
-             true)                                                             \
-    __GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
-             NUM_THREADS, true)                                                \
-    __GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
-             NUM_THREADS, true)                                                \
-    __GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
-             NUM_THREADS, true)                                                \
-    __GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, false, true, 4,             \
-             NUM_THREADS, true)
-
-template <typename scalar_t>
-MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
-                                int thread_m_blocks, int thread_n_blocks,
-                                int thread_k_blocks, bool m_block_size_8,
-                                bool has_act_order, bool has_zp,
-                                int group_blocks, int num_threads,
-                                bool is_zp_float) {
-  int num_bits = q_type.size_bits();
-  auto kernel = MarlinDefault;
-  if (false) {
-  }
-  GPTQ_GET_IF_M1(vllm::kU4B8, 8, 8, 256)
-  GPTQ_GET_IF_M1(vllm::kU4B8, 8, 4, 128)
-
-  GPTQ_GET_IF_M234(vllm::kU4B8, 16, 4, 256)
-  GPTQ_GET_IF_M234(vllm::kU4B8, 8, 4, 128)
-
-  GPTQ_GET_IF_M1(vllm::kU8B128, 8, 8, 256)
-  GPTQ_GET_IF_M1(vllm::kU8B128, 8, 4, 128)
-
-  GPTQ_GET_IF_M234(vllm::kU8B128, 16, 4, 256)
-  GPTQ_GET_IF_M234(vllm::kU8B128, 8, 4, 128)
-
-  AWQ_GET_IF_M1(vllm::kU4, 8, 8, 256)
-  AWQ_GET_IF_M1(vllm::kU4, 8, 4, 128)
-
-  AWQ_GET_IF_M234(vllm::kU4, 16, 4, 256)
-  AWQ_GET_IF_M234(vllm::kU4, 8, 4, 128)
-
-  return kernel;
-}
-
-template <typename scalar_t>
-exec_config_t determine_exec_config(const vllm::ScalarType& q_type, int prob_m,
-                                    int prob_n, int prob_k, int thread_m_blocks,
-                                    bool m_block_size_8, int num_bits,
-                                    int group_size, bool has_act_order,
-                                    bool is_k_full, bool has_zp,
-                                    bool is_zp_float, int max_shared_mem) {
-  exec_config_t exec_cfg = exec_config_t{1, thread_config_t{-1, -1, -1}};
-  thread_config_t* thread_configs = thread_m_blocks > 1
-                                        ? large_batch_thread_configs
-                                        : small_batch_thread_configs;
-  int thread_configs_size =
-      thread_m_blocks > 1
-          ? sizeof(large_batch_thread_configs) / sizeof(thread_config_t)
-          : sizeof(small_batch_thread_configs) / sizeof(thread_config_t);
-
-  int count = 0;
-  constexpr int device_max_reg_size = 255 * 1024;
-  for (int i = 0; i < thread_configs_size; i++) {
-    thread_config_t th_config = thread_configs[i];
-
-    if (!is_valid_config(th_config, thread_m_blocks, prob_m, prob_n, prob_k,
-                         num_bits, group_size, has_act_order, is_k_full, has_zp,
-                         is_zp_float, max_shared_mem)) {
-      continue;
-    }
-
-    int cache_size = get_kernel_cache_size(
-        th_config, thread_m_blocks, prob_m, prob_n, prob_k, num_bits,
-        group_size, has_act_order, is_k_full, has_zp, is_zp_float);
-
-    int group_blocks = 0;
-    if (!has_act_order) {
-      group_blocks = group_size == -1 ? -1 : group_size / 16;
-    }
-
-    auto kernel = get_marlin_kernel<scalar_t>(
-        q_type, thread_m_blocks, th_config.thread_n / 16,
-        th_config.thread_k / 16, m_block_size_8, has_act_order, has_zp,
-        group_blocks, th_config.num_threads, is_zp_float);
-
-    if (kernel == MarlinDefault) continue;
-
-    if (thread_m_blocks > 1) {
-      exec_cfg = {1, th_config};
-      break;
-    } else {
-      cudaFuncAttributes attr;
-      cudaFuncGetAttributes(&attr, kernel);
-      int reg_size = max(attr.numRegs, 1) * th_config.num_threads * 4;
-      int allow_count = min(device_max_reg_size / reg_size,
-                            max_shared_mem / (cache_size + 1024));
-      allow_count = max(min(allow_count, 4), 1);
-      if (allow_count > count) {
-        count = allow_count;
-        exec_cfg = {count, th_config};
-      };
-    }
-  }
-
-  return exec_cfg;
-}
-
-template <typename scalar_t>
-void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
-               void* zp, void* g_idx, void* perm, void* a_tmp,
-               void* sorted_token_ids, void* expert_ids,
-               void* num_tokens_past_padded, void* topk_weights,
-               int moe_block_size, int top_k, bool mul_topk_weights, bool is_ep,
-               int prob_m, int prob_n, int prob_k, void* workspace,
-               vllm::ScalarType const& q_type, bool has_act_order,
-               bool is_k_full, bool has_zp, int num_groups, int group_size,
-               int dev, cudaStream_t stream, int thread_k, int thread_n,
-               int sms, bool use_atomic_add, bool use_fp32_reduce,
-               bool is_zp_float) {
-  int thread_m_blocks = div_ceil(moe_block_size, 16);
-  bool m_block_size_8 = moe_block_size == 8;
-
-  if (has_zp) {
-    TORCH_CHECK(
-        q_type == vllm::kU4 || q_type == vllm::kU8,
-        "q_type must be u4 or u8 when has_zp = True. Got = ", q_type.str());
-  } else {
-    TORCH_CHECK(
-        q_type == vllm::kU4B8 || q_type == vllm::kU8B128,
-        "q_type must be uint4b8 or uint8b128 when has_zp = False. Got = ",
-        q_type.str());
-  }
-
-  TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m,
-              ", ", prob_n, ", ", prob_k, "]");
-
-  int group_blocks = 0;
-  if (has_act_order) {
-    if (is_k_full) {
-      TORCH_CHECK(group_size != -1);
-      group_blocks = group_size / 16;
-      TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
-                  " is not divisible by group_blocks = ", group_blocks);
-    } else {
-      TORCH_CHECK(group_size == 0);
-      group_blocks = 0;
-    }
-  } else {
-    if (group_size == -1) {
-      group_blocks = -1;
-    } else {
-      group_blocks = group_size / 16;
-      TORCH_CHECK(prob_k % group_blocks == 0, "prob_k = ", prob_k,
-                  " is not divisible by group_blocks = ", group_blocks);
-    }
-  }
-
-  int num_bits = q_type.size_bits();
-  const int4* A_ptr = (const int4*)A;
-  const int4* B_ptr = (const int4*)B;
-  int4* C_ptr = (int4*)C;
-  int4* C_tmp_ptr = (int4*)C_tmp;
-  const int4* s_ptr = (const int4*)s;
-  const int4* zp_ptr = (const int4*)zp;
-  const int* g_idx_ptr = (const int*)g_idx;
-  const int* perm_ptr = (const int*)perm;
-  int4* a_tmp_ptr = (int4*)a_tmp;
-  const int32_t* sorted_token_ids_ptr = (const int32_t*)sorted_token_ids;
-  const int32_t* expert_ids_ptr = (const int32_t*)expert_ids;
-  const int32_t* num_tokens_past_padded_ptr =
-      (const int32_t*)num_tokens_past_padded;
-  const float* topk_weights_ptr = (const float*)topk_weights;
-  int* locks = (int*)workspace;
-
-  if (has_act_order) {
-    // Permute A columns
-    auto kernel = permute_cols_kernel<8>;
-    if (moe_block_size == 8) {
-    } else if (moe_block_size == 16)
-      kernel = permute_cols_kernel<16>;
-    else if (moe_block_size == 32)
-      kernel = permute_cols_kernel<32>;
-    else if (moe_block_size == 48)
-      kernel = permute_cols_kernel<48>;
-    else if (moe_block_size == 64)
-      kernel = permute_cols_kernel<64>;
-    else
-      TORCH_CHECK(false, "unsupported moe_block_size ", moe_block_size);
-
-    // avoid ">>>" being formatted to "> > >"
-    // clang-format off
-    kernel<<<sms, default_threads, 0, stream>>>(
-        A_ptr, perm_ptr, a_tmp_ptr, sorted_token_ids_ptr, expert_ids_ptr,
-        num_tokens_past_padded_ptr, prob_m, prob_k, top_k);
-    // clang-format on
-    A_ptr = a_tmp_ptr;
-    prob_m = prob_m * top_k;
-    top_k = 1;
-
-    // If we have a full K, then we can run the non-act-order version of Marlin
-    // (since the weight rows are reordered by increasing group ids, and by
-    // having a full K, we have full original groups)
-    if (is_k_full) has_act_order = false;
-  }
-
-  int max_shared_mem = 0;
-  cudaDeviceGetAttribute(&max_shared_mem,
-                         cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
-  TORCH_CHECK(max_shared_mem > 0);
-
-  // Set thread config
-  exec_config_t exec_cfg;
-  thread_config_t thread_tfg;
-  if (thread_k != -1 && thread_n != -1) {
-    thread_tfg = thread_config_t{thread_k, thread_n, default_threads};
-    exec_cfg = exec_config_t{1, thread_tfg};
-    TORCH_CHECK(prob_n % thread_n == 0, "prob_n = ", prob_n,
-                " is not divisible by thread_n = ", thread_n);
-    TORCH_CHECK(prob_k % thread_k == 0, "prob_k = ", prob_k,
-                " is not divisible by thread_k = ", thread_k);
-  } else {
-    // Auto config
-    exec_cfg = determine_exec_config<scalar_t>(
-        q_type, prob_m, prob_n, prob_k, thread_m_blocks, m_block_size_8,
-        num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float,
-        max_shared_mem);
-    thread_tfg = exec_cfg.tb_cfg;
-  }
-
-  int num_threads = thread_tfg.num_threads;
-  thread_k = thread_tfg.thread_k;
-  thread_n = thread_tfg.thread_n;
-  int blocks = sms * exec_cfg.blocks_per_sm;
-  if (exec_cfg.blocks_per_sm > 1)
-    max_shared_mem = max_shared_mem / exec_cfg.blocks_per_sm - 1024;
-
-  int thread_k_blocks = thread_k / 16;
-  int thread_n_blocks = thread_n / 16;
-
-  TORCH_CHECK(is_valid_config(thread_tfg, thread_m_blocks, prob_m, prob_n,
-                              prob_k, num_bits, group_size, has_act_order,
-                              is_k_full, has_zp, is_zp_float, max_shared_mem),
-              "Invalid thread config: thread_m_blocks = ", thread_m_blocks,
-              ", thread_k = ", thread_tfg.thread_k,
-              ", thread_n = ", thread_tfg.thread_n,
-              ", num_threads = ", thread_tfg.num_threads, " for MKN = [",
-              prob_m, ", ", prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
-              ", group_size = ", group_size,
-              ", has_act_order = ", has_act_order, ", is_k_full = ", is_k_full,
-              ", has_zp = ", has_zp, ", is_zp_float = ", is_zp_float,
-              ", max_shared_mem = ", max_shared_mem);
-
-  auto kernel = get_marlin_kernel<scalar_t>(
-      q_type, thread_m_blocks, thread_n_blocks, thread_k_blocks, m_block_size_8,
-      has_act_order, has_zp, group_blocks, num_threads, is_zp_float);
-
-  if (kernel == MarlinDefault) {
-    TORCH_CHECK(false, "Unsupported shapes: MNK = [", prob_m, ", ", prob_n,
-                ", ", prob_k, "]", ", has_act_order = ", has_act_order,
-                ", num_groups = ", num_groups, ", group_size = ", group_size,
-                ", thread_m_blocks = ", thread_m_blocks,
-                ", thread_n_blocks = ", thread_n_blocks,
-                ", thread_k_blocks = ", thread_k_blocks,
-                ", num_bits = ", num_bits);
-  }
-
-  cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize,
-                       max_shared_mem);
-  // avoid ">>>" being formatted to "> > >"
-  // clang-format off
