[V1] Fix profiling.py

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

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

@@ -0,0 +1,11 @@
+# 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: 0.31
+  - name: "exact_match,flexible-extract"
+    value: 0.47
+limit: 1319
+num_fewshot: 5

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

@@ -4,7 +4,7 @@ Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
-Minitron-4B-Base-FP8.yaml
+Qwen1.5-MoE-W4A16-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-FP8W8.yaml
 Meta-Llama-3-8B-QQQ.yaml

.buildkite/scripts/run-benchmarks.sh

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

.buildkite/test-pipeline.yaml

@@ -163,11 +163,6 @@ steps:
   - tests/tracing
   commands:
   - pytest -v -s metrics
-  - "pip install \
-      'opentelemetry-sdk>=1.26.0,<1.27.0' \
-      'opentelemetry-api>=1.26.0,<1.27.0' \
-      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
@@ -292,6 +287,14 @@ steps:
   command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
   parallelism: 4
 
+- label: PyTorch Compilation Unit Tests
+  source_file_dependencies:
+    - vllm/
+    - tests/compile
+  commands:
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+
 - label: PyTorch Fullgraph Smoke Test # 9min
   source_file_dependencies:
   - vllm/
@@ -301,7 +304,6 @@ steps:
   # these tests need to be separated, cannot combine
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
-  - pytest -v -s compile/test_pass_manager.py
 
 - label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
@@ -389,7 +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
+    - 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'
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
@@ -426,7 +429,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal
     - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
+    - pytest -v -s models/decoder_only/vision_language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/vision_language -m core_model
     - pytest -v -s models/encoder_decoder/audio_language -m core_model
     - pytest -v -s models/encoder_decoder/language -m core_model
@@ -445,10 +448,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
     - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
-    # HACK - run phi3v tests separately to sidestep this transformers bug
-    # https://github.com/huggingface/transformers/issues/34307
-    - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
+    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/vision_language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'

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

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

.github/PULL_REQUEST_TEMPLATE.md

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

.pre-commit-config.yaml

@@ -122,6 +122,12 @@ repos:
     language: system
     always_run: true
     pass_filenames: false
+  - id: update-dockerfile-graph
+    name: Update Dockerfile dependency graph
+    entry: tools/update-dockerfile-graph.sh
+    language: script
+    files: ^docker/Dockerfile$
+    pass_filenames: false
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -230,6 +230,7 @@ set(VLLM_EXT_SRC
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
+  "csrc/attention/merge_attn_states.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"

README.md

@@ -15,11 +15,8 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
-[2025/04] We're hosting our first-ever *vLLM Asia Developer Day* in Singapore on *April 3rd*! This is a full-day event (9 AM - 9 PM SGT) in partnership with SGInnovate, AMD, and Embedded LLM. Meet the vLLM team and learn about LLM inference for RL, MI300X, and more! [Register Now](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)
-
----
-
 *Latest News* 🔥
+- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).