-  kernel<<<blocks, num_threads, max_shared_mem, stream>>>(
-      A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, zp_ptr, g_idx_ptr,
-      sorted_token_ids_ptr, expert_ids_ptr, num_tokens_past_padded_ptr,
-      topk_weights_ptr, top_k, mul_topk_weights, is_ep, num_groups, prob_m,
-      prob_n, prob_k, locks, use_atomic_add, use_fp32_reduce);
-  // clang-format on
-}
-
-}  // namespace MARLIN_NAMESPACE_NAME
-
-torch::Tensor moe_wna16_marlin_gemm(
-    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
-    torch::Tensor& b_q_weight, torch::Tensor& b_scales,
-    std::optional<torch::Tensor> const& b_zeros_or_none,
-    std::optional<torch::Tensor> const& g_idx_or_none,
-    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
-    torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
-    torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
-    int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
-    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
-    int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
-    bool is_zp_float) {
-  vllm::ScalarType const b_q_type = vllm::ScalarType::from_id(b_q_type_id);
-  int pack_factor = 32 / b_q_type.size_bits();
-
-  if (moe_block_size != 8) {
-    TORCH_CHECK(moe_block_size % 16 == 0,
-                "unsupported moe_block_size=", moe_block_size);
-    TORCH_CHECK(moe_block_size >= 16 && moe_block_size <= 64,
-                "unsupported moe_block_size=", moe_block_size);
-  }
-
-  // Verify A
-  TORCH_CHECK(a.size(0) == size_m, "Shape mismatch: a.size(0) = ", a.size(0),
-              ", size_m = ", size_m);
-  TORCH_CHECK(a.size(1) == size_k, "Shape mismatch: a.size(1) = ", a.size(1),
-              ", size_k = ", size_k);
-
-  // Verify B
-  TORCH_CHECK(
-      size_k % MARLIN_NAMESPACE_NAME::tile_size == 0, "size_k = ", size_k,
-      " is not divisible by tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
-  TORCH_CHECK((size_k / MARLIN_NAMESPACE_NAME::tile_size) == b_q_weight.size(1),
-              "Shape mismatch: b_q_weight.size(1) = ", b_q_weight.size(1),
-              ", size_k = ", size_k,
-              ", tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
-  TORCH_CHECK(
-      b_q_weight.size(2) % MARLIN_NAMESPACE_NAME::tile_size == 0,
-      "b_q_weight.size(2) = ", b_q_weight.size(2),
-      " is not divisible by tile_size = ", MARLIN_NAMESPACE_NAME::tile_size);
-  int actual_size_n =
-      (b_q_weight.size(2) / MARLIN_NAMESPACE_NAME::tile_size) * pack_factor;
-  TORCH_CHECK(size_n == actual_size_n, "size_n = ", size_n,
-              ", actual_size_n = ", actual_size_n);
-
-  // Verify device and strides
-  TORCH_CHECK(a.device().is_cuda(), "A is not on GPU");
-  TORCH_CHECK(a.is_contiguous(), "A is not contiguous");
-
-  TORCH_CHECK(b_q_weight.device().is_cuda(), "b_q_weight is not on GPU");
-  TORCH_CHECK(b_q_weight.is_contiguous(), "b_q_weight is not contiguous");
-
-  TORCH_CHECK(b_scales.device().is_cuda(), "b_scales is not on GPU");
-  TORCH_CHECK(b_scales.is_contiguous(), "b_scales is not contiguous");
-
-  // thread_k: `k` size of a thread_tile in `weights` (can usually be left as
-  // auto -1)
-  int thread_k = -1;
-  // thread_n: `n` size of a thread_tile in `weights` (can usually be left as
-  // auto -1)
-  int thread_n = -1;
-  // sms: number of SMs to use for the kernel
-  int sms = -1;
-  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, a.get_device());
-
-  // Alloc buffers
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
-  auto options = torch::TensorOptions().dtype(a.dtype()).device(a.device());
-  torch::Tensor c;
-  if (c_or_none.has_value()) {
-    c = c_or_none.value();
-    TORCH_CHECK(c.device().is_cuda(), "c is not on GPU");
-    TORCH_CHECK(c.is_contiguous(), "c is not contiguous");
-    TORCH_CHECK(c.size(0) == size_m * top_k,
-                "Shape mismatch: c.size(0) = ", c.size(0),
-                ", size_m * topk = ", size_m * top_k);
-    TORCH_CHECK(c.size(1) == size_n, "Shape mismatch: c.size(1) = ", c.size(1),
-                ", size_n = ", size_n);
-  } else {
-    c = torch::empty({size_m * top_k, size_n}, options);
-  }
-
-  // Alloc C tmp buffer that is going to be used for the global reduce
-  torch::Tensor c_tmp;
-  auto options_fp32 =
-      torch::TensorOptions().dtype(at::kFloat).device(a.device());
-  if (use_fp32_reduce && !use_atomic_add) {
-    // max num of threadblocks is sms * 4
-    long max_c_tmp_size = min(
-        (long)size_n * sorted_token_ids.size(0),
-        (long)sms * 4 * moe_block_size * MARLIN_NAMESPACE_NAME::max_thread_n);
-    if (moe_block_size == 8) max_c_tmp_size *= 2;
-    c_tmp = torch::empty({max_c_tmp_size}, options_fp32);
-  } else {
-    c_tmp = torch::empty({0}, options_fp32);
-  }
-
-  // Detect groupsize and act_order
-  int num_groups = -1;
-  int group_size = -1;
-
-  int rank = b_scales.sizes().size();
-  TORCH_CHECK(rank == 3, "b_scales rank = ", rank, " is not 3");
-  TORCH_CHECK(b_scales.size(2) == size_n, "b_scales dim 2 = ", b_scales.size(2),
-              " is not size_n = ", size_n);
-  num_groups = b_scales.size(1);
-
-  torch::Tensor g_idx, perm, a_tmp;
-  ;
-  if (g_idx_or_none.has_value() && perm_or_none.has_value()) {
-    g_idx = g_idx_or_none.value();
-    perm = perm_or_none.value();
-
-    TORCH_CHECK(g_idx.device().is_cuda(), "g_idx is not on GPU");
-    TORCH_CHECK(g_idx.is_contiguous(), "g_idx is not contiguous");
-    TORCH_CHECK(perm.device().is_cuda(), "perm is not on GPU");
-    TORCH_CHECK(perm.is_contiguous(), "perm is not contiguous");
-
-    // Verify g_idx and perm
-    TORCH_CHECK((g_idx.size(-1) == 0 && perm.size(-1) == 0) ||
-                    (g_idx.size(-1) == size_k && perm.size(-1) == size_k),
-                "Unexpected g_idx.size(-1) = ", g_idx.size(-1),
-                " and perm.size(-1) = ", perm.size(-1),
-                ", where size_k = ", size_k);
-  } else {
-    g_idx = torch::empty({0}, options);
-    perm = torch::empty({0}, options);
-    a_tmp = torch::empty({0}, options);
-  }
-  bool has_act_order = g_idx.size(-1) > 0 && perm.size(-1) > 0;
-
-  if (has_act_order) {
-    a_tmp = torch::empty({size_m * top_k, size_k}, options);
-    if (is_k_full) {
-      TORCH_CHECK(num_groups > 1, "For act_order, num_groups must be > 1");
-      TORCH_CHECK(size_k % num_groups == 0, "size_k = ", size_k,
-                  ", is not divisible by num_groups = ", num_groups);
-      group_size = size_k / num_groups;
-    } else {
-      group_size = 0;
-    }
-
-  } else {
-    a_tmp = torch::empty({0}, options);
-    if (num_groups > 1) {
-      TORCH_CHECK(
-          size_k % num_groups == 0, "size_k = ", size_k,
-          ", is not divisible by b_scales.size(1) = ", b_scales.size(1));
-      group_size = size_k / num_groups;
-    } else {
-      group_size = -1;
-    }
-  }
-
-  torch::Tensor b_zeros;
-  if (b_zeros_or_none.has_value()) {
-    b_zeros = b_zeros_or_none.value();
-    TORCH_CHECK(b_zeros.device().is_cuda(), "b_zeros is not on GPU");
-    TORCH_CHECK(b_zeros.is_contiguous(), "b_zeros is not contiguous");
-  } else {
-    b_zeros = torch::empty({0}, options);
-  }
-  bool has_zp = b_zeros.size(-1) > 0;
-
-  if (has_zp) {
-    TORCH_CHECK(
-        b_q_type == vllm::kU4,
-        "b_q_type must be u4 when has_zp = True. Got = ", b_q_type.str());
-  } else {
-    TORCH_CHECK(
-        b_q_type == vllm::kU4B8 || b_q_type == vllm::kU8B128,
-        "b_q_type must be uint4b8 or uint8b128 when has_zp = False. Got = ",
-        b_q_type.str());
-  }
-
-  if (has_zp && is_zp_float) {
-    TORCH_CHECK(a.scalar_type() == at::ScalarType::Half,
-                "Computation type must be float16 (half) when using float zero "
-                "points.");
-  }
-
-  // Verify b_zeros
-  if (has_zp) {
-    int rank = b_zeros.sizes().size();
-    TORCH_CHECK(rank == 3, "b_zeros rank = ", rank, " is not 3");
-    if (is_zp_float) {
-      TORCH_CHECK(b_zeros.size(2) == size_n,
-                  "b_zeros dim 2 = ", b_zeros.size(2),
-                  " is not size_n = ", size_n);
-      TORCH_CHECK(num_groups == b_zeros.size(1),
-                  "b_zeros dim 1 = ", b_zeros.size(1),
-                  " is not num_groups = ", num_groups);
-      TORCH_CHECK(num_groups != -1, "num_groups must be != -1");
-    } else {
-      TORCH_CHECK(b_zeros.size(1) == num_groups,
-                  "b_zeros dim 1 = ", b_zeros.size(1),
-                  " is not num_groups = ", num_groups);
-      TORCH_CHECK(b_zeros.size(2) == size_n / pack_factor,
-                  "b_zeros dim 2 = ", b_zeros.size(2),
-                  " is not size_n / pack_factor = ", size_n / pack_factor);
-    }
-  }
-
-  // Verify workspace size
-  TORCH_CHECK(size_n % MARLIN_NAMESPACE_NAME::min_thread_n == 0,
-              "size_n = ", size_n, ", is not divisible by min_thread_n = ",
-              MARLIN_NAMESPACE_NAME::min_thread_n);
-
-  int max_n_tiles = size_n / MARLIN_NAMESPACE_NAME::min_thread_n;
-  int min_workspace_size = min(
-      max_n_tiles * (int)(sorted_token_ids.size(0) / moe_block_size), sms * 4);
-  TORCH_CHECK(workspace.numel() >= min_workspace_size,
-              "workspace.numel = ", workspace.numel(),
-              " is below min_workspace_size = ", min_workspace_size);
-
-  int dev = a.get_device();
-  if (a.scalar_type() == at::ScalarType::Half) {
-    MARLIN_NAMESPACE_NAME::marlin_mm<half>(
-        a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
-        c_tmp.data_ptr<float>(), b_scales.data_ptr<at::Half>(),
-        b_zeros.data_ptr(), g_idx.data_ptr(), perm.data_ptr(),
-        a_tmp.data_ptr<at::Half>(), sorted_token_ids.data_ptr(),
-        expert_ids.data_ptr(), num_tokens_past_padded.data_ptr(),
-        topk_weights.data_ptr(), moe_block_size, top_k, mul_topk_weights, is_ep,
-        size_m, size_n, size_k, workspace.data_ptr(), b_q_type, has_act_order,
-        is_k_full, has_zp, num_groups, group_size, dev,
-        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
-        use_atomic_add, use_fp32_reduce, is_zp_float);
-  } else if (a.scalar_type() == at::ScalarType::BFloat16) {
-    MARLIN_NAMESPACE_NAME::marlin_mm<nv_bfloat16>(
-        a.data_ptr<at::BFloat16>(), b_q_weight.data_ptr(),
-        c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(),
-        b_scales.data_ptr<at::BFloat16>(), b_zeros.data_ptr(), g_idx.data_ptr(),
-        perm.data_ptr(), a_tmp.data_ptr<at::BFloat16>(),
-        sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
-        num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
-        moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
-        workspace.data_ptr(), b_q_type, has_act_order, is_k_full, has_zp,
-        num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
-        thread_k, thread_n, sms, use_atomic_add, use_fp32_reduce, is_zp_float);
-  } else {
-    TORCH_CHECK(false,
-                "moe_wna16_marlin_gemm only supports bfloat16 and float16");
-  }
-
-  return c;
-}
-
-#endif
-
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
-  m.impl("moe_wna16_marlin_gemm", &moe_wna16_marlin_gemm);
-}