benchmarks/README.md

@@ -204,6 +204,24 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --seed 42
 ```
 
+### Running With Sampling Parameters
+
+When using OpenAI-compatible backends such as `vllm`, optional sampling
+parameters can be specified. Example client command:
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
+```
+
 ---
 ## Example - Offline Throughput Benchmark
 

benchmarks/backend_request_func.py

@@ -497,3 +497,9 @@ ASYNC_REQUEST_FUNCS = {
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
 }
+
+OPENAI_COMPATIBLE_BACKENDS = [
+    k for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions,
+             async_request_openai_chat_completions)
+]

benchmarks/benchmark_dataset.py

@@ -288,7 +288,7 @@ def process_image(image: Any) -> Mapping[str, Any]:
 class RandomDataset(BenchmarkDataset):
     # Default values copied from benchmark_serving.py for the random dataset.
     DEFAULT_PREFIX_LEN = 0
-    DEFAULT_RANGE_RATIO = 1.0
+    DEFAULT_RANGE_RATIO = 0.0
     DEFAULT_INPUT_LEN = 1024
     DEFAULT_OUTPUT_LEN = 128
 
@@ -308,19 +308,32 @@ class RandomDataset(BenchmarkDataset):
         output_len: int = DEFAULT_OUTPUT_LEN,
         **kwargs,
     ) -> list[SampleRequest]:
+        # Enforce range_ratio < 1
+        assert range_ratio < 1.0, (
+            "random_range_ratio must be < 1.0 to ensure a valid sampling range"
+        )
+
         vocab_size = tokenizer.vocab_size
 
         prefix_token_ids = (np.random.randint(
             0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
 
-        input_low = int(input_len * range_ratio)
-        output_low = int(output_len * range_ratio)
+        # New sampling logic: [X * (1 - b), X * (1 + b)]
+        input_low = int(input_len * (1 - range_ratio))
+        input_high = int(input_len * (1 + range_ratio))
+        output_low = int(output_len * (1 - range_ratio))
+        output_high = int(output_len * (1 + range_ratio))
+
+        # Add logging for debugging
+        logger.info("Sampling input_len from [%s, %s]", input_low, input_high)
+        logger.info("Sampling output_len from [%s, %s]", output_low,
+                    output_high)
 
         input_lens = np.random.randint(input_low,
-                                       input_len + 1,
+                                       input_high + 1,
                                        size=num_requests)
         output_lens = np.random.randint(output_low,
-                                        output_len + 1,
+                                        output_high + 1,
                                         size=num_requests)
         offsets = np.random.randint(0, vocab_size, size=num_requests)
 
@@ -472,11 +485,11 @@ class SonnetDataset(BenchmarkDataset):
 
         # Determine how many poem lines to use.
         num_input_lines = round((input_len - base_offset) / avg_len)
-        num_prefix_lines = round((prefix_len - base_offset) / avg_len)
+        num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0)
         prefix_lines = self.data[:num_prefix_lines]
 
         samples = []
-        for _ in range(num_requests):
+        while len(samples) < num_requests:
             extra_lines = random.choices(self.data,
                                          k=num_input_lines - num_prefix_lines)
             prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
@@ -484,13 +497,14 @@ class SonnetDataset(BenchmarkDataset):
             prompt_formatted = tokenizer.apply_chat_template(
                 msg, add_generation_prompt=True, tokenize=False)
             prompt_len = len(tokenizer(prompt_formatted).input_ids)
-            samples.append(
-                SampleRequest(
-                    prompt=prompt_formatted
-                    if return_prompt_formatted else prompt,
-                    prompt_len=prompt_len,
-                    expected_output_len=output_len,
-                ))
+            if prompt_len <= input_len:
+                samples.append(
+                    SampleRequest(
+                        prompt=prompt_formatted
+                        if return_prompt_formatted else prompt,
+                        prompt_len=prompt_len,
+                        expected_output_len=output_len,
+                    ))
         return samples
 
 

benchmarks/benchmark_serving.py

@@ -34,7 +34,8 @@ from datetime import datetime
 from typing import Any, Optional
 
 import numpy as np
-from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+from backend_request_func import (ASYNC_REQUEST_FUNCS,
+                                  OPENAI_COMPATIBLE_BACKENDS, RequestFuncInput,
                                   RequestFuncOutput)
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
@@ -155,7 +156,7 @@ def calculate_metrics(
         if outputs[i].success:
             output_len = outputs[i].output_tokens
 
-            if output_len is None:
+            if not output_len:
                 # We use the tokenizer to count the number of output tokens
                 # for some serving backends instead of looking at
                 # len(outputs[i].itl) since multiple output tokens may be
@@ -260,6 +261,7 @@ async def benchmark(
     goodput_config_dict: dict[str, float],
     max_concurrency: Optional[int],
     lora_modules: Optional[Iterable[str]],
+    extra_body: Optional[dict],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -287,6 +289,7 @@ async def benchmark(
         logprobs=logprobs,
         multi_modal_content=test_mm_content,
         ignore_eos=ignore_eos,
+        extra_body=extra_body,
     )
 
     test_output = await request_func(request_func_input=test_input)
@@ -313,7 +316,8 @@ async def benchmark(
                                          output_len=test_output_len,
                                          logprobs=logprobs,
                                          multi_modal_content=test_mm_content,
-                                         ignore_eos=ignore_eos)
+                                         ignore_eos=ignore_eos,
+                                         extra_body=extra_body)
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
             print("Profiler started")
@@ -363,7 +367,8 @@ async def benchmark(
                                               output_len=output_len,
                                               logprobs=logprobs,
                                               multi_modal_content=mm_content,
-                                              ignore_eos=ignore_eos)
+                                              ignore_eos=ignore_eos,
+                                              extra_body=extra_body)
         tasks.append(
             asyncio.create_task(
                 limited_request_func(request_func_input=request_func_input,
@@ -652,6 +657,26 @@ def main(args: argparse.Namespace):
             raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
     goodput_config_dict = check_goodput_args(args)
 
+    # Collect the sampling parameters.
+    sampling_params = {
+        k: v
+        for k, v in {
+            "top_p": args.top_p,
+            "top_k": args.top_k,
+            "min_p": args.min_p,
+            "temperature": args.temperature
+        }.items() if v is not None
+    }
+
+    # Sampling parameters are only supported by openai-compatible backend.
+    if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS:
+        raise ValueError(
+            "Sampling parameters are only supported by openai-compatible "
+            "backends.")
+
+    if "temperature" not in sampling_params:
+        sampling_params["temperature"] = 0.0  # Default to greedy decoding.
+
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -678,6 +703,7 @@ def main(args: argparse.Namespace):
             goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
+            extra_body=sampling_params,
         ))
 
     # Save config and results to json
@@ -895,7 +921,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -903,7 +929,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -970,18 +996,23 @@ if __name__ == "__main__":
     random_group.add_argument(
         "--random-range-ratio",
         type=float,
-        default=1.0,
-        help="Range of sampled ratio of input/output length, "
-        "used only for random sampling.",
+        default=0.0,
+        help="Range ratio for sampling input/output length, "
+        "used only for random sampling. Must be in the range [0, 1) to define "
+        "a symmetric sampling range"
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
     random_group.add_argument(
         "--random-prefix-len",
         type=int,
         default=0,
-        help="Number of fixed prefix tokens before random "
-        " context. The length range of context in a random "
-        " request is [random-prefix-len, "
-        " random-prefix-len + random-prefix-len * random-range-ratio).")
+        help=("Number of fixed prefix tokens before the random context "
+              "in a request. "
+              "The total input length is the sum of `random-prefix-len` and "
+              "a random "
+              "context length sampled from [input_len * (1 - range_ratio), "
+              "input_len * (1 + range_ratio)]."),
+    )
 
     hf_group = parser.add_argument_group("hf dataset options")
     hf_group.add_argument("--hf-subset",
@@ -1000,6 +1031,33 @@ if __name__ == "__main__":
         "from the sampled HF dataset.",
     )
 
+    sampling_group = parser.add_argument_group("sampling parameters")
+    sampling_group.add_argument(
+        "--top-p",
+        type=float,
+        default=None,
+        help="Top-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--top-k",
+        type=int,
+        default=None,
+        help="Top-k sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--min-p",
+        type=float,
+        default=None,
+        help="Min-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--temperature",
+        type=float,
+        default=None,
+        help="Temperature sampling parameter. Only has effect on "
+        "openai-compatible backends. If not specified, default to greedy "
+        "decoding (i.e. temperature==0.0).")
+
     parser.add_argument(
         '--tokenizer-mode',
         type=str,

benchmarks/benchmark_serving_structured_output.py

@@ -11,7 +11,7 @@ On the client side, run:
         --model <your_model> \
         --dataset json \
         --structured-output-ratio 1.0 \
-        --structured-output-backend xgrammar \
+        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -130,10 +130,11 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
                         "description":
                         "An unique optional field to avoid cached schemas"
                     }
+        else:
+            json_schemas = [schema] * args.num_prompts
 
         def gen_prompt(index: int):
-            schema = json_schemas[index % len(json_schemas)]
-            return f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
+            return f"Generate an example of a user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
 
         def get_schema(index: int):
             return json_schemas[index % len(json_schemas)]
@@ -963,7 +964,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -971,7 +972,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -996,12 +997,14 @@ if __name__ == "__main__":
                         type=float,
                         default=1.0,
                         help="Ratio of Structured Outputs requests")
-    parser.add_argument(
-        "--structured-output-backend",
-        type=str,
-        choices=["outlines", "lm-format-enforcer", "xgrammar", "guidance"],
-        default="xgrammar",
-        help="Backend to use for structured outputs")
+    parser.add_argument("--structured-output-backend",
+                        type=str,
+                        choices=[
+                            "outlines", "lm-format-enforcer", "xgrammar",
+                            "guidance", "auto"
+                        ],
+                        default="auto",
+                        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -213,14 +213,17 @@ def run_hf(
     max_prompt_len = 0
     max_output_len = 0
     for i in range(len(requests)):
-        prompt, prompt_len, output_len = requests[i]
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
         # Add the prompt to the batch.
         batch.append(prompt)
         max_prompt_len = max(max_prompt_len, prompt_len)
         max_output_len = max(max_output_len, output_len)
         if len(batch) < max_batch_size and i != len(requests) - 1:
             # Check if we can add more requests to the batch.
-            _, next_prompt_len, next_output_len = requests[i + 1]
+            next_prompt_len = requests[i + 1].prompt_len
+            next_output_len = requests[i + 1].expected_output_len
             if (max(max_prompt_len, next_prompt_len) +
                     max(max_output_len, next_output_len)) <= 2048:
                 # We can add more requests to the batch.
@@ -591,18 +594,22 @@ if __name__ == "__main__":
         default=None,
         help="Path to the lora adapters to use. This can be an absolute path, "
         "a relative path, or a Hugging Face model identifier.")
-    parser.add_argument("--prefix-len",
-                        type=int,
-                        default=None,
-                        help="Number of prefix tokens per request."
-                        "This is for the RandomDataset and SonnetDataset")
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        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="Range of sampled ratio of input/output length, "
-        "used only for RandomDataSet.",
+        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)].",
     )
 
     # hf dtaset

benchmarks/kernels/benchmark_moe.py

@@ -553,6 +553,9 @@ def main(args: argparse.Namespace):
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
+        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/cpu_extension.cmake

@@ -33,8 +33,6 @@ endif()
 
 if(MACOSX_FOUND)
     list(APPEND CXX_COMPILE_FLAGS
-        "-Xpreprocessor"
-        "-fopenmp"
         "-DVLLM_CPU_EXTENSION")
 else()
     list(APPEND CXX_COMPILE_FLAGS

collect_env.py

@@ -105,8 +105,14 @@ def run(command):
     else:
         enc = locale.getpreferredencoding()
     output = raw_output.decode(enc)
+    if command == 'nvidia-smi topo -m':
+        # don't remove the leading whitespace of `nvidia-smi topo -m`
+        #   because they are meaningful
+        output = output.rstrip()
+    else:
+        output = output.strip()
     err = raw_err.decode(enc)
-    return rc, output.strip(), err.strip()
+    return rc, output, err.strip()
 
 
 def run_and_read_all(run_lambda, command):

csrc/attention/merge_attn_states.cu

@@ -0,0 +1,173 @@
+#include <optional>
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <algorithm>
+
+#include "attention_dtypes.h"
+#include "attention_utils.cuh"
+
+namespace vllm {
+
+// Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+// can be used to combine partial attention results (in the split-KV case)
+template <typename scalar_t, const uint NUM_THREADS>
+__global__ void merge_attn_states_kernel(
+    scalar_t* output, float* output_lse, const scalar_t* prefix_output,
+    const float* prefix_lse, const scalar_t* suffix_output,
+    const float* suffix_lse, const uint num_tokens, const uint num_heads,
+    const uint head_size) {
+  using pack_128b_t = uint4;
+  const uint pack_size = 16 / sizeof(scalar_t);
+  const uint threads_per_head = head_size / pack_size;
+
+  const uint global_idx = blockIdx.x * NUM_THREADS + threadIdx.x;
+  const uint token_head_threads = num_tokens * num_heads * threads_per_head;