csrc/moe/moe_wna16.cu

@@ -13,6 +13,7 @@
 template <typename scalar_t, int bit, int GROUPS>
 __global__ void moe_wna16_gemm_kernel(
     const scalar_t* __restrict__ input, scalar_t* __restrict__ output,
+
     const uint32_t* __restrict__ qweight, const scalar_t* __restrict__ scales,
     const uint32_t* __restrict__ qzeros,
 
@@ -53,6 +54,8 @@ __global__ void moe_wna16_gemm_kernel(
       if (token_index / top_k >= size_m) break;
 
       num_valid_tokens = m + 1;
+      if (blockIdx.z == 0 && offset_n < size_n)
+        output[token_index * size_n + offset_n] = Dtype::int2num(0);
 
       if (expert_id != -1) {
         int k_per_thread = DIVIDE(BLOCK_SIZE_K, BLOCK_SIZE_N);
@@ -281,7 +284,8 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              int64_t BLOCK_SIZE_M, int64_t BLOCK_SIZE_N,
                              int64_t BLOCK_SIZE_K, int64_t bit) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
-  output.zero_();
+  auto options =
+      torch::TensorOptions().dtype(input.dtype()).device(input.device());
 
   const int num_experts = b_qweight.size(0);
   const int size_m = input.size(0);
@@ -298,9 +302,9 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
   const uint32_t* b_qzeros_ptr;
   if (b_qzeros.has_value())
     b_qzeros_ptr = (const uint32_t*)b_qzeros.value().data_ptr<uint8_t>();
-  const float* topk_weights_ptr = nullptr;
+  const float* topk_weights_ptr;
   if (topk_weights.has_value())
-    topk_weights_ptr = (const float*)topk_weights.value().data_ptr<float>();
+    topk_weights_ptr = (const float*)topk_weights.value().data_ptr();
 
   int groups_per_block_row = BLOCK_SIZE_K / group_size;
   TORCH_CHECK(bit == 4 || bit == 8, "bit must be 4 or 8");

csrc/moe/torch_bindings.cpp

@@ -43,17 +43,14 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
   m.impl("moe_wna16_gemm", torch::kCUDA, &moe_wna16_gemm);
 
   m.def(
-      "moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none,"
-      "Tensor! b_q_weight, Tensor! b_scales, Tensor? b_zeros_or_none,"
-      "Tensor? g_idx_or_none, Tensor? perm_or_none, Tensor! workspace,"
-      "Tensor sorted_token_ids,"
-      "Tensor! expert_ids, Tensor! num_tokens_past_padded,"
-      "Tensor! topk_weights, int moe_block_size, int top_k, "
-      "bool mul_topk_weights, bool is_ep, int b_q_type_id,"
-      "int size_m, int size_n, int size_k,"
-      "bool is_full_k, bool use_atomic_add,"
-      "bool use_fp32_reduce, bool is_zp_float) -> Tensor");
-
+      "marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "
+      "Tensor! topk_weights, Tensor! topk_ids, Tensor! b_scales, Tensor! "
+      "b_zeros, Tensor! g_idx, Tensor! perm, Tensor! workspace, "
+      "int b_q_type, SymInt size_m, "
+      "SymInt size_n, SymInt size_k, bool is_k_full, int num_experts, int "
+      "topk, "
+      "int moe_block_size, bool replicate_input, bool apply_weights)"
+      " -> Tensor");
   // conditionally compiled so impl registration is in source file
 
 #endif

csrc/ops.h

@@ -128,12 +128,6 @@ void advance_step_flashinfer(
     torch::Tensor& paged_kv_indices, torch::Tensor& paged_kv_indptr,
     torch::Tensor& paged_kv_last_page_len, torch::Tensor& block_table_bounds);
 
-void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
-                        torch::Tensor const& q_pe,
-                        torch::Tensor const& kv_c_and_k_pe_cache,
-                        torch::Tensor const& seq_lens,
-                        torch::Tensor const& page_table, double scale);
-
 torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor);
 