+
+  if (global_idx >= token_head_threads) return;
+
+  // global_idx -> token_idx + head_idx + pack_idx
+  const uint token_head_idx = global_idx / threads_per_head;
+  const uint pack_idx = global_idx % threads_per_head;
+
+  const uint token_idx = token_head_idx / num_heads;
+  const uint head_idx = token_head_idx % num_heads;
+
+  const uint pack_offset = pack_idx * pack_size;  // (0~15)*8, etc.
+  const uint head_offset =
+      token_idx * num_heads * head_size + head_idx * head_size;
+  const scalar_t* prefix_head_ptr = prefix_output + head_offset;
+  const scalar_t* suffix_head_ptr = suffix_output + head_offset;
+  scalar_t* output_head_ptr = output + head_offset;
+
+  float p_lse = prefix_lse[head_idx * num_tokens + token_idx];
+  float s_lse = suffix_lse[head_idx * num_tokens + token_idx];
+  p_lse = std::isinf(p_lse) ? -std::numeric_limits<float>::infinity() : p_lse;
+  s_lse = std::isinf(s_lse) ? -std::numeric_limits<float>::infinity() : s_lse;
+
+  const float max_lse = fmaxf(p_lse, s_lse);
+  p_lse = p_lse - max_lse;
+  s_lse = s_lse - max_lse;
+  const float p_se = expf(p_lse);
+  const float s_se = expf(s_lse);
+  const float out_se = p_se + s_se;
+  const float p_scale = p_se / out_se;
+  const float s_scale = s_se / out_se;
+
+  if (pack_offset < head_size) {
+    // Pack 128b load
+    pack_128b_t p_out_pack = reinterpret_cast<const pack_128b_t*>(
+        prefix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t s_out_pack = reinterpret_cast<const pack_128b_t*>(
+        suffix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t o_out_pack;
+
+#pragma unroll
+    for (uint i = 0; i < pack_size; ++i) {
+      // Always use float for FMA to keep high precision.
+      // half(uint16_t), bfloat16, float -> float.
+      const float p_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&p_out_pack)[i]);
+      const float s_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&s_out_pack)[i]);
+      // fma: a * b + c = p_out_f * p_scale + (s_out_f * s_scale)
+      const float o_out_f = p_out_f * p_scale + (s_out_f * s_scale);
+      // float -> half(uint16_t), bfloat16, float.
+      vllm::from_float(reinterpret_cast<scalar_t*>(&o_out_pack)[i], o_out_f);
+    }
+
+    // Pack 128b storage
+    reinterpret_cast<pack_128b_t*>(output_head_ptr)[pack_offset / pack_size] =
+        o_out_pack;
+  }
+  // We only need to write to output_lse once per head.
+  if (output_lse != nullptr && pack_idx == 0) {
+    float out_lse = logf(out_se) + max_lse;
+    output_lse[head_idx * num_tokens + token_idx] = out_lse;
+  }
+}
+
+}  // namespace vllm
+
+// The following macro is used to dispatch the conversion function based on
+// the output data type. The FN is a macro that calls a function with
+// template<typename scalar_t>.
+#define DISPATCH_BY_SCALAR_DTYPE(scalar_dtype, fn)                      \
+  {                                                                     \
+    if (scalar_dtype == at::ScalarType::Float) {                        \
+      fn(float);                                                        \
+    } else if (scalar_dtype == at::ScalarType::Half) {                  \
+      fn(uint16_t);                                                     \
+    } else if (scalar_dtype == at::ScalarType::BFloat16) {              \
+      fn(__nv_bfloat16);                                                \
+    } else {                                                            \
+      TORCH_CHECK(false, "Unsupported data type of O: ", scalar_dtype); \
+    }                                                                   \
+  }
+
+#define LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS)                     \
+  {                                                                         \
+    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
+ * into the output tensor. NUM_TOKENS: n, NUM_HEADS: h, HEAD_SIZE: d
+ *
+ * @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,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,d] The log-sum-exp values for the suffix attention
+ * states.
+ */
+template <typename scalar_t>
+void merge_attn_states_launcher(torch::Tensor& output,
+                                std::optional<torch::Tensor> output_lse,
+                                const torch::Tensor& prefix_output,
+                                const torch::Tensor& prefix_lse,
+                                const torch::Tensor& suffix_output,
+                                const torch::Tensor& suffix_lse) {
+  constexpr uint NUM_THREADS = 128;
+  const uint num_tokens = output.size(0);
+  const uint num_heads = output.size(1);
+  const uint head_size = output.size(2);
+  const uint pack_size = 16 / sizeof(scalar_t);
+  TORCH_CHECK(head_size % pack_size == 0,
+              "headsize must be multiple of pack_size:", pack_size);
+  float* output_lse_ptr = nullptr;
+  if (output_lse.has_value()) {
+    output_lse_ptr = output_lse.value().data_ptr<float>();
+  }
+  // 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);
+
+  LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS);
+}
+
+#define CALL_MERGE_ATTN_STATES_LAUNCHER(scalar_t)                           \
+  {                                                                         \
+    merge_attn_states_launcher<scalar_t>(output, output_lse, prefix_output, \
+                                         prefix_lse, suffix_output,         \
+                                         suffix_lse);                       \
+  }
+
+void merge_attn_states(torch::Tensor& output,
+                       std::optional<torch::Tensor> output_lse,
+                       const torch::Tensor& prefix_output,
+                       const torch::Tensor& prefix_lse,
+                       const torch::Tensor& suffix_output,
+                       const torch::Tensor& suffix_lse) {
+  DISPATCH_BY_SCALAR_DTYPE(output.dtype(), CALL_MERGE_ATTN_STATES_LAUNCHER);
+}

csrc/cpu/utils.cpp

@@ -4,6 +4,11 @@
   #include <string>
   #include <sched.h>
 #endif
+#if __GLIBC__ == 2 && __GLIBC_MINOR__ < 30
+  #include <unistd.h>
+  #include <sys/syscall.h>
+  #define gettid() syscall(SYS_gettid)
+#endif
 
 #include "cpu_types.hpp"
 

csrc/mamba/causal_conv1d/causal_conv1d.cu

@@ -422,7 +422,7 @@ void causal_conv1d_fwd_kernel(ConvParamsBase params) {
         int final_state_position =  ((seqlen - (kWidth - 1)) - (n_chunks - 1) * kChunkSize);
         // in case the final state is separated between the last "smem_exchange" and 
         // and the one before it (chunk = n_chunks - 1 and chunk = n_chunks - 2), 
-        // (which occurs when `final_state_position` is a non-positivie index)
+        // (which occurs when `final_state_position` is a non-positive index)
         // we load the correct data from smem_exchange from both chunks, the last chunk iteration and the one before it
         if (conv_states != nullptr && final_state_position < 0 && seqlen > kWidth){
             input_t vals_load[kNElts] = {0};

csrc/ops.h

@@ -52,6 +52,15 @@ void paged_attention_v2(
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step);
 
+#ifndef USE_ROCM
+void merge_attn_states(torch::Tensor& output,
+                       std::optional<torch::Tensor> output_lse,
+                       const torch::Tensor& prefix_output,
+                       const torch::Tensor& prefix_lse,
+                       const torch::Tensor& suffix_output,
+                       const torch::Tensor& suffix_lse);
+#endif
+
 void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
               double epsilon);
 

csrc/quantization/gguf/moe.cuh

@@ -129,7 +129,7 @@ static __device__ __forceinline__ void moe_q(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q4_0 64
+  #define MOE_X_Q4_0 8
   #define MOE_Y_Q4_0 128
   #define NWARPS_Q4_0 8
 #else
@@ -190,7 +190,7 @@ static void ggml_moe_q4_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q4_1 64
+  #define MOE_X_Q4_1 8
   #define MOE_Y_Q4_1 128
   #define NWARPS_Q4_1 8
 #else
@@ -251,7 +251,7 @@ static void ggml_moe_q4_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q5_0 64
+  #define MOE_X_Q5_0 8
   #define MOE_Y_Q5_0 128
   #define NWARPS_Q5_0 8
 #else
@@ -312,7 +312,7 @@ static void ggml_moe_q5_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q5_1 64
+  #define MOE_X_Q5_1 8
   #define MOE_Y_Q5_1 128
   #define NWARPS_Q5_1 8
 #else
@@ -373,7 +373,7 @@ static void ggml_moe_q5_1_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q8_0 64
+  #define MOE_X_Q8_0 8
   #define MOE_Y_Q8_0 128
   #define NWARPS_Q8_0 8
 #else
@@ -434,7 +434,7 @@ static void ggml_moe_q8_0_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q2_K 64
+  #define MOE_X_Q2_K 8
   #define MOE_Y_Q2_K 128
   #define NWARPS_Q2_K 8
 #else
@@ -495,7 +495,7 @@ static void ggml_moe_q2_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q3_K 64
+  #define MOE_X_Q3_K 8
   #define MOE_Y_Q3_K 128
   #define NWARPS_Q3_K 8
 #else
@@ -556,7 +556,7 @@ static void ggml_moe_q3_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q4_K 64
+  #define MOE_X_Q4_K 8
   #define MOE_Y_Q4_K 128
   #define NWARPS_Q4_K 8
 #else
@@ -617,7 +617,7 @@ static void ggml_moe_q4_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q5_K 64
+  #define MOE_X_Q5_K 8
   #define MOE_Y_Q5_K 128
   #define NWARPS_Q5_K 8
 #else
@@ -678,7 +678,7 @@ static void ggml_moe_q5_K_q8_1_cuda(
 }
 
 #if defined(USE_ROCM)
-  #define MOE_X_Q6_K 64
+  #define MOE_X_Q6_K 8
   #define MOE_Y_Q6_K 128
   #define NWARPS_Q6_K 8
 #else

csrc/quantization/gptq_marlin/gptq_marlin.cu

@@ -1785,7 +1785,7 @@ __global__ void Marlin(
             <<<blocks, NUM_THREADS, max_shared_mem, stream>>>(                 \
                 A_ptr, B_ptr, C_ptr, C_tmp_ptr, s_ptr, zp_ptr, g_idx_ptr,      \
                 num_groups, prob_m, prob_n, prob_k, lda, locks,                \
-                use_atomic_add, use_fp32_reduce);                              \
+                part_use_atomic_add, use_fp32_reduce);                         \
       }                                                                        \
     }
 
@@ -2215,6 +2215,10 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* s,
       thread_m_blocks = exec_cfg.max_m_blocks;
     }
 
+    // atomic add reduce have better performance only when m * n is small
+    bool part_use_atomic_add =
+        use_atomic_add && div_ceil(prob_m, 64) * prob_n <= 2048;
+
     if (false) {
     }
     GPTQ_CALL_IF(vllm::kU4B8, 16, 4, 256)

csrc/torch_bindings.cpp

@@ -64,6 +64,21 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    int blocksparse_head_sliding_step) -> ()");
   ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);
 
+#ifndef USE_ROCM
+  // Merge attn states
+  // Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+  // can be used to combine partial attention results (in the split-KV case)
+  ops.def(
+      "merge_attn_states("
+      "    Tensor! output,"
+      "    Tensor!? output_lse,"
+      "    Tensor prefix_output,"
+      "    Tensor prefix_lse,"
+      "    Tensor suffix_output,"
+      "    Tensor suffix_lse) -> ()");
+  ops.impl("merge_attn_states", torch::kCUDA, &merge_attn_states);
+#endif
+
   // Activation ops
   // Activation function used in SwiGLU.
   ops.def("silu_and_mul(Tensor! out, Tensor input) -> ()");

docker/Dockerfile.cpu

@@ -18,6 +18,8 @@ WORKDIR /workspace/
 ARG PYTHON_VERSION=3.12
 ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
 
+ENV LD_PRELOAD=""
+
 # Install minimal dependencies and uv
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt,sharing=locked \
@@ -32,6 +34,7 @@ ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
 
 ENV PATH="/root/.local/bin:$PATH"
 ENV VIRTUAL_ENV="/opt/venv"
+ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
 RUN uv venv --python ${PYTHON_VERSION} --seed ${VIRTUAL_ENV}
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 

docker/Dockerfile.hpu

@@ -1,4 +1,4 @@
-FROM vault.habana.ai/gaudi-docker/1.19.1/ubuntu22.04/habanalabs/pytorch-installer-2.5.1:latest
+FROM vault.habana.ai/gaudi-docker/1.20.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
 
 COPY ./ /workspace/vllm
 

docker/Dockerfile.neuron

@@ -1,6 +1,6 @@
 # default base image
 # https://gallery.ecr.aws/neuron/pytorch-inference-neuronx
-ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.21.0-ubuntu22.04"
+ARG BASE_IMAGE="public.ecr.aws/neuron/pytorch-inference-neuronx:2.5.1-neuronx-py310-sdk2.22.0-ubuntu22.04"
 
 FROM $BASE_IMAGE
 
@@ -21,9 +21,9 @@ VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}/vllm
 
 RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
-RUN python3 -m pip install sentencepiece transformers==4.45.2 -U
-RUN python3 -m pip install neuronx-cc==2.16.345.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas tenacity
+RUN python3 -m pip install sentencepiece transformers==4.48.0 -U
+RUN python3 -m pip install neuronx-cc==2.17.194.0 --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
 RUN python3 -m pip install pytest
 
 # uninstall transformers-neuronx package explicitly to avoid version conflict

docs/source/community/meetups.md

@@ -4,6 +4,7 @@
 
 We host regular meetups in San Francisco Bay Area every 2 months. We will share the project updates from the vLLM team and have guest speakers from the industry to share their experience and insights. Please find the materials of our previous meetups below:
 
+- [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day), April 3rd 2025. [[Slides]](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
 - [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama), March 27th 2025. [[Slides]](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
 - [The first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg), March 16th 2025. [[Slides]](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
 - [The East Coast vLLM Meetup](https://lu.ma/7mu4k4xx), March 11th 2025. [[Slides]](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0)

docs/source/contributing/model/multimodal.md

@@ -79,6 +79,17 @@ Further update the model as follows:
             return inputs_embeds
     ```
 
+- Implement {meth}`~vllm.model_executor.models.interfaces.SupportsMultiModal.get_language_model` getter to provide stable access to the underlying language model.
+
+    ```python
+    class YourModelForImage2Seq(nn.Module):
+        ...
+
+        def get_language_model(self) -> torch.nn.Module:
+            # Change `language_model` according to your implementation.
+            return self.language_model
+    ```
+
 - Once the above steps are done, update the model class with the {class}`~vllm.model_executor.models.interfaces.SupportsMultiModal` interface.
 
   ```diff
@@ -110,17 +121,21 @@ def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
     return {"image": None, "video": 1}
 ```
 
-### Maximum number of placeholder feature tokens
+## 3. Specify dummy inputs
 
-Also, override the abstract method {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item`
-to return the maximum number of placeholder feature tokens per input item for each modality.
+Then, inherit {class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` to construct dummy inputs for
+HF processing as well as memory profiling.
 
-When calling the model, the output embeddings from the visual encoder are assigned to the input positions
-containing placeholder feature tokens. Therefore, the number of placeholder feature tokens should be equal
-to the size of the output embeddings.
+### For memory profiling
 
-:::::{tab-set}
-::::{tab-item} Basic example: LLaVA
+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 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
 :sync: llava
 
 Looking at the code of HF's `LlavaForConditionalGeneration`:
@@ -229,7 +244,7 @@ def get_num_image_tokens(
 ```
 