 #ifndef USE_ROCM

csrc/quantization/cutlass_w8a8/moe/moe_data.cu

@@ -46,26 +46,14 @@ __global__ void compute_expert_offsets(
 }
 
 __global__ void compute_arg_sorts(const int* __restrict__ topk_ids,
-                                  const int32_t* __restrict__ expert_offsets,
                                   int32_t* input_permutation,
                                   int32_t* output_permutation,
                                   int32_t* atomic_buffer, const int topk_length,
                                   const int topk) {
-  int const blk_expert_id = blockIdx.x;
-  int const num_experts = gridDim.x;
-  int32_t const num_tokens = expert_offsets[num_experts];
+  int expert_id = blockIdx.x;
 
   for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
-    int const expert_id = topk_ids[i];
-    if (expert_id == -1 && blockIdx.x == 0) {
-      // output_permutation is used to re-order the moe outputs. It is
-      // used as c2 = c2[c_map], where c2 is a torch.tensor that is the
-      // output of the cutlass kernels and c_map is the output_permutation.
-      // c2 is initialized to zeros, therefore by setting the output_permutation
-      // to num_tokens, we are guaranteed to fill the moe outputs to zero
-      // for "invalid" topk_ids.
-      output_permutation[i] = num_tokens;
-    } else if (expert_id == blk_expert_id) {
+    if (topk_ids[i] == expert_id) {
       int start = atomicAdd(&atomic_buffer[expert_id], 1);
       input_permutation[start] = i / topk;
       output_permutation[i] = start;
@@ -95,7 +83,6 @@ void get_cutlass_moe_mm_data_caller(
       static_cast<int32_t*>(atomic_buffer.data_ptr()), num_experts);
   compute_arg_sorts<<<num_experts, num_threads, 0, stream>>>(
       static_cast<const int32_t*>(topk_ids.data_ptr()),
-      static_cast<const int32_t*>(expert_offsets.data_ptr()),
       static_cast<int32_t*>(input_permutation.data_ptr()),
       static_cast<int32_t*>(output_permutation.data_ptr()),
       static_cast<int32_t*>(atomic_buffer.data_ptr()), topk_ids.numel(),

csrc/quantization/cutlass_w8a8/scaled_mm_c2x_sm89_fp8_dispatch.cuh

@@ -336,7 +336,7 @@ inline void cutlass_gemm_sm89_fp8_dispatch(torch::Tensor& out,
 
   uint32_t const m = a.size(0);
   uint32_t const mp2 =
-      std::max(static_cast<uint32_t>(16), next_pow_2(m));  // next power of 2
+      std::max(static_cast<uint32_t>(32), next_pow_2(m));  // next power of 2
 
   if (mp2 <= 16) {
     // M in [1, 16]

csrc/quantization/cutlass_w8a8/scaled_mm_c2x_sm89_int8_dispatch.cuh

@@ -321,7 +321,7 @@ inline void cutlass_gemm_sm89_int8_dispatch(torch::Tensor& out,
 
   uint32_t const m = a.size(0);
   uint32_t const mp2 =
-      std::max(static_cast<uint32_t>(16), next_pow_2(m));  // next power of 2
+      std::max(static_cast<uint32_t>(32), next_pow_2(m));  // next power of 2
 
   if (mp2 <= 16) {
     // M in [1, 16]

csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu

@@ -134,7 +134,7 @@ typename T::Gemm::Arguments args_from_options(
   using StrideB = typename T::StrideB;
   using StrideD = typename T::StrideD;
   using Sm100BlkScaledConfig =
-      typename T::Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
+      typename T::Gemm::GemmKernel::CollectiveMainloop::Sm100BlkScaledConfig;
 
   int m = static_cast<int>(M);
   int n = static_cast<int>(N);

csrc/quantization/gptq_marlin/marlin.cuh

@@ -9,11 +9,7 @@
 #include <cuda_runtime.h>
 #include <iostream>
 
-#ifndef MARLIN_NAMESPACE_NAME
-  #define MARLIN_NAMESPACE_NAME marlin
-#endif
-
-namespace MARLIN_NAMESPACE_NAME {
+namespace marlin {
 
 // Marlin params
 
@@ -27,7 +23,6 @@ static constexpr int pipe_stages =
 
 static constexpr int min_thread_n = 64;
 static constexpr int min_thread_k = 64;
-static constexpr int max_thread_n = 256;
 
 static constexpr int tile_size = 16;
 static constexpr int max_par = 16;
@@ -89,4 +84,4 @@ __device__ inline void cp_async_wait() {
 
 #endif
 
-}  // namespace MARLIN_NAMESPACE_NAME
+}  // namespace marlin

csrc/quantization/gptq_marlin/marlin_dtypes.cuh

@@ -5,11 +5,7 @@
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
 
-#ifndef MARLIN_NAMESPACE_NAME
-  #define MARLIN_NAMESPACE_NAME marlin
-#endif
-
-namespace MARLIN_NAMESPACE_NAME {
+namespace marlin {
 
 template <typename scalar_t>
 class ScalarType {};
@@ -58,7 +54,7 @@ class ScalarType<nv_bfloat16> {
   using FragS = Vec<nv_bfloat162, 1>;
   using FragZP = Vec<nv_bfloat162, 4>;
 
-#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
   static __device__ float inline num2float(const nv_bfloat16 x) {
     return __bfloat162float(x);
   }
@@ -78,6 +74,6 @@ class ScalarType<nv_bfloat16> {
 #endif
 };
 
-}  // namespace MARLIN_NAMESPACE_NAME
+}  // namespace marlin
 
 #endif

csrc/rocm/ops.h

@@ -2,15 +2,6 @@
 
 #include <torch/all.h>
 
-torch::Tensor LLMM1(at::Tensor& in_a, at::Tensor& in_b,
-                    const int64_t rows_per_block);
-
-torch::Tensor wvSplitK(at::Tensor& in_a, at::Tensor& in_b,
-                       const int64_t CuCount);
-
-void wvSplitKQ(at::Tensor& in_a, at::Tensor& in_b, at::Tensor& out_c,
-               at::Tensor& scale_a, at::Tensor& scale_b, const int64_t CuCount);
-
 void paged_attention(torch::Tensor& out, torch::Tensor& exp_sums,
                      torch::Tensor& max_logits, torch::Tensor& tmp_out,
                      torch::Tensor& query, torch::Tensor& key_cache,

csrc/rocm/torch_bindings.cpp

@@ -14,24 +14,6 @@
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, rocm_ops) {
   // vLLM custom ops for rocm
 
-  // Custom gemm op for matrix-vector multiplication
-  rocm_ops.def(
-      "LLMM1(Tensor in_a, Tensor in_b, int rows_per_block) -> "
-      "Tensor");
-  rocm_ops.impl("LLMM1", torch::kCUDA, &LLMM1);
-
-  // Custom gemm op for skinny matrix-matrix multiplication
-  rocm_ops.def(
-      "wvSplitK(Tensor in_a, Tensor in_b, int CuCount) -> "
-      "Tensor");
-  rocm_ops.impl("wvSplitK", torch::kCUDA, &wvSplitK);
-
-  // wvSplitK for fp8
-  rocm_ops.def(
-      "wvSplitKQ(Tensor in_a, Tensor in_b, Tensor! out_c, Tensor scale_a, "
-      "          Tensor scale_b, int CuCount) -> ()");
-  rocm_ops.impl("wvSplitKQ", torch::kCUDA, &wvSplitKQ);
-
   // Custom attention op
   // Compute the attention between an input query and the cached
   // keys/values using PagedAttention.

csrc/torch_bindings.cpp

@@ -130,13 +130,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       ") -> ()");
   ops.impl("advance_step_flashinfer", torch::kCUDA, &advance_step_flashinfer);
 
-  // Compute MLA decode using cutlass.
-  ops.def(
-      "cutlass_mla_decode(Tensor! out, Tensor q_nope, Tensor q_pe,"
-      "                   Tensor kv_c_and_k_pe_cache, Tensor seq_lens,"
-      "                   Tensor page_table, float scale) -> ()");
-  ops.impl("cutlass_mla_decode", torch::kCUDA, &cutlass_mla_decode);
-
   // Layernorm
   // Apply Root Mean Square (RMS) Normalization to the input tensor.
   ops.def(

docker/Dockerfile

@@ -162,9 +162,6 @@ ENV UV_HTTP_TIMEOUT=500
 COPY requirements/lint.txt requirements/lint.txt
 COPY requirements/test.txt requirements/test.txt
 COPY requirements/dev.txt requirements/dev.txt
-# Workaround for #17068
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system mamba-ssm==2.2.4 --no-build-isolation
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/dev.txt
 #################### DEV IMAGE ####################
@@ -243,8 +240,6 @@ if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
     uv pip install --system https://github.com/flashinfer-ai/flashinfer/releases/download/v0.2.1.post2/flashinfer_python-0.2.1.post2+cu124torch2.6-cp38-abi3-linux_x86_64.whl ; \
 fi
 COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
 
 # Although we build Flashinfer with AOT mode, there's still
 # some issues w.r.t. JIT compilation. Therefore we need to
@@ -268,9 +263,6 @@ ADD . /vllm-workspace/
 ENV UV_HTTP_TIMEOUT=500
 
 # install development dependencies (for testing)
-# Workaround for #17068
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system mamba-ssm==2.2.4 --no-build-isolation
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/dev.txt
 
@@ -297,7 +289,6 @@ RUN mv vllm test_docs/
 #################### OPENAI API SERVER ####################
 # base openai image with additional requirements, for any subsequent openai-style images
 FROM vllm-base AS vllm-openai-base
-ARG TARGETPLATFORM
 
 # This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
 # Reference: https://github.com/astral-sh/uv/pull/1694

docker/Dockerfile.cpu

@@ -121,7 +121,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 ADD ./tests/ ./tests/
 ADD ./examples/ ./examples/
 ADD ./benchmarks/ ./benchmarks/
-ADD ./vllm/collect_env.py .
 