 Notice that the number of image tokens doesn't depend on the image width and height.
-So, we can calculate the maximum number of image tokens using any image size:
+We can simply use a dummy `image_size`:
 
 ```python
 def get_image_size_with_most_features(self) -> ImageSize:
@@ -237,33 +252,35 @@ def get_image_size_with_most_features(self) -> ImageSize:
     width = height = hf_config.image_size
     return ImageSize(width=width, height=height)
 
-def get_max_image_tokens(self) -> int:
-    target_width, target_height = self.get_image_size_with_most_features()
-
-    return self.get_num_image_tokens(
-        image_width=target_width,
-        image_height=target_height,
-    )
-```
-
-And thus, we can override the method as:
-
-```python
-def get_mm_max_tokens_per_item(
+def get_dummy_processor_inputs(
     self,
     seq_len: int,
     mm_counts: Mapping[str, int],
-) -> Mapping[str, int]:
-    return {"image": self.get_max_image_tokens()}
+) -> ProcessorInputs:
+    num_images = mm_counts.get("image", 0)
+
+    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()
+
+    mm_data = {
+        "image":
+        self._get_dummy_images(width=target_width,
+                               height=target_height,
+                               num_images=num_images)
+    }
+
+    return ProcessorInputs(
+        prompt_text=image_token * num_images,
+        mm_data=mm_data,
+    )
 ```
 
-:::{note}
-Our [actual code](gh-file:vllm/model_executor/models/llava.py) is more abstracted to support vision encoders other than CLIP.
 :::
 
-::::
-
-::::{tab-item} Non-consecutive feature tokens: Fuyu
+:::{tab-item} No input placeholders: Fuyu
 :sync: fuyu
 
 Looking at the code of HF's `FuyuForCausalLM`:
@@ -383,188 +400,16 @@ num_patches_per_dim_w = image_width // patch_width
 num_patches = num_patches_per_dim_h * num_patches_per_dim_w
 ```
 
-We can calculate this in vLLM using this code:
-
-```python
-def get_num_image_patches(
-    self,
-    *,
-    image_width: int,
-    image_height: int,
-) -> int:
-    image_processor = self.get_image_processor()
-    target_width = image_processor.size["width"]
-    target_height = image_processor.size["height"]
-    patch_width = image_processor.patch_size["width"]
-    patch_height = image_processor.patch_size["height"]
-
-    if not (image_width <= target_width and image_height <= target_height):
-        height_scale_factor = target_height / image_height
-        width_scale_factor = target_width / image_width
-        optimal_scale_factor = min(height_scale_factor, width_scale_factor)
-
-        image_height = int(image_height * optimal_scale_factor)
-        image_width = int(image_width * optimal_scale_factor)
-
-    ncols = math.ceil(image_width / patch_width)
-    nrows = math.ceil(image_height / patch_height)
-    return ncols * nrows
-```
-
-These image patches correspond to placeholder tokens (`|SPEAKER|`). However, the processor also
-inserts newline tokens (`|NEWLINE|`) as shown here:
-
-```python
-# https://github.com/huggingface/transformers/blob/v4.48.3/src/transformers/models/fuyu/image_processing_fuyu.py#L654-L670
-tensor_of_image_ids = torch.full(
-    [num_patches], image_placeholder_id, dtype=torch.int32, device=image_input.device
-)
-patches = self.patchify_image(image=image.unsqueeze(0)).squeeze(0)
-assert num_patches == patches.shape[0]
-
-if variable_sized:
-    # Now terminate each line with |NEWLINE|.
-    tensor_of_image_ids = tensor_of_image_ids.reshape(-1, image_width // patch_width)
-    newline_ids = torch.full(
-        [tensor_of_image_ids.shape[0], 1],
-        image_newline_id,
-        dtype=torch.int32,
-        device=image_input.device,
-    )
-    tensor_of_image_ids = torch.cat([tensor_of_image_ids, newline_ids], dim=1)
-    tensor_of_image_ids = tensor_of_image_ids.reshape(-1)
-```
-
-So, the layout of tokens for an image is:
-
-```
-|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
-|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
-...
-|SPEAKER||SPEAKER|...|SPEAKER||NEWLINE|
-```
-
-This makes the placeholder tokens non-consecutive in the prompt.
-Since vLLM requires the feature tokens to be consecutive, **we also treat the newline tokens as feature tokens**.
-
-So overall, the total number of feature tokens is
-
-```python
-def get_num_image_tokens(
-    self,
-    *,
-    image_width: int,
-    image_height: int,
-) -> int:
-    image_processor = self.get_image_processor()
-    target_width = image_processor.size["width"]
-    target_height = image_processor.size["height"]
-    patch_width = image_processor.patch_size["width"]
-    patch_height = image_processor.patch_size["height"]
-
-    if not (image_width <= target_width and image_height <= target_height):
-        height_scale_factor = target_height / image_height
-        width_scale_factor = target_width / image_width
-        optimal_scale_factor = min(height_scale_factor, width_scale_factor)
-
-        image_height = int(image_height * optimal_scale_factor)
-        image_width = int(image_width * optimal_scale_factor)
-
-    ncols = math.ceil(image_width / patch_width)
-    nrows = math.ceil(image_height / patch_height)
-    return (ncols + 1) * nrows
-```
-
-To calculate the maximum number of image tokens, recall that input images are first resized
-to fit within `image_processor.size`. The maximum possible dimensions of the image before
-being converted into patches is therefore equal to `image_processor.size`.
+These image patches correspond to placeholder tokens (`|SPEAKER|`). So, we just need to maximize the number of image patches. Since input images are first resized
+to fit within `image_processor.size`, we can maximize the number of image patches by inputting an image with size equal to `image_processor.size`.
 