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \

docker/Dockerfile.nightly_torch

@@ -1,307 +0,0 @@
-# The vLLM Dockerfile is used to construct vLLM image against torch nightly that can be directly used for testing
-
-# for torch nightly, cuda >=12.6 is required,
-# use 12.8 due to FlashAttention issue with cuda 12.6 (https://github.com/vllm-project/vllm/issues/15435#issuecomment-2775924628)
-ARG CUDA_VERSION=12.8.0
-#
-#################### BASE BUILD IMAGE ####################
-# prepare basic build environment
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
-ARG CUDA_VERSION=12.8.0
-ARG PYTHON_VERSION=3.12
-ARG TARGETPLATFORM
-ENV DEBIAN_FRONTEND=noninteractive
-# Install Python and other dependencies
-RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
-    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
-    && apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl sudo \
-    && add-apt-repository ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
-    && python3 --version \
-    && python3 -m pip --version
-# Install uv for faster pip installs
-RUN --mount=type=cache,target=/root/.cache/uv \
-    python3 -m pip install uv
-
-# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
-# Reference: https://github.com/astral-sh/uv/pull/1694
-ENV UV_HTTP_TIMEOUT=500
-
-# Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
-# as it was causing spam when compiling the CUTLASS kernels
-RUN apt-get install -y gcc-10 g++-10
-RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
-RUN <<EOF
-gcc --version
-EOF
-
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
-WORKDIR /workspace
-
-# install build and runtime dependencies
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-
-# install build and runtime dependencies without stable torch version
-RUN python3 use_existing_torch.py
-
-# install torch nightly
-ARG PINNED_TORCH_VERSION
-RUN --mount=type=cache,target=/root/.cache/uv \
-    if [ -n "$PINNED_TORCH_VERSION" ]; then \
-      pkgs="$PINNED_TORCH_VERSION"; \
-    else \
-      pkgs="torch torchaudio torchvision"; \
-    fi && \
-    uv pip install --system $pkgs --index-url https://download.pytorch.org/whl/nightly/cu128
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system numba==0.61.2
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-uv pip install --system -r requirements/common.txt
-
-# must put before installing xformers, so it can install the correct version of xfomrers.
-ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
-
-# Build xformers with cuda and torch nightly
-# following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# todo(elainewy): cache xformers build result for faster build
-ARG max_jobs=16
-ENV MAX_JOBS=${max_jobs}
-ARG XFORMERS_COMMIT=f2de641ef670510cadab099ce6954031f52f191c
-
-ENV CCACHE_DIR=/root/.cache/ccache
-RUN --mount=type=cache,target=/root/.cache/ccache \
-     --mount=type=cache,target=/root/.cache/uv \
-    echo 'git clone xformers...' \
-    && git clone https://github.com/facebookresearch/xformers.git --recursive \
-    && cd xformers \
-    && git checkout ${XFORMERS_COMMIT} \
-    && git submodule update --init --recursive \
-    && echo 'finish git clone xformers...' \
-    && rm -rf build \
-    && python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
-    && cd .. \
-    && rm -rf xformers
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system xformers-dist/*.whl --verbose
-
-# build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
-# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-RUN cat torch_build_versions.txt
-
-# cuda arch list used by torch
-# can be useful for `test`
-# explicitly set the list to avoid issues with torch 2.2
-# see https://github.com/pytorch/pytorch/pull/123243
-
-# Override the arch list for flash-attn to reduce the binary size
-ARG vllm_fa_cmake_gpu_arches='80-real;90-real'
-ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
-#################### BASE BUILD IMAGE ####################
-
-#################### WHEEL BUILD IMAGE ####################
-FROM base AS build
-ARG TARGETPLATFORM
-
-# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
-# Reference: https://github.com/astral-sh/uv/pull/1694
-ENV UV_HTTP_TIMEOUT=500
-
-COPY . .
-
-RUN python3 use_existing_torch.py
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/build.txt
-
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
-
-# Max jobs used by Ninja to build extensions
-ARG max_jobs=16
-ENV MAX_JOBS=${max_jobs}
-ARG nvcc_threads=2
-ENV NVCC_THREADS=$nvcc_threads
-
-ARG USE_SCCACHE
-ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
-ARG SCCACHE_REGION_NAME=us-west-2
-ARG SCCACHE_S3_NO_CREDENTIALS=0
-
-# if USE_SCCACHE is set, use sccache to speed up compilation
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,source=.git,target=.git \
-    if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..." \
-        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
-        && tar -xzf sccache.tar.gz \
-        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
-        && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
-        && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
-        && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
-        && export SCCACHE_IDLE_TIMEOUT=0 \
-        && export CMAKE_BUILD_TYPE=Release \
-        && sccache --show-stats \
-        && python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38 \
-        && sccache --show-stats; \
-    fi
-
-ENV CCACHE_DIR=/root/.cache/ccache
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,source=.git,target=.git  \
-    if [ "$USE_SCCACHE" != "1" ]; then \
-        # Clean any existing CMake artifacts
-        rm -rf .deps && \
-        mkdir -p .deps && \
-        python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38; \
-    fi
-
-#################### WHEEL BUILD IMAGE ####################
-
-################### VLLM INSTALLED IMAGE ####################
-# Setup clean environment for vLLM and its dependencies for test and api server using ubuntu22.04 with AOT flashinfer
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 AS vllm-base
-# prepare for environment starts
-ARG CUDA_VERSION=12.8.0
-ARG PYTHON_VERSION=3.12
-WORKDIR /vllm-workspace
-ENV DEBIAN_FRONTEND=noninteractive
-ARG TARGETPLATFORM
-
-RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
-    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
-
-# Install Python and other dependencies
-RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
-    && echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
-    && apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim python3-pip \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
-    && add-apt-repository ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv libibverbs-dev \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION} \
-    && python3 --version && python3 -m pip --version
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    python3 -m pip install uv
-
-# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
-# Reference: https://github.com/astral-sh/uv/pull/1694
-ENV UV_HTTP_TIMEOUT=500
-
-# Workaround for https://github.com/openai/triton/issues/2507 and
-# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
-# this won't be needed for future versions of this docker image
-# or future versions of triton.
-RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
-
-# get the nightly torch version used in the build to make sure the version is the same
-COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128
-
-# install the vllm wheel
-RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/vllm-dist \
-    --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system vllm-dist/*.whl --verbose
-
-# install xformers again for the new environment
-RUN --mount=type=bind,from=base,src=/workspace/xformers-dist,target=/vllm-workspace/xformers-dist \
-    --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system /vllm-workspace/xformers-dist/*.whl --verbose
-
-ARG torch_cuda_arch_list='8.0;8.6;8.9;9.0'
-
-# install package for build flashinfer
-# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
-RUN pip install setuptools==75.6.0 packaging==23.2 ninja==1.11.1.3 build==1.2.2.post1
-
-
-# build flashinfer for torch nightly from source around 10 mins
-# release version: v0.2.2.post1
-# todo(elainewy): cache flashinfer build result for faster build
-ENV CCACHE_DIR=/root/.cache/ccache
-RUN --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=cache,target=/root/.cache/uv \
-    echo "git clone flashinfer..." \
-    && git clone --recursive https://github.com/flashinfer-ai/flashinfer.git \
-    && cd flashinfer \
-    && git checkout v0.2.2.post1 \
-    && git submodule update --init --recursive \
-    && echo "finish git clone flashinfer..." \
-    && rm -rf build \
-    && export TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list} \
-    && FLASHINFER_ENABLE_AOT=1 python3 setup.py bdist_wheel --dist-dir=../flashinfer-dist --verbose \
-    && cd .. \
-    && rm -rf flashinfer
-
-# install flashinfer
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system flashinfer-dist/*.whl --verbose
-
-# install common packages
-COPY requirements/common.txt requirements/common.txt
-COPY use_existing_torch.py use_existing_torch.py
-COPY pyproject.toml pyproject.toml
-
-COPY examples examples
-COPY benchmarks benchmarks
-COPY ./vllm/collect_env.py .
-
-RUN python3 use_existing_torch.py
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
-
-################### VLLM INSTALLED IMAGE ####################
-
-
-#################### UNITTEST IMAGE #############################
-FROM vllm-base as test
-COPY tests/ tests/
-
-# install build and runtime dependencies without stable torch version
-COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
-
-# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
-# Reference: https://github.com/astral-sh/uv/pull/1694
-ENV UV_HTTP_TIMEOUT=500
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -e tests/vllm_test_utils
-
-# enable fast downloads from hf (for testing)
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system hf_transfer
-ENV HF_HUB_ENABLE_HF_TRANSFER 1
-
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/nightly_torch_test.txt
-
-#################### UNITTEST IMAGE #############################
-

docker/Dockerfile.ppc64le

@@ -126,16 +126,13 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 FROM base-builder AS cv-builder
 
 ARG MAX_JOBS
-ARG OPENCV_VERSION=86
-# patch for version 4.11.0.86
-ARG OPENCV_PATCH=97f3f39
+ARG OPENCV_VERSION=84
 ARG ENABLE_HEADLESS=1
 RUN --mount=type=cache,target=/root/.cache/uv \
     source /opt/rh/gcc-toolset-13/enable && \
     git clone --recursive https://github.com/opencv/opencv-python.git -b ${OPENCV_VERSION} && \
     cd opencv-python && \
-    sed -i -E -e 's/"setuptools.+",/"setuptools",/g' pyproject.toml && \
-    cd opencv && git cherry-pick --no-commit $OPENCV_PATCH && cd .. && \
+    sed -i 's/"setuptools==59.2.0",/"setuptools<70.0",/g' pyproject.toml && \
     python -m build --wheel --installer=uv --outdir /opencvwheels/
 
 ###############################################################
@@ -151,15 +148,9 @@ COPY --from=arrow-builder /tmp/control /dev/null
 COPY --from=cv-builder /tmp/control /dev/null
 
 ARG VLLM_TARGET_DEVICE=cpu
-ARG GRPC_PYTHON_BUILD_SYSTEM_OPENSSL=1
 
 # this step installs vllm and populates uv cache
 # with all the transitive dependencies
-RUN --mount=type=cache,target=/root/.cache/uv \
-    source /opt/rh/gcc-toolset-13/enable && \
-    git clone https://github.com/huggingface/xet-core.git && cd xet-core/hf_xet/ && \
-    uv pip install maturin && \
-    uv build --wheel --out-dir /hf_wheels/
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
     --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
@@ -168,7 +159,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     source /opt/rh/gcc-toolset-13/enable && \
     uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl && \
     sed -i -e 's/.*torch.*//g' /src/pyproject.toml /src/requirements/*.txt && \
-    uv pip install pandas pythran pybind11 /hf_wheels/*.whl && \
+    uv pip install pandas pythran pybind11 && \
     # sentencepiece.pc is in some pkgconfig inside uv cache
     export PKG_CONFIG_PATH=$(find / -type d -name "pkgconfig" 2>/dev/null | tr '\n' ':') && \
     uv pip install -r /src/requirements/common.txt -r /src/requirements/cpu.txt -r /src/requirements/build.txt --no-build-isolation && \
@@ -256,9 +247,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,from=torch-builder,source=/torchwheels/,target=/torchwheels/,ro \
     --mount=type=bind,from=arrow-builder,source=/arrowwheels/,target=/arrowwheels/,ro \
     --mount=type=bind,from=cv-builder,source=/opencvwheels/,target=/opencvwheels/,ro \
-    --mount=type=bind,from=vllmcache-builder,source=/hf_wheels/,target=/hf_wheels/,ro \
     --mount=type=bind,from=vllmcache-builder,source=/vllmwheel/,target=/vllmwheel/,ro \
-    HOME=/root uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl /hf_wheels/*.whl /vllmwheel/*.whl
+    HOME=/root uv pip install /opencvwheels/*.whl /arrowwheels/*.whl /torchwheels/*.whl /vllmwheel/*.whl
 
 COPY ./ /workspace/vllm
 WORKDIR /workspace/vllm

docker/Dockerfile.rocm_base

@@ -12,7 +12,7 @@ ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
 ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
 ARG FA_BRANCH="1a7f4dfa"
 ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
-ARG AITER_BRANCH="7e1ed08"
+ARG AITER_BRANCH="8970b25b"
 ARG AITER_REPO="https://github.com/ROCm/aiter.git"
 
 FROM ${BASE_IMAGE} AS base

docker/Dockerfile.s390x

@@ -58,7 +58,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     cd ../../python && \
     export PYARROW_PARALLEL=4 && \
     export ARROW_BUILD_TYPE=release && \
-    uv pip install -r requirements-build.txt && \
+    uv pip install -r requirements/build.txt && \
     python setup.py build_ext --build-type=$ARROW_BUILD_TYPE --bundle-arrow-cpp bdist_wheel
 
 FROM python-install AS numa-build
@@ -96,22 +96,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install -v torch==${TORCH_VERSION} --extra-index-url https://download.pytorch.org/whl/nightly/cpu && \
     python setup.py bdist_wheel
 
-FROM python-install AS hf-xet-builder
-# Install hf-xet
-WORKDIR /tmp
-ENV CARGO_HOME=/root/.cargo
-ENV RUSTUP_HOME=/root/.rustup
-ENV PATH="$CARGO_HOME/bin:$RUSTUP_HOME/bin:$PATH"
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,from=rust,source=/root/.cargo,target=/root/.cargo,rw \
-    --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
-    git clone https://github.com/huggingface/xet-core.git && \
-    cd xet-core/hf_xet/ && \
-    uv pip install maturin patchelf && \
-    python -m maturin build --release --out dist && \
-    mkdir -p /tmp/hf-xet/dist && \
-    cp dist/*.whl /tmp/hf-xet/dist/
-
 # Final build stage
 FROM python-install AS vllm-cpu
 ARG PYTHON_VERSION
@@ -136,15 +120,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,from=rust,source=/root/.rustup,target=/root/.rustup,rw \
     --mount=type=bind,from=pyarrow,source=/tmp/arrow/python/dist,target=/tmp/arrow-wheels \
     --mount=type=bind,from=torch-vision,source=/tmp/vision/dist,target=/tmp/vision-wheels/ \
-    --mount=type=bind,from=hf-xet-builder,source=/tmp/hf-xet/dist,target=/tmp/hf-xet-wheels/ \
      sed -i '/^torch/d' requirements/build.txt && \
      ARROW_WHL_FILE=$(ls /tmp/arrow-wheels/pyarrow-*.whl | head -n 1) && \
      VISION_WHL_FILE=$(ls /tmp/vision-wheels/*.whl | head -n 1) && \
-     HF_XET_WHL_FILE=$(ls /tmp/hf-xet-wheels/*.whl | head -n 1) && \
     uv pip install -v \    
         $ARROW_WHL_FILE  \
         $VISION_WHL_FILE \
-        $HF_XET_WHL_FILE \
         --extra-index-url https://download.pytorch.org/whl/nightly/cpu \
         --index-strategy unsafe-best-match \
         -r requirements/build.txt \
@@ -168,5 +149,4 @@ USER 2000
 WORKDIR /home/vllm
 
 # Set the default entrypoint
-ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]
-
+ENTRYPOINT ["python", "-m", "vllm.entrypoints.openai.api_server"]

docs/source/conf.py

@@ -177,11 +177,6 @@ def linkcode_resolve(domain, info):
         for part in info['fullname'].split('.'):
             obj = getattr(obj, part)
 
-            # Skip decorator wrappers by checking if the object is a function
-            # and has a __wrapped__ attribute (which decorators typically set)
-            while hasattr(obj, '__wrapped__'):
-                obj = obj.__wrapped__
-
             if not (inspect.isclass(obj) or inspect.isfunction(obj)
                     or inspect.ismethod(obj)):
                 obj = obj.__class__  # Get the class of the instance

docs/source/contributing/model/multimodal.md

@@ -128,9 +128,11 @@ HF processing as well as memory profiling.
 
 ### For memory profiling
 
-Override the abstract methods {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_text` and {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_mm_data` to construct dummy inputs for memory profiling. These dummy inputs should result in the worst-case memory usage of the model so that vLLM can reserve the correct amount of memory for it.
+Override the abstract method {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`
+to construct dummy inputs for memory profiling. This dummy input should result in the worst-case memory usage of
+the model so that vLLM can reserve the correct amount of memory for it.
 
-Assuming that the memory usage increases with the number of tokens, the dummy inputs can be constructed to maximize the number of output embeddings, which is the same number as placeholder feature tokens.
+Assuming that the memory usage increases with the number of tokens, the dummy input can be constructed to maximize the number of output embeddings, which is the same number as placeholder feature tokens.
 
 ::::{tab-set}
 :::{tab-item} Basic example: LLaVA
@@ -242,45 +244,38 @@ def get_num_image_tokens(
 ```
 
 Notice that the number of image tokens doesn't depend on the image width and height.
-We can simply use a dummy `image_size` to calculate the multimodal profiling data:
+We can simply use a dummy `image_size`:
 
 ```python
-# NOTE: In actuality, this is usually implemented as part of the
-# model's subclass of `BaseProcessingInfo`, but we show it as is
-# here for simplicity.
 def get_image_size_with_most_features(self) -> ImageSize:
     hf_config = self.get_hf_config()
     width = height = hf_config.image_size
     return ImageSize(width=width, height=height)
 
-def get_dummy_mm_data(
+def get_dummy_processor_inputs(
     self,
     seq_len: int,
     mm_counts: Mapping[str, int],
-) -> MultiModalDataDict:
+) -> ProcessorInputs:
     num_images = mm_counts.get("image", 0)
 
-    target_width, target_height = \
-        self.info.get_image_size_with_most_features()
+    processor = self.info.get_hf_processor()
+    image_token = processor.image_token
+  
+    hf_config = self.get_hf_config()
+    target_width, target_height = self.info.get_image_size_with_most_features()
 
-    return {
+    mm_data = {
         "image":
         self._get_dummy_images(width=target_width,
                                height=target_height,
                                num_images=num_images)
     }
-```
 
-For the text, we simply expand the multimodal image token from the model config to match the desired number of images.
-
-```python
-def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
-    num_images = mm_counts.get("image", 0)
-
-    processor = self.info.get_hf_processor()
-    image_token = processor.image_token
-
-    return image_token * num_images
+    return ProcessorInputs(
+        prompt_text=image_token * num_images,
+        mm_data=mm_data,
+    )
 ```
 
 :::
@@ -417,30 +412,29 @@ def get_image_size_with_most_features(self) -> ImageSize:
 
 Fuyu does not expect image placeholders in the inputs to HF processor, so
 the dummy prompt text is empty regardless of the number of images.
+Otherwise, the logic of this method is very similar to LLaVA:
 
 ```python
-def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
-    return ""
-```
-
-For the multimodal image profiling data, the logic is very similar to LLaVA:
-
-```python
-def get_dummy_mm_data(
+def get_dummy_processor_inputs(
     self,
     seq_len: int,
     mm_counts: Mapping[str, int],
-) -> MultiModalDataDict:
+) -> ProcessorInputs:
     target_width, target_height = \
         self.info.get_image_size_with_most_features()
     num_images = mm_counts.get("image", 0)
 
-    return {
+    mm_data = {
         "image":
         self._get_dummy_images(width=target_width,
-                               height=target_height,
-                               num_images=num_images)
+                                height=target_height,
+                                num_images=num_images)
     }
+
+    return ProcessorInputs(
+        prompt_text="",
+        mm_data=mm_data,
+    )
 ```
 
 :::

docs/source/deployment/docker.md

@@ -19,18 +19,6 @@ $ docker run --runtime nvidia --gpus all \
     --model mistralai/Mistral-7B-v0.1
 ```
 
-This image can also be used with other container engines such as [Podman](https://podman.io/).
-
-```console
-$ podman run --gpus all \
-  -v ~/.cache/huggingface:/root/.cache/huggingface \
-  --env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
-  -p 8000:8000 \
-  --ipc=host \
-  vllm/vllm-openai:latest \
-  --model mistralai/Mistral-7B-v0.1
-```
-
 You can add any other <project:#engine-args> you need after the image tag (`vllm/vllm-openai:latest`).
 