 ```python
 def get_image_size_with_most_features(self) -> ImageSize:
     image_processor = self.get_image_processor()
     return ImageSize(width=image_processor.size["width"],
                         height=image_processor.size["height"])
-
-def get_max_image_tokens(self) -> int:
-    target_width, target_height = self.get_image_size_with_most_features()
-
-    return self.get_num_image_tokens(
-        image_width=target_width,
-        image_height=target_height,
-    )
 ```
 
-And thus, we can override the method as:
-
-```python
-def get_mm_max_tokens_per_item(
-    self,
-    seq_len: int,
-    mm_counts: Mapping[str, int],
-) -> Mapping[str, int]:
-    return {"image": self.get_max_image_tokens()}
-```
-
-:::{note}
-Our [actual code](gh-file:vllm/model_executor/models/fuyu.py) returns `ncols` and `nrows` directly instead of the total token count.
-This is because `ncols` and `nrows` are used to specify the layout of the feature tokens (as shown in Step 4 of this guide).
-:::
-
-::::
-:::::
-
-## 3. Specify dummy inputs
-
-Then, inherit {class}`~vllm.multimodal.profiling.BaseDummyInputsBuilder` to construct dummy inputs for
-HF processing as well as memory profiling.
-
-### For memory profiling
-
-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 input can be constructed based
-on the code for {meth}`~vllm.multimodal.processing.BaseProcessingInfo.get_mm_max_tokens_per_item`.
-
-::::{tab-set}
-:::{tab-item} Basic example: LLaVA
-:sync: llava
-
-Making use of the `get_image_size_with_most_features` method implemented in Step 2:
-
-```python
-def get_dummy_processor_inputs(
-    self,
-    seq_len: int,
-    mm_counts: Mapping[str, int],
-) -> ProcessorInputs:
-    num_images = mm_counts.get("image", 0)
-
-    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()
-
-    mm_data = {
-        "image":
-        self._get_dummy_images(width=target_width,
-                               height=target_height,
-                               num_images=num_images)
-    }
-
-    return ProcessorInputs(
-        prompt_text=image_token * num_images,
-        mm_data=mm_data,
-    )
-```
-
-:::
-
-:::{tab-item} No input placeholders: Fuyu
-:sync: fuyu
-
 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:
@@ -860,8 +705,8 @@ prompt_tokens, prompts_length = _tokenize_prompts_with_image_and_batch(
 )
 ```
 
-To accommodate this, instead of a string you can return an instance of {class}`~vllm.multimodal.processing.PromptUpdateDetails`
-with different `full` and `feature` attributes:
+To assign the vision embeddings to only the image tokens, instead of a string
+you can return an instance of {class}`~vllm.multimodal.processing.PromptUpdateDetails`:
 
 ```python
 hf_config = self.info.get_hf_config()
@@ -879,9 +724,9 @@ def get_replacement_fuyu(item_idx: int):
     image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
                     [_NEWLINE_TOKEN_ID]) * nrows
 
-    return PromptUpdateDetails(
-        full=image_tokens + [bos_token_id],
-        features=image_tokens,
+    return PromptUpdateDetails.select_token_id(
+        image_tokens + [bos_token_id],
+        embed_token_id=_IMAGE_TOKEN_ID,
     )
 ```
 
@@ -914,9 +759,9 @@ def _get_prompt_updates(
         image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
                         [_NEWLINE_TOKEN_ID]) * nrows
 
-        return PromptUpdateDetails(
-            full=image_tokens + [bos_token_id],
-            features=image_tokens,
+        return PromptUpdateDetails.select_token_id(
+            image_tokens + [bos_token_id],
+            embed_token_id=_IMAGE_TOKEN_ID,
         )
 
     return [

docs/source/deployment/docker.md

@@ -34,11 +34,11 @@ If you need to use those dependencies (having accepted the license terms),
 create a custom Dockerfile on top of the base image with an extra layer that installs them:
 
 ```Dockerfile
-FROM vllm/vllm-openai:v0.8.2
+FROM vllm/vllm-openai:v0.8.3
 
-# e.g. install the `audio` and `video` optional dependencies
+# e.g. install the `audio` optional dependencies
 # NOTE: Make sure the version of vLLM matches the base image!
-RUN uv pip install --system vllm[audio,video]==0.8.2
+RUN uv pip install --system vllm[audio]==0.8.3
 ```
 
 :::

docs/source/deployment/k8s.md

@@ -46,6 +46,7 @@ metadata:
 type: Opaque
 data:
   token: $(HF_TOKEN)
+EOF
 ```
 
 Next, start the vLLM server as a Kubernetes Deployment and Service:

docs/source/design/mm_processing.md

@@ -8,7 +8,7 @@ Here are the main features of {class}`~vllm.multimodal.processing.BaseMultiModal
 
 ## Prompt Update Detection
 
-One of the main responsibilies of HF processor is to update the prompt with placeholder tokens. For example:
+One of the main responsibilities of HF processor is to update the prompt with placeholder tokens. For example:
 
 - Insert feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size) at the start of the string.
 - Replace existing input placeholder tokens (e.g. `<image>` for a single image) with feature placeholder tokens (e.g. `<image><image>...<image>`, the number of which equals to the feature size).

docs/source/design/v1/torch_compile.md

@@ -126,7 +126,7 @@ Unfortunately, because auto-tuning takes quite a long time (from seconds to minu
 
 ## Cudagraph Capture
 
-vLLM's V1 architecture uses piecewise cudagraph. The full computation graph is split as mentioned above, and we only capture the cudagraph for the piece of graph between attention operations (including the first graph before any attention operation, and the last graph after all the attention operation). This is based on a common observation: computation between attentions are usually token-wise and easy to deal with for cudagraph; while the attention operation is non-trival to be cudagraph compatible. Thus, by running the attention operation in eager mode while the rest operations in cudagraph, we keep the flexibility of the attention operation.
+vLLM's V1 architecture uses piecewise cudagraph. The full computation graph is split as mentioned above, and we only capture the cudagraph for the piece of graph between attention operations (including the first graph before any attention operation, and the last graph after all the attention operation). This is based on a common observation: computation between attentions are usually token-wise and easy to deal with for cudagraph; while the attention operation is non-trivial to be cudagraph compatible. Thus, by running the attention operation in eager mode while the rest operations in cudagraph, we keep the flexibility of the attention operation.
 
 The piecewise cudagraph also has fine-grained memory management. The purpose is to only exclude the attention kernel from cudagraph, while keeping all the rest modules and the memory allocation operations in the cudagraph. This is why the attention operation in V1 has the output tensor as the input of the attention.
 

docs/source/features/quantization/bnb.md

@@ -19,17 +19,20 @@ And usually, these repositories have a config.json file that includes a quantiza
 
 ## Read quantized checkpoint
 
+For pre-quantized checkpoints, vLLM will try to infer the quantization method from the config file, so you don't need to explicitly specify the quantization argument.
+
 ```python
 from vllm import LLM
 import torch
 # unsloth/tinyllama-bnb-4bit is a pre-quantized checkpoint.
 model_id = "unsloth/tinyllama-bnb-4bit"
-llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True, \
-quantization="bitsandbytes")
+llm = LLM(model=model_id, dtype=torch.bfloat16, trust_remote_code=True)
 ```
 
 ## Inflight quantization: load as 4bit quantization
 
+For inflight 4bit quantization with BitsAndBytes, you need to explicitly specify the quantization argument.
+
 ```python
 from vllm import LLM
 import torch
@@ -40,7 +43,7 @@ quantization="bitsandbytes")
 
 ## OpenAI Compatible Server
 
-Append the following to your 4bit model arguments:
+Append the following to your model arguments for 4bit inflight quantization:
 
 ```console
 --quantization bitsandbytes

docs/source/features/quantization/gguf.md

@@ -29,7 +29,7 @@ vllm serve ./tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf --tokenizer TinyLlama/TinyLlam
 We recommend using the tokenizer from base model instead of GGUF model. Because the tokenizer conversion from GGUF is time-consuming and unstable, especially for some models with large vocab size.
 :::
 
-GGUF assumes that huggingface can convert the metadata to a config file. In case huggingface doesn't support your model you can manually create a config and pass it as hf-confing-path
+GGUF assumes that huggingface can convert the metadata to a config file. In case huggingface doesn't support your model you can manually create a config and pass it as hf-config-path
 
 ```console
 # If you model is not supported by huggingface you can manually provide a huggingface compatible config path

docs/source/features/quantization/index.md

@@ -18,4 +18,5 @@ int8
 fp8
 quark
 quantized_kvcache
+torchao
 :::

docs/source/features/quantization/supported_hardware.md

@@ -62,7 +62,7 @@ The table below shows the compatibility of various quantization implementations
   * ❌
   * ✅︎
   * ❌
-  * ❌
+  * ✅︎
 - * FP8 (W8A8)
   * ❌
   * ❌

docs/source/features/quantization/torchao.md

@@ -0,0 +1,34 @@
+# TorchAO
+
+TorchAO is an architecture optimization library for PyTorch, it provides high performance dtypes, optimization techniques and kernels for inference and training, featuring composability with native PyTorch features like torch.compile, FSDP etc.. Some benchmark numbers can be found [here](https://github.com/pytorch/ao/tree/main/torchao/quantization#benchmarks).
+
+We recommend installing the latest torchao nightly with
+
+```console
+# Install the latest TorchAO nightly build
+# Choose the CUDA version that matches your system (cu126, cu128, etc.)
+pip install --pre torchao>=10.0.0 --index-url https://download.pytorch.org/whl/nightly/cu126
+```
+
+## Quantizing HuggingFace Models
+You can quantize your own huggingface model with torchao, e.g. [transformers](https://huggingface.co/docs/transformers/main/en/quantization/torchao) and [diffusers](https://huggingface.co/docs/diffusers/en/quantization/torchao), and save the checkpoint to huggingface hub like [this](https://huggingface.co/jerryzh168/llama3-8b-int8wo) with the following example code:
+
+```Python
+import torch
+from transformers import TorchAoConfig, AutoModelForCausalLM, AutoTokenizer
+from torchao.quantization import Int8WeightOnlyConfig
+
+model_name = "meta-llama/Meta-Llama-3-8B"
+quantization_config = TorchAoConfig(Int8WeightOnlyConfig())
+quantized_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto", quantization_config=quantization_config)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+input_text = "What are we having for dinner?"
+input_ids = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+hub_repo = # YOUR HUB REPO ID
+tokenizer.push_to_hub(hub_repo)
+quantized_model.push_to_hub(hub_repo, safe_serialization=False)
+```
+
+Alternatively, you can use the TorchAO Quantization space for quantizing models with a simple UI.
+See: https://huggingface.co/spaces/medmekk/TorchAO_Quantization

docs/source/generate_examples.py

@@ -17,6 +17,7 @@ def fix_case(text: str) -> str:
         "cli": "CLI",
         "cpu": "CPU",
         "llm": "LLM",
+        "mae": "MAE",
         "tpu": "TPU",
         "aqlm": "AQLM",
         "gguf": "GGUF",
@@ -24,6 +25,7 @@ def fix_case(text: str) -> str:
         "rlhf": "RLHF",
         "vllm": "vLLM",
         "openai": "OpenAI",
+        "lmcache": "LMCache",
         "multilora": "MultiLoRA",
         "mlpspeculator": "MLPSpeculator",
         r"fp\d+": lambda x: x.group(0).upper(),  # e.g. fp16, fp32

docs/source/getting_started/installation/cpu/apple.inc.md

@@ -12,7 +12,7 @@ There are no pre-built wheels or images for this device, so you must build vLLM
 
 - OS: `macOS Sonoma` or later
 - SDK: `XCode 15.4` or later with Command Line Tools
-- Compiler: `Apple Clang >= 15.0.0` and `Apple Clang < 17.0.0`
+- Compiler: `Apple Clang >= 15.0.0`
 