 :::{note}

docs/source/deployment/frameworks/anything-llm.md

@@ -1,47 +0,0 @@
-(deployment-anything-llm)=
-
-# Anything LLM
-
-[Anything LLM](https://github.com/Mintplex-Labs/anything-llm) is a full-stack application that enables you to turn any document, resource, or piece of content into context that any LLM can use as references during chatting.
-
-It allows you to deploy a large language model (LLM) server with vLLM as the backend, which exposes OpenAI-compatible endpoints.
-
-## Prerequisites
-
-- Setup vLLM environment
-
-## Deploy
-
-- Start the vLLM server with the supported chat completion model, e.g.
-
-```console
-vllm serve Qwen/Qwen1.5-32B-Chat-AWQ --max-model-len 4096
-```
-
-- Download and install [Anything LLM desktop](https://anythingllm.com/desktop).
-
-- On the bottom left of open settings, AI Prooviders --> LLM:
-  - LLM Provider: Generic OpenAI
-  - Base URL: http://{vllm server host}:{vllm server port}/v1
-  - Chat Model Name: `Qwen/Qwen1.5-32B-Chat-AWQ`
-
-:::{image} /assets/deployment/anything-llm-provider.png
-:::
-
-- Back to home page, New Workspace --> create `vllm` workspace, and start to chat:
-
-:::{image} /assets/deployment/anything-llm-chat-without-doc.png
-:::
-
-- Click the upload button:
-  - upload the doc
-  - select the doc and move to the workspace
-  - save and embed
-
-:::{image} /assets/deployment/anything-llm-upload-doc.png
-:::
-
-- Chat again:
-
-:::{image} /assets/deployment/anything-llm-chat-with-doc.png
-:::

docs/source/deployment/frameworks/index.md

@@ -3,14 +3,12 @@
 :::{toctree}
 :maxdepth: 1
 
-anything-llm
 bentoml
 cerebrium
 dstack
 helm
 lws
 modal
-open-webui
 skypilot
 triton
 :::

docs/source/deployment/frameworks/open-webui.md

@@ -1,29 +0,0 @@
-(deployment-open-webui)=
-
-# Open WebUI
-
-1. Install the [Docker](https://docs.docker.com/engine/install/)
-
-2. Start the vLLM server with the supported chat completion model, e.g.
-
-```console
-vllm serve qwen/Qwen1.5-0.5B-Chat
-```
-
-1. Start the [Open WebUI](https://github.com/open-webui/open-webui) docker container (replace the vllm serve host and vllm serve port):
-
-```console
-docker run -d -p 3000:8080 \
---name open-webui \
--v open-webui:/app/backend/data \
--e OPENAI_API_BASE_URL=http://<vllm serve host>:<vllm serve port>/v1 \
---restart always \
-ghcr.io/open-webui/open-webui:main
-```
-
-1. Open it in the browser: <http://open-webui-host:3000/>
-
-On the top of the web page, you can see the model `qwen/Qwen1.5-0.5B-Chat`.
-
-:::{image} /assets/deployment/open_webui.png
-:::

docs/source/deployment/integrations/production-stack.md

@@ -16,7 +16,7 @@ Ensure that you have a running Kubernetes environment with GPU (you can follow [
 
 ## Deployment using vLLM production stack
 
-The standard vLLM production stack is installed using a Helm chart. You can run this [bash script](https://github.com/vllm-project/production-stack/blob/main/utils/install-helm.sh) to install Helm on your GPU server.
+The standard vLLM production stack install uses a Helm chart. You can run this [bash script](https://github.com/vllm-project/production-stack/blob/main/tutorials/install-helm.sh) to install Helm on your GPU server.
 
 To install the vLLM production stack, run the following commands on your desktop:
 

docs/source/deployment/security.md

@@ -1,58 +0,0 @@
-# Security Guide
-
-## Inter-Node Communication
-
-All communications between nodes in a multi-node vLLM deployment are **insecure by default** and must be protected by placing the nodes on an isolated network. This includes:
-
-1. PyTorch Distributed communications
-2. KV cache transfer communications
-3. Tensor, Pipeline, and Data parallel communications
-
-### Configuration Options for Inter-Node Communications
-
-The following options control inter-node communications in vLLM:
-
-1. **Environment Variables:**
-   - `VLLM_HOST_IP`: Sets the IP address for vLLM processes to communicate on
-
-2. **KV Cache Transfer Configuration:**
-   - `--kv-ip`: The IP address for KV cache transfer communications (default: 127.0.0.1)
-   - `--kv-port`: The port for KV cache transfer communications (default: 14579)
-
-3. **Data Parallel Configuration:**
-   - `data_parallel_master_ip`: IP of the data parallel master (default: 127.0.0.1)
-   - `data_parallel_master_port`: Port of the data parallel master (default: 29500)
-
-### Notes on PyTorch Distributed
-
-vLLM uses PyTorch's distributed features for some inter-node communication. For
-detailed information about PyTorch Distributed security considerations, please
-refer to the [PyTorch Security
-Guide](https://github.com/pytorch/pytorch/security/policy#using-distributed-features).
-
-Key points from the PyTorch security guide:
-- PyTorch Distributed features are intended for internal communication only
-- They are not built for use in untrusted environments or networks
-- No authorization protocol is included for performance reasons
-- Messages are sent unencrypted
-- Connections are accepted from anywhere without checks
-
-### Security Recommendations
-
-1. **Network Isolation:**
-   - Deploy vLLM nodes on a dedicated, isolated network
-   - Use network segmentation to prevent unauthorized access
-   - Implement appropriate firewall rules
-
-2. **Configuration Best Practices:**
-   - Always set `VLLM_HOST_IP` to a specific IP address rather than using defaults
-   - Configure firewalls to only allow necessary ports between nodes
-
-3. **Access Control:**
-   - Restrict physical and network access to the deployment environment
-   - Implement proper authentication and authorization for management interfaces
-   - Follow the principle of least privilege for all system components
-
-## Reporting Security Vulnerabilities
-
-If you believe you have found a security vulnerability in vLLM, please report it following the project's security policy. For more information on how to report security issues and the project's security policy, please see the [vLLM Security Policy](https://github.com/vllm-project/vllm/blob/main/SECURITY.md).

docs/source/design/mm_processing.md

@@ -47,7 +47,7 @@ Moreover, since the tokenized text has not passed through the HF processor, we h
 
 ### Dummy text
 
-We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_text`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
+We work around the first issue by requiring each model to define how to generate dummy text based on the number of multi-modal inputs, via {meth}`~vllm.multimodal.profiling.BaseDummyInputsBuilder.get_dummy_processor_inputs`. This lets us generate dummy text corresponding to the multi-modal inputs and input them together to obtain the processed multi-modal data.
 
 (mm-automatic-prompt-updating)=
 

docs/source/design/v1/metrics.md

@@ -66,8 +66,8 @@ vLLM also provides [a reference example](https://docs.vllm.ai/en/latest/getting_
 The subset of metrics exposed in the Grafana dashboard gives us an indication of which metrics are especially important:
 
 - `vllm:e2e_request_latency_seconds_bucket` - End to end request latency measured in seconds
-- `vllm:prompt_tokens_total` - Prompt Tokens
-- `vllm:generation_tokens_total` - Generation Tokens
+- `vllm:prompt_tokens_total` - Prompt Tokens/Sec
+- `vllm:generation_tokens_total` - Generation Tokens/Sec
 - `vllm:time_per_output_token_seconds` - Inter token latency (Time Per Output Token, TPOT) in second.
 - `vllm:time_to_first_token_seconds` - Time to First Token (TTFT) latency in seconds.
 - `vllm:num_requests_running` (also, `_swapped` and `_waiting`) - Number of requests in RUNNING, WAITING, and SWAPPED state
@@ -86,17 +86,6 @@ See [the PR which added this Dashboard](gh-pr:2316) for interesting and useful b
 
 Prometheus support was initially added [using the aioprometheus library](gh-pr:1890), but a switch was made quickly to [prometheus_client](gh-pr:2730). The rationale is discussed in both linked PRs.
 
-With the switch to `aioprometheus`, we lost a `MetricsMiddleware` to track HTTP metrics, but this was reinstated [using prometheus_fastapi_instrumentator](gh-pr:15657):
-
-```bash
-$ curl http://0.0.0.0:8000/metrics 2>/dev/null  | grep -P '^http_(?!.*(_bucket|_created|_sum)).*'
-http_requests_total{handler="/v1/completions",method="POST",status="2xx"} 201.0
-http_request_size_bytes_count{handler="/v1/completions"} 201.0
-http_response_size_bytes_count{handler="/v1/completions"} 201.0
-http_request_duration_highr_seconds_count 201.0
-http_request_duration_seconds_count{handler="/v1/completions",method="POST"} 201.0
-```
-
 ### Multi-process Mode
 
 In v0, metrics are collected in the engine core process and we use multi-process mode to make them available in the API server process. See <gh-pr:7279>.

docs/source/design/v1/torch_compile.md

@@ -99,7 +99,7 @@ This time, Inductor compilation is completely bypassed, and we will load from di
 
 The above example just uses Inductor to compile for a general shape (i.e. symbolic shape). We can also use Inductor to compile for some of the specific shapes, for example:
 
-`vllm serve meta-llama/Llama-3.2-1B --compilation_config "{'compile_sizes': [1, 2, 4, 8]}"`
+`VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.2-1B --compilation_config "{'compile_sizes': [1, 2, 4, 8]}"`
 
 Then it will also compile a specific kernel just for batch size `1, 2, 4, 8`. At this time, all of the shapes in the computation graph are static and known, and we will turn on auto-tuning to tune for max performance. This can be slow when you run it for the first time, but the next time you run it, we can directly bypass the tuning and run the tuned kernel.
 