 ## Set up using Python
 
@@ -51,14 +51,6 @@ If the build has error like the following snippet where standard C++ headers can
       1 error generated.
 ```
 
-If run with error like the following snippet you need to check clang version and install a compatible version.
-
-```text
-AttributeError: '_OpNamespace' '_C' object has no attribute 'silu_and_mul'
-```
-
-More information can be found in <gh-issue:15941>.
-
 ## Set up using Docker
 
 ### Pre-built images

docs/source/getting_started/v1_user_guide.md

@@ -156,10 +156,3 @@ vLLM V1 is currently optimized for decoder-only transformers. Models requiring
   cross-attention between separate encoder and decoder are not yet supported (e.g., `BartForConditionalGeneration`, `MllamaForConditionalGeneration`).
 
 For a complete list of supported models, see the [list of supported models](https://docs.vllm.ai/en/latest/models/supported_models.html).
-
-## Frequently Asked Questions
-
-**I'm using vLLM V1 and I'm getting CUDA OOM errors. What should I do?**
-The default `max_num_seqs` has been raised from `256` in V0 to `1024` in V1. If you encounter CUDA OOM only when using V1 engine, try setting a lower value of `max_num_seqs` or `gpu_memory_utilization`.
-
-On the other hand, if you get an error about insufficient memory for the cache blocks, you should increase `gpu_memory_utilization` as this indicates that your GPU has sufficient memory but you're not allocating enough to vLLM for KV cache blocks.

docs/source/models/extensions/tensorizer.md

@@ -9,7 +9,7 @@ shorter Pod startup times and CPU memory usage. Tensor encryption is also suppor
 
 For more information on CoreWeave's Tensorizer, please refer to
 [CoreWeave's Tensorizer documentation](https://github.com/coreweave/tensorizer). For more information on serializing a vLLM model, as well a general usage guide to using Tensorizer with vLLM, see
-the [vLLM example script](https://docs.vllm.ai/en/stable/getting_started/examples/offline_inference/tensorize_vllm_model.html).
+the [vLLM example script](https://docs.vllm.ai/en/latest/getting_started/examples/tensorize_vllm_model.html).
 
 :::{note}
 Note that to use this feature you will need to install `tensorizer` by running `pip install vllm[tensorizer]`.

docs/source/models/supported_models.md

@@ -24,7 +24,7 @@ vLLM also supports model implementations that are available in Transformers. Thi
 
 To check if the modeling backend is Transformers, you can simply do this:
 
-```python 
+```python
 from vllm import LLM
 llm = LLM(model=..., task="generate")  # Name or path of your model
 llm.apply_model(lambda model: print(type(model)))
@@ -55,7 +55,7 @@ If your model is neither supported natively by vLLM or Transformers, you can sti
 Simply set `trust_remote_code=True` and vLLM will run any model on the Model Hub that is compatible with Transformers.
 Provided that the model writer implements their model in a compatible way, this means that you can run new models before they are officially supported in Transformers or vLLM!
 
-```python 
+```python
 from vllm import LLM
 llm = LLM(model=..., task="generate", trust_remote_code=True)  # Name or path of your model
 llm.apply_model(lambda model: print(model.__class__))
@@ -160,6 +160,35 @@ If vLLM successfully returns text (for generative models) or hidden states (for
 Otherwise, please refer to [Adding a New Model](#new-model) for instructions on how to implement your model in vLLM.
 Alternatively, you can [open an issue on GitHub](https://github.com/vllm-project/vllm/issues/new/choose) to request vLLM support.
 
+#### Using a proxy
+
+Here are some tips for loading/downloading models from Hugging Face using a proxy:
+
+- Set the proxy globally for your session (or set it in the profile file):
+
+```shell
+export http_proxy=http://your.proxy.server:port
+export https_proxy=http://your.proxy.server:port
+```
+
+- Set the proxy for just the current command:
+
+```shell
+https_proxy=http://your.proxy.server:port huggingface-cli download <model_name>
+
+# or use vllm cmd directly
+https_proxy=http://your.proxy.server:port  vllm serve <model_name> --disable-log-requests
+```
+
+- Set the proxy in Python interpreter:
+
+```python
+import os
+
+os.environ['http_proxy'] = 'http://your.proxy.server:port'
+os.environ['https_proxy'] = 'http://your.proxy.server:port'
+```
+
 ### ModelScope
 
 To use models from [ModelScope](https://www.modelscope.cn) instead of Hugging Face Hub, set an environment variable:
@@ -233,9 +262,9 @@ See [this page](#generative-models) for more information on how to use generativ
   * `facebook/bart-base`, `facebook/bart-large-cnn`, etc.
   *
   *
-- * `ChatGLMModel`
+- * `ChatGLMModel`, `ChatGLMForConditionalGeneration`
   * ChatGLM
-  * `THUDM/chatglm2-6b`, `THUDM/chatglm3-6b`, etc.
+  * `THUDM/chatglm2-6b`, `THUDM/chatglm3-6b`, `ShieldLM-6B-chatglm3`, etc.
   * ✅︎
   * ✅︎
 - * `CohereForCausalLM`, `Cohere2ForCausalLM`
@@ -303,6 +332,11 @@ See [this page](#generative-models) for more information on how to use generativ
   * `THUDM/glm-4-9b-chat-hf`, etc.
   * ✅︎
   * ✅︎
+- * `Glm4ForCausalLM`
+  * GLM-4-0414
+  * `THUDM/GLM-4-32B-Chat-0414`, etc.
+  * ✅︎
+  * ✅︎
 - * `GPT2LMHeadModel`
   * GPT-2
   * `gpt2`, `gpt2-xl`, etc.
@@ -478,6 +512,16 @@ See [this page](#generative-models) for more information on how to use generativ
   * `Qwen/Qwen1.5-MoE-A2.7B`, `Qwen/Qwen1.5-MoE-A2.7B-Chat`, etc.
   *
   * ✅︎
+- * `Qwen3ForCausalLM`
+  * Qwen3
+  * `Qwen/Qwen3-8B`, etc.
+  * ✅︎
+  * ✅︎
+- * `Qwen3MoeForCausalLM`
+  * Qwen3MoE
+  * `Qwen/Qwen3-MoE-15B-A2B`, etc.
+  * ✅︎
+  * ✅︎
 - * `StableLmForCausalLM`
   * StableLM
   * `stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.
@@ -840,6 +884,13 @@ See [this page](#generative-models) for more information on how to use generativ
   *
   * ✅︎
   * ✅︎
+- * `Llama4ForConditionalGeneration`
+  * Llama-4-17B-Omni-Instruct
+  * T + I<sup>+</sup>
+  * `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8`, `meta-llama/Llama-4-Maverick-17B-128E-Instruct`, etc.
+  *
+  * ✅︎
+  * ✅︎
 - * `LlavaForConditionalGeneration`
   * LLaVA-1.5
   * T + I<sup>E+</sup>
@@ -973,6 +1024,13 @@ See [this page](#generative-models) for more information on how to use generativ
   *
   * ✅︎
   * ✅︎
+- * `SmolVLMForConditionalGeneration`
+  * SmolVLM2
+  * T + I
+  * `SmolVLM2-2.2B-Instruct`
+  *
+  * ✅︎
+  * ✅︎
 - * `UltravoxModel`
   * Ultravox
   * T + A<sup>E+</sup>
@@ -989,9 +1047,6 @@ See [this page](#generative-models) for more information on how to use generativ
 <sup>+</sup> Multiple items can be inputted per text prompt for this modality.
 
 :::{important}
-To use Gemma3 series models, you have to install Hugging Face Transformers library from source via
-`pip install git+https://github.com/huggingface/transformers`.
-
 Pan-and-scan image pre-processing is currently supported on V0 (but not V1).
 You can enable it by passing `--mm-processor-kwargs '{"do_pan_and_scan": True}'`.
 :::
@@ -1128,5 +1183,5 @@ We have the following levels of testing for models:
 
 1. **Strict Consistency**: We compare the output of the model with the output of the model in the HuggingFace Transformers library under greedy decoding. This is the most stringent test. Please refer to [models tests](https://github.com/vllm-project/vllm/blob/main/tests/models) for the models that have passed this test.
 2. **Output Sensibility**: We check if the output of the model is sensible and coherent, by measuring the perplexity of the output and checking for any obvious errors. This is a less stringent test.
-3. **Runtime Functionality**: We check if the model can be loaded and run without errors. This is the least stringent test. Please refer to [functionality tests](gh-dir:tests) and [examples](gh-dir:main/examples) for the models that have passed this test.
+3. **Runtime Functionality**: We check if the model can be loaded and run without errors. This is the least stringent test. Please refer to [functionality tests](gh-dir:tests) and [examples](gh-dir:examples) for the models that have passed this test.
 4. **Community Feedback**: We rely on the community to provide feedback on the models. If a model is broken or not working as expected, we encourage users to raise issues to report it or open pull requests to fix it. The rest of the models fall under this category.

examples/offline_inference/audio_language.py

@@ -47,7 +47,7 @@ def run_minicpmo(question: str, audio_count: int) -> ModelRequestData:
         model=model_name,
         trust_remote_code=True,
         max_model_len=4096,
-        max_num_seqs=5,
+        max_num_seqs=2,
         limit_mm_per_prompt={"audio": audio_count},
     )
 
@@ -199,13 +199,6 @@ def main(args):
     engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
     llm = LLM(**engine_args)
 
-    # To maintain code compatibility in this script, we add LoRA here.
-    # You can also add LoRA using:
-    # llm.generate(prompts, lora_request=lora_request,...)
-    if req_data.lora_requests:
-        for lora_request in req_data.lora_requests:
-            llm.llm_engine.add_lora(lora_request=lora_request)
-
     # We set temperature to 0.2 so that outputs can be different
     # even when all prompts are identical when running batch inference.
     sampling_params = SamplingParams(temperature=0.2,
@@ -226,8 +219,15 @@ def main(args):
     if args.num_prompts > 1:
         # Batch inference
         inputs = [inputs] * args.num_prompts
+    # Add LoRA request if applicable
+    lora_request = (req_data.lora_requests *
+                    args.num_prompts if req_data.lora_requests else None)
 
-    outputs = llm.generate(inputs, sampling_params=sampling_params)
+    outputs = llm.generate(
+        inputs,
+        sampling_params=sampling_params,
+        lora_request=lora_request,
+    )
 
     for o in outputs:
         generated_text = o.outputs[0].text

examples/offline_inference/eagle.py

@@ -7,89 +7,103 @@ from transformers import AutoTokenizer
 
 from vllm import LLM, SamplingParams
 
-parser = argparse.ArgumentParser()
 
-parser.add_argument(
-    "--dataset",
-    type=str,
-    default="./examples/data/gsm8k.jsonl",
-    help="downloaded from the eagle repo " \
-    "https://github.com/SafeAILab/EAGLE/blob/main/eagle/data/"
-)
-parser.add_argument("--max_num_seqs", type=int, default=8)
-parser.add_argument("--num_prompts", type=int, default=80)
-parser.add_argument("--num_spec_tokens", type=int, default=2)
-parser.add_argument("--tp", type=int, default=1)
-parser.add_argument("--draft_tp", type=int, default=1)
-parser.add_argument("--enforce_eager", action='store_true')
-parser.add_argument("--enable_chunked_prefill", action='store_true')
-parser.add_argument("--max_num_batched_tokens", type=int, default=2048)
-parser.add_argument("--temp", type=float, default=0)
+def load_prompts(dataset_path, num_prompts):
+    if os.path.exists(dataset_path):
+        prompts = []
+        try:
+            with open(dataset_path) as f:
+                for line in f:
+                    data = json.loads(line)
+                    prompts.append(data["turns"][0])
+        except Exception as e:
+            print(f"Error reading dataset: {e}")
+            return []
+    else:
+        prompts = [
+            "The future of AI is", "The president of the United States is"
+        ]
 
-args = parser.parse_args()
+    return prompts[:num_prompts]
 
-print(args)
 
-model_dir = "meta-llama/Meta-Llama-3-8B-Instruct"
-eagle_dir = "abhigoyal/EAGLE-LLaMA3-Instruct-8B-vllm"
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default="./examples/data/gsm8k.jsonl",
+        help="downloaded from the eagle repo " \
+        "https://github.com/SafeAILab/EAGLE/blob/main/eagle/data/"
+    )
+    parser.add_argument("--max_num_seqs", type=int, default=8)
+    parser.add_argument("--num_prompts", type=int, default=80)
+    parser.add_argument("--num_spec_tokens", type=int, default=2)
+    parser.add_argument("--tp", type=int, default=1)
+    parser.add_argument("--draft_tp", type=int, default=1)
+    parser.add_argument("--enforce_eager", action='store_true')
+    parser.add_argument("--enable_chunked_prefill", action='store_true')
+    parser.add_argument("--max_num_batched_tokens", type=int, default=2048)
+    parser.add_argument("--temp", type=float, default=0)
+    args = parser.parse_args()
 
-max_model_len = 2048
+    model_dir = "meta-llama/Meta-Llama-3-8B-Instruct"
+    eagle_dir = "abhigoyal/EAGLE-LLaMA3-Instruct-8B-vllm"
 
-tokenizer = AutoTokenizer.from_pretrained(model_dir)
+    max_model_len = 2048
 
-if os.path.exists(args.dataset):
-    prompts = []
-    num_prompts = args.num_prompts
-    with open(args.dataset) as f:
-        for line in f:
-            data = json.loads(line)
-            prompts.append(data["turns"][0])
-else:
-    prompts = ["The future of AI is", "The president of the United States is"]
+    tokenizer = AutoTokenizer.from_pretrained(model_dir)
 
-prompts = prompts[:args.num_prompts]
-num_prompts = len(prompts)
+    prompts = load_prompts(args.dataset, args.num_prompts)
 
-prompt_ids = [
-    tokenizer.apply_chat_template([{
-        "role": "user",
-        "content": prompt
-    }],
-                                  add_generation_prompt=True)
-    for prompt in prompts
-]
+    prompt_ids = [
+        tokenizer.apply_chat_template([{
+            "role": "user",
+            "content": prompt
+        }],
+                                      add_generation_prompt=True)
+        for prompt in prompts
+    ]
 
-llm = LLM(
-    model=model_dir,
-    trust_remote_code=True,
-    tensor_parallel_size=args.tp,
-    enable_chunked_prefill=args.enable_chunked_prefill,
-    max_num_batched_tokens=args.max_num_batched_tokens,
-    enforce_eager=args.enforce_eager,
-    max_model_len=max_model_len,
-    max_num_seqs=args.max_num_seqs,
-    gpu_memory_utilization=0.8,
-    speculative_config={
-        "model": eagle_dir,
-        "num_speculative_tokens": args.num_spec_tokens,
-        "draft_tensor_parallel_size": args.draft_tp,
-        "max_model_len": max_model_len,
-    },
-    disable_log_stats=False,
-)
+    llm = LLM(
+        model=model_dir,
+        trust_remote_code=True,
+        tensor_parallel_size=args.tp,
+        enable_chunked_prefill=args.enable_chunked_prefill,
+        max_num_batched_tokens=args.max_num_batched_tokens,
+        enforce_eager=args.enforce_eager,
+        max_model_len=max_model_len,
+        max_num_seqs=args.max_num_seqs,
+        gpu_memory_utilization=0.8,
+        speculative_config={
+            "method": "eagle",
+            "model": eagle_dir,
+            "num_speculative_tokens": args.num_spec_tokens,
+            "draft_tensor_parallel_size": args.draft_tp,
+            "max_model_len": max_model_len,
+        },
+        disable_log_stats=False,
+    )
 
-sampling_params = SamplingParams(temperature=args.temp, max_tokens=256)
+    sampling_params = SamplingParams(temperature=args.temp, max_tokens=256)
 
-outputs = llm.generate(prompt_token_ids=prompt_ids,
-                       sampling_params=sampling_params)
+    outputs = llm.generate(prompt_token_ids=prompt_ids,
+                           sampling_params=sampling_params)
 
-# calculate the average number of accepted tokens per forward pass, +1 is
-# to account for the token from the target model that's always going to be
-# accepted
-acceptance_counts = [0] * (args.num_spec_tokens + 1)
-for output in outputs:
-    for step, count in enumerate(output.metrics.spec_token_acceptance_counts):
-        acceptance_counts[step] += count
+    # calculate the average number of accepted tokens per forward pass, +1 is
+    # to account for the token from the target model that's always going to be
+    # accepted
+    acceptance_counts = [0] * (args.num_spec_tokens + 1)
+    for output in outputs:
+        for step, count in enumerate(
+                output.metrics.spec_token_acceptance_counts):
+            acceptance_counts[step] += count
 
-print(f"mean acceptance length: \
-    {sum(acceptance_counts) / acceptance_counts[0]:.2f}")
+    print("-" * 50)
+    print(f"mean acceptance length: \
+        {sum(acceptance_counts) / acceptance_counts[0]:.2f}")
+    print("-" * 50)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/embed_jina_embeddings_v3.py

@@ -0,0 +1,50 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from argparse import Namespace
+
+from vllm import LLM, EngineArgs
+from vllm.utils import FlexibleArgumentParser
+
+
+def main(args: Namespace):
+    # Sample prompts.
+    prompts = [
+        "Follow the white rabbit.",  # English
+        "Sigue al conejo blanco.",  # Spanish
+        "Suis le lapin blanc.",  # French
+        "跟着白兔走。",  # Chinese
+        "اتبع الأرنب الأبيض.",  # Arabic
+        "Folge dem weißen Kaninchen.",  # German
+    ]
+
+    # Create an LLM.
+    # You should pass task="embed" for embedding models
+    model = LLM(**vars(args))
+
+    # Generate embedding. The output is a list of EmbeddingRequestOutputs.
+    # Only text matching task is supported for now. See #16120
+    outputs = model.embed(prompts)
+
+    # Print the outputs.
+    print("\nGenerated Outputs:")
+    print("Only text matching task is supported for now. See #16120")
+    print("-" * 60)
+    for prompt, output in zip(prompts, outputs):
+        embeds = output.outputs.embedding
+        embeds_trimmed = ((str(embeds[:16])[:-1] +
+                           ", ...]") if len(embeds) > 16 else embeds)
+        print(f"Prompt: {prompt!r} \n"
+              f"Embeddings for text matching: {embeds_trimmed} "
+              f"(size={len(embeds)})")
+        print("-" * 60)
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser = EngineArgs.add_cli_args(parser)
+    # Set example specific arguments
+    parser.set_defaults(model="jinaai/jina-embeddings-v3",
+                        task="embed",
+                        trust_remote_code=True)
+    args = parser.parse_args()
+    main(args)

examples/offline_inference/encoder_decoder.py

@@ -75,8 +75,6 @@ prompts = [
     enc_dec_prompt1, enc_dec_prompt2, enc_dec_prompt3
 ] + zipped_prompt_list
 
-print(prompts)
-
 # Create a sampling params object.
 sampling_params = SamplingParams(
     temperature=0,
@@ -91,10 +89,13 @@ sampling_params = SamplingParams(
 outputs = llm.generate(prompts, sampling_params)
 
 # Print the outputs.
-for output in outputs:
+print("-" * 50)
+for i, output in enumerate(outputs):
     prompt = output.prompt
     encoder_prompt = output.encoder_prompt
     generated_text = output.outputs[0].text
-    print(f"Encoder prompt: {encoder_prompt!r}, "
-          f"Decoder prompt: {prompt!r}, "
+    print(f"Output {i+1}:")
+    print(f"Encoder prompt: {encoder_prompt!r}\n"
+          f"Decoder prompt: {prompt!r}\n"
           f"Generated text: {generated_text!r}")
+    print("-" * 50)

examples/offline_inference/encoder_decoder_multimodal.py

@@ -56,7 +56,7 @@ def run_florence2():
 def run_mllama():
     engine_args = EngineArgs(
         model="meta-llama/Llama-3.2-11B-Vision-Instruct",
-        max_model_len=4096,
+        max_model_len=8192,
         max_num_seqs=2,
         limit_mm_per_prompt={"image": 1},
         dtype="half",

examples/offline_inference/llm_engine_example.py

@@ -1,5 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
-
+"""
+This file demonstrates using the `LLMEngine`
+for processing prompts with various sampling parameters.
+"""
 import argparse
 
 from vllm import EngineArgs, LLMEngine, RequestOutput, SamplingParams
@@ -26,6 +29,7 @@ def process_requests(engine: LLMEngine,
     """Continuously process a list of prompts and handle the outputs."""
     request_id = 0
 
+    print('-' * 50)
     while test_prompts or engine.has_unfinished_requests():
         if test_prompts:
             prompt, sampling_params = test_prompts.pop(0)
@@ -37,6 +41,7 @@ def process_requests(engine: LLMEngine,
         for request_output in request_outputs:
             if request_output.finished:
                 print(request_output)
+                print('-' * 50)
 
 
 def initialize_engine(args: argparse.Namespace) -> LLMEngine:

examples/offline_inference/mistral-small.py

@@ -13,9 +13,14 @@ from vllm.sampling_params import SamplingParams
 # - Server:
 #
 # ```bash
+# # Mistral format
 # vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 \
 #   --tokenizer-mode mistral --config-format mistral --load-format mistral \
 #   --limit-mm-per-prompt 'image=4' --max-model-len 16384
+#
+# # HF format
+# vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 \
+#   --limit-mm-per-prompt 'image=4' --max-model-len 16384
 # ```
 #
 # - Client:
@@ -44,19 +49,22 @@ from vllm.sampling_params import SamplingParams
 #     python demo.py simple
 #     python demo.py advanced
 
+# Lower max_model_len and/or max_num_seqs on low-VRAM GPUs.
+# These scripts have been tested on 2x L40 GPUs
+
 
 def run_simple_demo(args: argparse.Namespace):
     model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
     sampling_params = SamplingParams(max_tokens=8192)
 
-    # Lower max_model_len and/or max_num_seqs on low-VRAM GPUs.
     llm = LLM(
         model=model_name,
-        tokenizer_mode="mistral",
-        config_format="mistral",
-        load_format="mistral",
+        tokenizer_mode="mistral" if args.format == "mistral" else "auto",
+        config_format="mistral" if args.format == "mistral" else "auto",
+        load_format="mistral" if args.format == "mistral" else "auto",
         max_model_len=4096,
         max_num_seqs=2,
+        tensor_parallel_size=2,
         disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
     )
 
@@ -82,23 +90,25 @@ def run_simple_demo(args: argparse.Namespace):
         },
     ]
     outputs = llm.chat(messages, sampling_params=sampling_params)
-
+    print("-" * 50)
     print(outputs[0].outputs[0].text)
+    print("-" * 50)
 
 
 def run_advanced_demo(args: argparse.Namespace):
     model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
-    max_img_per_msg = 5
+    max_img_per_msg = 3
     max_tokens_per_img = 4096
 
     sampling_params = SamplingParams(max_tokens=8192, temperature=0.7)
     llm = LLM(
         model=model_name,
-        tokenizer_mode="mistral",
-        config_format="mistral",
-        load_format="mistral",
+        tokenizer_mode="mistral" if args.format == "mistral" else "auto",
+        config_format="mistral" if args.format == "mistral" else "auto",
+        load_format="mistral" if args.format == "mistral" else "auto",
         limit_mm_per_prompt={"image": max_img_per_msg},
         max_model_len=max_img_per_msg * max_tokens_per_img,
+        tensor_parallel_size=2,
         disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
     )
 
@@ -153,7 +163,9 @@ def run_advanced_demo(args: argparse.Namespace):
     ]
 
     outputs = llm.chat(messages=messages, sampling_params=sampling_params)
+    print("-" * 50)
     print(outputs[0].outputs[0].text)
+    print("-" * 50)
 
 
 def main():
@@ -166,6 +178,11 @@ def main():
         help="Specify the demo mode: 'simple' or 'advanced'",
     )
 
+    parser.add_argument('--format',
+                        choices=["mistral", "hf"],
+                        default="mistral",
+                        help='Specify the format of the model to load.')
+
     parser.add_argument(
         '--disable-mm-preprocessor-cache',
         action='store_true',

examples/offline_inference/mlpspeculator.py

@@ -1,4 +1,11 @@
 # SPDX-License-Identifier: Apache-2.0
+"""
+This file demonstrates the usage of text generation with an LLM model,
+comparing the performance with and without speculative decoding.
+
+Note that still not support `v1`:
+VLLM_USE_V1=0 python examples/offline_inference/mlpspeculator.py
+"""
 
 import gc
 import time
@@ -7,7 +14,7 @@ from vllm import LLM, SamplingParams
 
 
 def time_generation(llm: LLM, prompts: list[str],
-                    sampling_params: SamplingParams):
+                    sampling_params: SamplingParams, title: str):
     # Generate texts from the prompts. The output is a list of RequestOutput
     # objects that contain the prompt, generated text, and other information.
     # Warmup first
@@ -16,11 +23,15 @@ def time_generation(llm: LLM, prompts: list[str],
     start = time.time()
     outputs = llm.generate(prompts, sampling_params)
     end = time.time()
-    print((end - start) / sum([len(o.outputs[0].token_ids) for o in outputs]))
+    print("-" * 50)
+    print(title)
+    print("time: ",
+          (end - start) / sum(len(o.outputs[0].token_ids) for o in outputs))
     # Print the outputs.
     for output in outputs:
         generated_text = output.outputs[0].text
         print(f"text: {generated_text!r}")
+        print("-" * 50)
 
 
 if __name__ == "__main__":
@@ -41,8 +52,7 @@ if __name__ == "__main__":
     # Create an LLM without spec decoding
     llm = LLM(model="meta-llama/Llama-2-13b-chat-hf")
 
-    print("Without speculation")
-    time_generation(llm, prompts, sampling_params)
+    time_generation(llm, prompts, sampling_params, "Without speculation")
 
     del llm
     gc.collect()
@@ -55,5 +65,4 @@ if __name__ == "__main__":
         },
     )
 
-    print("With speculation")
-    time_generation(llm, prompts, sampling_params)
+    time_generation(llm, prompts, sampling_params, "With speculation")

examples/offline_inference/multilora_inference.py

@@ -61,6 +61,7 @@ def process_requests(engine: LLMEngine,
     """Continuously process a list of prompts and handle the outputs."""
     request_id = 0
 
+    print("-" * 50)
     while test_prompts or engine.has_unfinished_requests():
         if test_prompts:
             prompt, sampling_params, lora_request = test_prompts.pop(0)
@@ -75,6 +76,7 @@ def process_requests(engine: LLMEngine,
         for request_output in request_outputs:
             if request_output.finished:
                 print(request_output)
+                print("-" * 50)
 
 
 def initialize_engine() -> LLMEngine:

examples/offline_inference/neuron.py

@@ -12,27 +12,36 @@ prompts = [
 # Create a sampling params object.
 sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
-# Create an LLM.
-llm = LLM(
-    model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
-    max_num_seqs=8,
-    # The max_model_len and block_size arguments are required to be same as
-    # max sequence length when targeting neuron device.
-    # Currently, this is a known limitation in continuous batching support
-    # in transformers-neuronx.
-    # TODO(liangfu): Support paged-attention in transformers-neuronx.
-    max_model_len=1024,
-    block_size=1024,
-    # The device can be automatically detected when AWS Neuron SDK is installed.
-    # The device argument can be either unspecified for automated detection,
-    # or explicitly assigned.
-    device="neuron",
-    tensor_parallel_size=2)
-# Generate texts from the prompts. The output is a list of RequestOutput objects
-# that contain the prompt, generated text, and other information.
-outputs = llm.generate(prompts, sampling_params)
-# Print the outputs.
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+def main():
+    # Create an LLM.
+    llm = LLM(
+        model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+        max_num_seqs=8,
+        # The max_model_len and block_size arguments are required to be same as
+        # max sequence length when targeting neuron device.
+        # Currently, this is a known limitation in continuous batching support
+        # in transformers-neuronx.
+        # TODO(liangfu): Support paged-attention in transformers-neuronx.
+        max_model_len=1024,
+        block_size=1024,
+        # ruff: noqa: E501
+        # The device can be automatically detected when AWS Neuron SDK is installed.
+        # The device argument can be either unspecified for automated detection,
+        # or explicitly assigned.
+        device="neuron",
+        tensor_parallel_size=2)
+    # Generate texts from the prompts. The output is a list of RequestOutput objects
+    # that contain the prompt, generated text, and other information.
+    outputs = llm.generate(prompts, sampling_params)
+    # Print the outputs.
+    print("-" * 50)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/neuron_int8_quantization.py

@@ -22,31 +22,40 @@ prompts = [
 # Create a sampling params object.
 sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
-# Create an LLM.
-llm = LLM(
-    model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
-    max_num_seqs=8,
-    # The max_model_len and block_size arguments are required to be same as
-    # max sequence length when targeting neuron device.
-    # Currently, this is a known limitation in continuous batching support
-    # in transformers-neuronx.
-    # TODO(liangfu): Support paged-attention in transformers-neuronx.
-    max_model_len=2048,
-    block_size=2048,
-    # The device can be automatically detected when AWS Neuron SDK is installed.
-    # The device argument can be either unspecified for automated detection,
-    # or explicitly assigned.
-    device="neuron",
-    quantization="neuron_quant",
-    override_neuron_config={
-        "cast_logits_dtype": "bfloat16",
-    },
-    tensor_parallel_size=2)
-# Generate texts from the prompts. The output is a list of RequestOutput objects
-# that contain the prompt, generated text, and other information.
-outputs = llm.generate(prompts, sampling_params)
-# Print the outputs.
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+def main():
+    # Create an LLM.
+    llm = LLM(
+        model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+        max_num_seqs=8,
+        # The max_model_len and block_size arguments are required to be same as
+        # max sequence length when targeting neuron device.
+        # Currently, this is a known limitation in continuous batching support
+        # in transformers-neuronx.
+        # TODO(liangfu): Support paged-attention in transformers-neuronx.
+        max_model_len=2048,
+        block_size=2048,
+        # ruff: noqa: E501
+        # The device can be automatically detected when AWS Neuron SDK is installed.
+        # The device argument can be either unspecified for automated detection,
+        # or explicitly assigned.
+        device="neuron",
+        quantization="neuron_quant",
+        override_neuron_config={
+            "cast_logits_dtype": "bfloat16",
+        },
+        tensor_parallel_size=2)
+    # Generate texts from the prompts. The output is a list of RequestOutput objects
+    # that contain the prompt, generated text, and other information.
+    outputs = llm.generate(prompts, sampling_params)
+    # Print the outputs.
+    print("-" * 50)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/prefix_caching.py

@@ -31,55 +31,62 @@ generating_prompts = [prefix + prompt for prompt in prompts]
 # Create a sampling params object.
 sampling_params = SamplingParams(temperature=0.0)
 
-# Create an LLM without prefix caching as a baseline.
-regular_llm = LLM(model="facebook/opt-125m", gpu_memory_utilization=0.4)
 
-print("Results without `enable_prefix_caching`")
+def main():
+    # Create an LLM without prefix caching as a baseline.
+    regular_llm = LLM(model="facebook/opt-125m", gpu_memory_utilization=0.4)
 
-# Generate texts from the prompts. The output is a list of RequestOutput objects
-# that contain the prompt, generated text, and other information.
-outputs = regular_llm.generate(generating_prompts, sampling_params)
+    print("Results without `enable_prefix_caching`")
 
-regular_generated_texts = []
-# Print the outputs.
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    regular_generated_texts.append(generated_text)
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    # ruff: noqa: E501
+    # Generate texts from the prompts. The output is a list of RequestOutput objects
+    # that contain the prompt, generated text, and other information.
+    outputs = regular_llm.generate(generating_prompts, sampling_params)
 
-print("-" * 80)
+    regular_generated_texts = []
+    # Print the outputs.
+    print("-" * 50)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        regular_generated_texts.append(generated_text)
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
 
-# Destroy the LLM object and free up the GPU memory.
-del regular_llm
-cleanup_dist_env_and_memory()
+    # Destroy the LLM object and free up the GPU memory.
+    del regular_llm
+    cleanup_dist_env_and_memory()
 
-# Create an LLM with prefix caching enabled.
-prefix_cached_llm = LLM(model="facebook/opt-125m",
-                        enable_prefix_caching=True,
-                        gpu_memory_utilization=0.4)
+    # Create an LLM with prefix caching enabled.
+    prefix_cached_llm = LLM(model="facebook/opt-125m",
+                            enable_prefix_caching=True,
+                            gpu_memory_utilization=0.4)
 
-# Warmup so that the shared prompt's KV cache is computed.
-prefix_cached_llm.generate(generating_prompts[0], sampling_params)
+    # Warmup so that the shared prompt's KV cache is computed.
+    prefix_cached_llm.generate(generating_prompts[0], sampling_params)
 
-# Generate with prefix caching.
-outputs = prefix_cached_llm.generate(generating_prompts, sampling_params)
+    # Generate with prefix caching.
+    outputs = prefix_cached_llm.generate(generating_prompts, sampling_params)
 
-print("Results with `enable_prefix_caching`")
+    print("Results with `enable_prefix_caching`")
 
-cached_generated_texts = []
-# Print the outputs. You should see the same outputs as before.
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    cached_generated_texts.append(generated_text)
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+    cached_generated_texts = []
+    # Print the outputs. You should see the same outputs as before.
+    print("-" * 50)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        cached_generated_texts.append(generated_text)
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
 
-print("-" * 80)
+    # Compare the results and display the speedup
+    generated_same = all([
+        regular_generated_texts[i] == cached_generated_texts[i]
+        for i in range(len(prompts))
+    ])
+    print(f"Generated answers are the same: {generated_same}")
 
-# Compare the results and display the speedup
-generated_same = all([
-    regular_generated_texts[i] == cached_generated_texts[i]
-    for i in range(len(prompts))
-])
-print(f"Generated answers are the same: {generated_same}")
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/profiling.py

@@ -12,7 +12,7 @@ from typing import Any, Optional, TypeAlias
 import torch
 import tqdm
 
-from vllm import LLM, SamplingParams
+from vllm import LLM, SamplingParams, envs
 from vllm.engine.arg_utils import EngineArgs
 from vllm.profiler import layerwise_profile
 from vllm.utils import FlexibleArgumentParser
@@ -234,9 +234,8 @@ def run_profile(context: ProfileContext, csv_output: Optional[str],
             sampling_params.max_tokens = next(output_len_generator)
             assert isinstance(sampling_params.max_tokens, int)
 
-            prompt_token_ids = torch.randint(
-                llm.llm_engine.model_config.get_vocab_size(),
-                size=(prompt_len, )).tolist()
+            prompt_token_ids = torch.randint(llm.get_tokenizer().vocab_size,
+                                             size=(prompt_len, )).tolist()
 
             llm.llm_engine.add_request(
                 request_id=f"seq{i}",
@@ -262,8 +261,13 @@ def run_profile(context: ProfileContext, csv_output: Optional[str],
 
     decode_profs = []
     for _ in tqdm.tqdm(range(num_steps_to_profile - 1)):
-        num_running_seqs = llm.llm_engine.scheduler[
-            0].get_num_unfinished_seq_groups()
+        if envs.VLLM_USE_V1:
+            num_running_seqs = llm.llm_engine.scheduler[
+                0].get_num_unfinished_requests()
+        else:
+            num_running_seqs = llm.llm_engine.scheduler[
+                0].get_num_unfinished_seq_groups()
+
         with layerwise_profile(
                 num_running_seqs=num_running_seqs) as decode_prof:
             llm.llm_engine.step()

examples/offline_inference/reproduciblity.py

@@ -19,8 +19,6 @@ SEED = 42
 # because it is almost impossible to make the scheduling deterministic in the
 # online serving setting.
 
-llm = LLM(model="facebook/opt-125m", seed=SEED)