@@ -134,6 +134,6 @@ The cudagraphs are captured and managed by the compiler backend, and replayed wh
 
 By default, vLLM will try to determine a set of sizes to capture cudagraph. You can also override it using the config `cudagraph_capture_sizes`:
 
-`vllm serve meta-llama/Llama-3.2-1B --compilation-config "{'cudagraph_capture_sizes': [1, 2, 4, 8]}"`
+`VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.2-1B --compilation_config "{'cudagraph_capture_sizes': [1, 2, 4, 8]}"`
 
 Then it will only capture cudagraph for the specified sizes. It can be useful to have fine-grained control over the cudagraph capture.

docs/source/features/disagg_prefill.md

@@ -21,11 +21,11 @@ Disaggregated prefill DOES NOT improve throughput.
 
 ## Usage example
 
-Please refer to <gh-file:examples/online_serving/disaggregated_prefill.sh> for the example usage of disaggregated prefilling.
+Please refer to `examples/online_serving/disaggregated_prefill.sh` for the example usage of disaggregated prefilling.
 
 ## Benchmarks
 
-Please refer to <gh-file:benchmarks/disagg_benchmarks> for disaggregated prefilling benchmarks.
+Please refer to `benchmarks/disagg_benchmarks/` for disaggregated prefilling benchmarks.
 
 ## Development
 

docs/source/features/lora.md

@@ -106,18 +106,19 @@ curl http://localhost:8000/v1/completions \
 
 ## Dynamically serving LoRA Adapters
 
-In addition to serving LoRA adapters at server startup, the vLLM server supports dynamically configuring LoRA adapters at runtime through dedicated API endpoints and plugins. This feature can be particularly useful when the flexibility to change models on-the-fly is needed.
+In addition to serving LoRA adapters at server startup, the vLLM server now supports dynamically loading and unloading
+LoRA adapters at runtime through dedicated API endpoints. This feature can be particularly useful when the flexibility
+to change models on-the-fly is needed.
 
 Note: Enabling this feature in production environments is risky as users may participate in model adapter management.
 
-To enable dynamic LoRA configuration, ensure that the environment variable `VLLM_ALLOW_RUNTIME_LORA_UPDATING`
-is set to `True`.
+To enable dynamic LoRA loading and unloading, ensure that the environment variable `VLLM_ALLOW_RUNTIME_LORA_UPDATING`
+is set to `True`. When this option is enabled, the API server will log a warning to indicate that dynamic loading is active.
 
 ```bash
 export VLLM_ALLOW_RUNTIME_LORA_UPDATING=True
 ```
 
-### Using API Endpoints
 Loading a LoRA Adapter:
 
 To dynamically load a LoRA adapter, send a POST request to the `/v1/load_lora_adapter` endpoint with the necessary
@@ -152,58 +153,6 @@ curl -X POST http://localhost:8000/v1/unload_lora_adapter \
 }'
 ```
 
-### Using Plugins
-Alternatively, you can use the LoRAResolver plugin to dynamically load LoRA adapters. LoRAResolver plugins enable you to load LoRA adapters from both local and remote sources such as local file system and S3. On every request, when there's a new model name that hasn't been loaded yet, the LoRAResolver will try to resolve and load the corresponding LoRA adapter.
-
-You can set up multiple LoRAResolver plugins if you want to load LoRA adapters from different sources. For example, you might have one resolver for local files and another for S3 storage. vLLM will load the first LoRA adapter that it finds.
-
-You can either install existing plugins or implement your own.
-
-Steps to implement your own LoRAResolver plugin:
-1. Implement the LoRAResolver interface.
-
-    Example of a simple S3 LoRAResolver implementation:
-
-    ```python
-    import os
-    import s3fs
-    from vllm.lora.request import LoRARequest
-    from vllm.lora.resolver import LoRAResolver
-
-    class S3LoRAResolver(LoRAResolver):
-        def __init__(self):
-            self.s3 = s3fs.S3FileSystem()
-            self.s3_path_format = os.getenv("S3_PATH_TEMPLATE")
-            self.local_path_format = os.getenv("LOCAL_PATH_TEMPLATE")
-
-        async def resolve_lora(self, base_model_name, lora_name):
-            s3_path = self.s3_path_format.format(base_model_name=base_model_name, lora_name=lora_name)
-            local_path = self.local_path_format.format(base_model_name=base_model_name, lora_name=lora_name)
-
-            # Download the LoRA from S3 to the local path
-            await self.s3._get(
-                s3_path, local_path, recursive=True, maxdepth=1
-            )
-
-            lora_request = LoRARequest(
-                lora_name=lora_name,
-                lora_path=local_path,
-                lora_int_id=abs(hash(lora_name))
-            )
-            return lora_request
-    ```
-
-2. Register LoRAResolver plugin.
-
-     ```python
-    from vllm.lora.resolver import LoRAResolverRegistry
-
-    s3_resolver = S3LoRAResolver()
-    LoRAResolverRegistry.register_resolver("s3_resolver", s3_resolver)
-    ```
-
-    For more details, refer to the [vLLM's Plugins System](../design/plugin_system.md).
-
 ## New format for `--lora-modules`
 
 In the previous version, users would provide LoRA modules via the following format, either as a key-value pair or in JSON format. For example:

docs/source/features/quantization/auto_awq.md

@@ -6,13 +6,13 @@ To create a new 4-bit quantized model, you can leverage [AutoAWQ](https://github
 Quantization reduces the model's precision from BF16/FP16 to INT4 which effectively reduces the total model memory footprint.
 The main benefits are lower latency and memory usage.
 
-You can quantize your own models by installing AutoAWQ or picking one of the [6500+ models on Huggingface](https://huggingface.co/models?search=awq).
+You can quantize your own models by installing AutoAWQ or picking one of the [6500+ models on Huggingface](https://huggingface.co/models?sort=trending&search=awq).
 
 ```console
 pip install autoawq
 ```
 
-After installing AutoAWQ, you are ready to quantize a model. Please refer to the [AutoAWQ documentation](https://casper-hansen.github.io/AutoAWQ/examples/#basic-quantization) for further details. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
+After installing AutoAWQ, you are ready to quantize a model. Please refer to the `AutoAWQ documentation <https://casper-hansen.github.io/AutoAWQ/examples/#basic-quantization>`_ for further details. Here is an example of how to quantize `mistralai/Mistral-7B-Instruct-v0.2`:
 
 ```python
 from awq import AutoAWQForCausalLM

docs/source/features/quantization/bitblas.md

@@ -1,48 +0,0 @@
-(bitblas)=
-
-# BitBLAS
-
-vLLM now supports [BitBLAS](https://github.com/microsoft/BitBLAS) for more efficient and flexible model inference. Compared to other quantization frameworks, BitBLAS provides more precision combinations.
-
-:::{note}
-Ensure your hardware supports the selected `dtype` (`torch.bfloat16` or `torch.float16`).
-Most recent NVIDIA GPUs support `float16`, while `bfloat16` is more common on newer architectures like Ampere or Hopper.
-For details see [supported hardware](https://docs.vllm.ai/en/latest/features/quantization/supported_hardware.html).
-:::
-
-Below are the steps to utilize BitBLAS with vLLM.
-
-```console
-pip install bitblas>=0.1.0
-```
-
-vLLM reads the model's config file and supports pre-quantized checkpoints.
-
-You can find pre-quantized models on:
-
-- [Hugging Face (BitBLAS)](https://huggingface.co/models?search=bitblas)
-- [Hugging Face (GPTQ)](https://huggingface.co/models?search=gptq)
-
-Usually, these repositories have a `quantize_config.json` file that includes a `quantization_config` section.
-
-## Read bitblas format checkpoint
-
-```python
-from vllm import LLM
-import torch
-
-# "hxbgsyxh/llama-13b-4bit-g-1-bitblas" is a pre-quantized checkpoint.
-model_id = "hxbgsyxh/llama-13b-4bit-g-1-bitblas"
-llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True, quantization="bitblas")
-```
-
-## Read gptq format checkpoint
-
-```python
-from vllm import LLM
-import torch
-
-# "hxbgsyxh/llama-13b-4bit-g-1" is a pre-quantized checkpoint.
-model_id = "hxbgsyxh/llama-13b-4bit-g-1"
-llm = LLM(model=model_id, dtype=torch.float16, trust_remote_code=True, quantization="bitblas", max_model_len=1024)
-```

docs/source/features/quantization/bnb.md

@@ -14,7 +14,7 @@ pip install bitsandbytes>=0.45.3
 
 vLLM reads the model's config file and supports both in-flight quantization and pre-quantized checkpoint.
 
-You can find bitsandbytes quantized models on <https://huggingface.co/models?search=bitsandbytes>.
+You can find bitsandbytes quantized models on <https://huggingface.co/models?other=bitsandbytes>.
 And usually, these repositories have a config.json file that includes a quantization_config section.
 
 ## Read quantized checkpoint