-
 prompts = [
     "Hello, my name is",
     "The president of the United States is",
@@ -29,8 +27,17 @@ prompts = [
 ]
 sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
-outputs = llm.generate(prompts, sampling_params)
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+
+def main():
+    llm = LLM(model="facebook/opt-125m", seed=SEED)
+    outputs = llm.generate(prompts, sampling_params)
+    print("-" * 50)
+    for output in outputs:
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/rlhf.py

@@ -85,11 +85,13 @@ sampling_params = SamplingParams(temperature=0)
 
 outputs = ray.get(llm.generate.remote(prompts, sampling_params))
 
+print("-" * 50)
 for output in outputs:
     prompt = output.prompt
     generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, "
+    print(f"Prompt: {prompt!r}\n"
           f"Generated text: {generated_text!r}")
+    print("-" * 50)
 
 # set up the communication between the training process
 # and the inference engine.
@@ -120,8 +122,10 @@ assert all(ray.get(llm.collective_rpc.remote("check_weights_changed")))
 # use the updated model to generate texts, they will be nonsense
 # because the weights are all zeros.
 outputs_updated = ray.get(llm.generate.remote(prompts, sampling_params))
+print("-" * 50)
 for output in outputs_updated:
     prompt = output.prompt
     generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, "
+    print(f"Prompt: {prompt!r}\n"
           f"Generated text: {generated_text!r}")
+    print("-" * 50)

examples/offline_inference/simple_profiling.py

@@ -32,10 +32,12 @@ if __name__ == "__main__":
     llm.stop_profile()
 
     # Print the outputs.
+    print("-" * 50)
     for output in outputs:
         prompt = output.prompt
         generated_text = output.outputs[0].text
-        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        print("-" * 50)
 
     # Add a buffer to wait for profiler in the background process
     # (in case MP is on) to finish writing profiling output.

examples/offline_inference/structured_outputs.py

@@ -1,4 +1,11 @@
 # SPDX-License-Identifier: Apache-2.0
+"""
+This file demonstrates the example usage of guided decoding 
+to generate structured outputs using vLLM. It shows how to apply 
+different guided decoding techniques such as Choice, Regex, JSON schema, 
+and Grammar to produce structured and formatted results 
+based on specific prompts.
+"""
 
 from enum import Enum
 
@@ -7,26 +14,21 @@ from pydantic import BaseModel
 from vllm import LLM, SamplingParams
 from vllm.sampling_params import GuidedDecodingParams
 
-llm = LLM(model="Qwen/Qwen2.5-3B-Instruct", max_model_len=100)
-
 # Guided decoding by Choice (list of possible options)
-guided_decoding_params = GuidedDecodingParams(choice=["Positive", "Negative"])
-sampling_params = SamplingParams(guided_decoding=guided_decoding_params)
-outputs = llm.generate(
-    prompts="Classify this sentiment: vLLM is wonderful!",
-    sampling_params=sampling_params,
-)
-print(outputs[0].outputs[0].text)
+guided_decoding_params_choice = GuidedDecodingParams(
+    choice=["Positive", "Negative"])
+sampling_params_choice = SamplingParams(
+    guided_decoding=guided_decoding_params_choice)
+prompt_choice = "Classify this sentiment: vLLM is wonderful!"
 
 # Guided decoding by Regex
-guided_decoding_params = GuidedDecodingParams(regex="\w+@\w+\.com\n")
-sampling_params = SamplingParams(guided_decoding=guided_decoding_params,
-                                 stop=["\n"])
-prompt = ("Generate an email address for Alan Turing, who works in Enigma."
-          "End in .com and new line. Example result:"
-          "alan.turing@enigma.com\n")
-outputs = llm.generate(prompts=prompt, sampling_params=sampling_params)
-print(outputs[0].outputs[0].text)
+guided_decoding_params_regex = GuidedDecodingParams(regex=r"\w+@\w+\.com\n")
+sampling_params_regex = SamplingParams(
+    guided_decoding=guided_decoding_params_regex, stop=["\n"])
+prompt_regex = (
+    "Generate an email address for Alan Turing, who works in Enigma."
+    "End in .com and new line. Example result:"
+    "alan.turing@enigma.com\n")
 
 
 # Guided decoding by JSON using Pydantic schema
@@ -44,37 +46,54 @@ class CarDescription(BaseModel):
 
 
 json_schema = CarDescription.model_json_schema()
-
-guided_decoding_params = GuidedDecodingParams(json=json_schema)
-sampling_params = SamplingParams(guided_decoding=guided_decoding_params)
-prompt = ("Generate a JSON with the brand, model and car_type of"
-          "the most iconic car from the 90's")
-outputs = llm.generate(
-    prompts=prompt,
-    sampling_params=sampling_params,
-)
-print(outputs[0].outputs[0].text)
+guided_decoding_params_json = GuidedDecodingParams(json=json_schema)
+sampling_params_json = SamplingParams(
+    guided_decoding=guided_decoding_params_json)
+prompt_json = ("Generate a JSON with the brand, model and car_type of"
+               "the most iconic car from the 90's")
 
 # Guided decoding by Grammar
 simplified_sql_grammar = """
-    ?start: select_statement
-
-    ?select_statement: "SELECT " column_list " FROM " table_name
-
-    ?column_list: column_name ("," column_name)*
-
-    ?table_name: identifier
-
-    ?column_name: identifier
-
-    ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
+root ::= select_statement
+select_statement ::= "SELECT " column " from " table " where " condition
+column ::= "col_1 " | "col_2 "
+table ::= "table_1 " | "table_2 "
+condition ::= column "= " number
+number ::= "1 " | "2 "
 """
-guided_decoding_params = GuidedDecodingParams(grammar=simplified_sql_grammar)
-sampling_params = SamplingParams(guided_decoding=guided_decoding_params)
-prompt = ("Generate an SQL query to show the 'username' and 'email'"
-          "from the 'users' table.")
-outputs = llm.generate(
-    prompts=prompt,
-    sampling_params=sampling_params,
-)
-print(outputs[0].outputs[0].text)
+guided_decoding_params_grammar = GuidedDecodingParams(
+    grammar=simplified_sql_grammar)
+sampling_params_grammar = SamplingParams(
+    guided_decoding=guided_decoding_params_grammar)
+prompt_grammar = ("Generate an SQL query to show the 'username' and 'email'"
+                  "from the 'users' table.")
+
+
+def format_output(title: str, output: str):
+    print(f"{'-' * 50}\n{title}: {output}\n{'-' * 50}")
+
+
+def generate_output(prompt: str, sampling_params: SamplingParams, llm: LLM):
+    outputs = llm.generate(prompts=prompt, sampling_params=sampling_params)
+    return outputs[0].outputs[0].text
+
+
+def main():
+    llm = LLM(model="Qwen/Qwen2.5-3B-Instruct", max_model_len=100)
+
+    choice_output = generate_output(prompt_choice, sampling_params_choice, llm)
+    format_output("Guided decoding by Choice", choice_output)
+
+    regex_output = generate_output(prompt_regex, sampling_params_regex, llm)
+    format_output("Guided decoding by Regex", regex_output)
+
+    json_output = generate_output(prompt_json, sampling_params_json, llm)
+    format_output("Guided decoding by JSON", json_output)
+
+    grammar_output = generate_output(prompt_grammar, sampling_params_grammar,
+                                     llm)
+    format_output("Guided decoding by Grammar", grammar_output)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/torchrun_example.py

@@ -36,11 +36,13 @@ llm = LLM(
 outputs = llm.generate(prompts, sampling_params)
 
 # all ranks will have the same outputs
+print("-" * 50)
 for output in outputs:
     prompt = output.prompt
     generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, "
+    print(f"Prompt: {prompt!r}\n"
           f"Generated text: {generated_text!r}")
+    print("-" * 50)
 """
 Further tips:
 

examples/offline_inference/tpu.py

@@ -16,14 +16,22 @@ N = 1
 # Currently, top-p sampling is disabled. `top_p` should be 1.0.
 sampling_params = SamplingParams(temperature=0, top_p=1.0, n=N, max_tokens=16)
 
-# Set `enforce_eager=True` to avoid ahead-of-time compilation.
-# In real workloads, `enforace_eager` should be `False`.
-llm = LLM(model="Qwen/Qwen2-1.5B-Instruct",
-          max_num_batched_tokens=64,
-          max_num_seqs=4)
-outputs = llm.generate(prompts, sampling_params)
-for output, answer in zip(outputs, answers):
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-    assert generated_text.startswith(answer)
+
+def main():
+    # Set `enforce_eager=True` to avoid ahead-of-time compilation.
+    # In real workloads, `enforace_eager` should be `False`.
+    llm = LLM(model="Qwen/Qwen2-1.5B-Instruct",
+              max_num_batched_tokens=64,
+              max_num_seqs=4)
+    outputs = llm.generate(prompts, sampling_params)
+    print("-" * 50)
+    for output, answer in zip(outputs, answers):
+        prompt = output.prompt
+        generated_text = output.outputs[0].text
+        print(f"Prompt: {prompt!r}\nGenerated text: {generated_text!r}")
+        assert generated_text.startswith(answer)
+        print("-" * 50)
+
+
+if __name__ == "__main__":
+    main()

examples/offline_inference/vision_language.py

@@ -8,6 +8,7 @@ on HuggingFace model repository.
 """
 import os
 import random
+from contextlib import contextmanager
 from dataclasses import asdict
 from typing import NamedTuple, Optional
 
@@ -298,6 +299,34 @@ def run_idefics3(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+# SmolVLM2-2.2B-Instruct
+def run_smolvlm(questions: list[str], modality: str) -> ModelRequestData:
+    assert modality == "image"
+    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=2,
+        enforce_eager=True,
+        mm_processor_kwargs={
+            "max_image_size": {
+                "longest_edge": 384
+            },
+        },
+        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+    )
+    prompts = [
+        (f"<|im_start|>User:<image>{question}<end_of_utterance>\nAssistant:")
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # InternVL
 def run_internvl(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -556,7 +585,7 @@ def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
     # The configuration below has been confirmed to launch on a single L40 GPU.
     engine_args = EngineArgs(
         model=model_name,
-        max_model_len=4096,
+        max_model_len=8192,
         max_num_seqs=2,
         disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
     )
@@ -582,6 +611,42 @@ def run_mllama(questions: list[str], modality: str) -> ModelRequestData:
     )
 
 
+def run_llama4(questions: list[str], modality: str):
+    assert modality == "image"
+
+    model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=4,
+        tensor_parallel_size=8,
+        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
+        gpu_memory_utilization=0.4,
+    )
+
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    messages = [[{
+        "role":
+        "user",
+        "content": [{
+            "type": "image"
+        }, {
+            "type": "text",
+            "text": f"{question}"
+        }]
+    }] for question in questions]
+    prompts = tokenizer.apply_chat_template(messages,
+                                            add_generation_prompt=True,
+                                            tokenize=False)
+    stop_token_ids = None
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+        stop_token_ids=stop_token_ids,
+    )
+
+
 # Molmo
 def run_molmo(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -907,6 +972,7 @@ model_example_map = {
     "minicpmv": run_minicpmv,
     "mistral3": run_mistral3,
     "mllama": run_mllama,
+    "llama4": run_llama4,
     "molmo": run_molmo,
     "NVLM_D": run_nvlm_d,
     "paligemma": run_paligemma,
@@ -918,6 +984,7 @@ model_example_map = {
     "qwen2_vl": run_qwen2_vl,
     "qwen2_5_vl": run_qwen2_5_vl,
     "skywork_chat": run_skyworkr1v,
+    "smolvlm": run_smolvlm,
 }
 
 
@@ -989,6 +1056,20 @@ def apply_image_repeat(image_repeat_prob, num_prompts, data,
     return inputs
 
 
+@contextmanager
+def time_counter(enable: bool):
+    if enable:
+        import time
+        start_time = time.time()
+        yield
+        elapsed_time = time.time() - start_time
+        print("-" * 50)
+        print("-- generate time = {}".format(elapsed_time))
+        print("-" * 50)
+    else:
+        yield
+
+
 def main(args):
     model = args.model_type
     if model not in model_example_map:
@@ -1047,19 +1128,22 @@ def main(args):
                 },
             } for i in range(args.num_prompts)]
 
-    if args.time_generate:
-        import time
-        start_time = time.time()
-        outputs = llm.generate(inputs, sampling_params=sampling_params)
-        elapsed_time = time.time() - start_time
-        print("-- generate time = {}".format(elapsed_time))
+    # Add LoRA request if applicable
+    lora_request = (req_data.lora_requests *
+                    args.num_prompts if req_data.lora_requests else None)
 
-    else:
-        outputs = llm.generate(inputs, sampling_params=sampling_params)
+    with time_counter(args.time_generate):
+        outputs = llm.generate(
+            inputs,
+            sampling_params=sampling_params,
+            lora_request=lora_request,
+        )
 
+    print("-" * 50)
     for o in outputs:
         generated_text = o.outputs[0].text
         print(generated_text)
+        print("-" * 50)
 
 
 if __name__ == "__main__":

examples/offline_inference/vision_language_embedding.py

@@ -143,8 +143,10 @@ def run_encode(model: str, modality: QueryModality, seed: Optional[int]):
         "multi_modal_data": mm_data,
     })
 
+    print("-" * 50)
     for output in outputs:
         print(output.outputs.embedding)
+        print("-" * 50)
 
 
 def main(args: Namespace):

examples/offline_inference/vision_language_multi_image.py

@@ -22,6 +22,16 @@ QUESTION = "What is the content of each image?"
 IMAGE_URLS = [
     "https://upload.wikimedia.org/wikipedia/commons/d/da/2015_Kaczka_krzy%C5%BCowka_w_wodzie_%28samiec%29.jpg",
     "https://upload.wikimedia.org/wikipedia/commons/7/77/002_The_lion_king_Snyggve_in_the_Serengeti_National_Park_Photo_by_Giles_Laurent.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/2/26/Ultramarine_Flycatcher_%28Ficedula_superciliaris%29_Naggar%2C_Himachal_Pradesh%2C_2013_%28cropped%29.JPG",
+    "https://upload.wikimedia.org/wikipedia/commons/thumb/e/e5/Anim1754_-_Flickr_-_NOAA_Photo_Library_%281%29.jpg/2560px-Anim1754_-_Flickr_-_NOAA_Photo_Library_%281%29.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/d/d4/Starfish%2C_Caswell_Bay_-_geograph.org.uk_-_409413.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/6/69/Grapevinesnail_01.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/thumb/0/0b/Texas_invasive_Musk_Thistle_1.jpg/1920px-Texas_invasive_Musk_Thistle_1.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/thumb/7/7a/Huskiesatrest.jpg/2880px-Huskiesatrest.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/thumb/6/68/Orange_tabby_cat_sitting_on_fallen_leaves-Hisashi-01A.jpg/1920px-Orange_tabby_cat_sitting_on_fallen_leaves-Hisashi-01A.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/3/30/George_the_amazing_guinea_pig.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/thumb/1/1f/Oryctolagus_cuniculus_Rcdo.jpg/1920px-Oryctolagus_cuniculus_Rcdo.jpg",
+    "https://upload.wikimedia.org/wikipedia/commons/9/98/Horse-and-pony.jpg",
 ]
 
 
@@ -217,6 +227,33 @@ def load_idefics3(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_smolvlm(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
+
+    # The configuration below has been confirmed to launch on a single L40 GPU.
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=8192,
+        max_num_seqs=16,
+        enforce_eager=True,
+        limit_mm_per_prompt={"image": len(image_urls)},
+        mm_processor_kwargs={
+            "max_image_size": {
+                "longest_edge": 384
+            },
+        },
+    )
+
+    placeholders = "\n".join(f"Image-{i}: <image>\n"
+                             for i, _ in enumerate(image_urls, start=1))
+    prompt = f"<|im_start|>User:{placeholders}\n{question}<end_of_utterance>\nAssistant:"  # noqa: E501
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_internvl(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "OpenGVLab/InternVL2-2B"
 
@@ -253,6 +290,42 @@ def load_internvl(question: str, image_urls: list[str]) -> ModelRequestData:
     )
 
 
+def load_llama4(question: str, image_urls: list[str]) -> ModelRequestData:
+    model_name = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=131072,
+        tensor_parallel_size=8,
+        limit_mm_per_prompt={"image": len(image_urls)},
+    )
+
+    placeholders = [{"type": "image", "image": url} for url in image_urls]
+    messages = [{
+        "role":
+        "user",
+        "content": [
+            *placeholders,
+            {
+                "type": "text",
+                "text": question
+            },
+        ],
+    }]
+
+    processor = AutoProcessor.from_pretrained(model_name)
+
+    prompt = processor.apply_chat_template(messages,
+                                           tokenize=False,
+                                           add_generation_prompt=True)
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+        image_data=[fetch_image(url) for url in image_urls],
+    )
+
+
 def load_mistral3(question: str, image_urls: list[str]) -> ModelRequestData:
     model_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
 
@@ -281,8 +354,8 @@ def load_mllama(question: str, image_urls: list[str]) -> ModelRequestData:
     # The configuration below has been confirmed to launch on a single L40 GPU.
     engine_args = EngineArgs(
         model=model_name,
-        max_model_len=4096,
-        max_num_seqs=16,
+        max_model_len=8192,
+        max_num_seqs=2,
         limit_mm_per_prompt={"image": len(image_urls)},
     )
 
@@ -567,6 +640,7 @@ model_example_map = {
     "h2ovl_chat": load_h2ovl,
     "idefics3": load_idefics3,
     "internvl_chat": load_internvl,
+    "llama4": load_llama4,
     "mistral3": load_mistral3,
     "mllama": load_mllama,
     "NVLM_D": load_nvlm_d,
@@ -576,6 +650,7 @@ model_example_map = {
     "qwen_vl_chat": load_qwen_vl_chat,
     "qwen2_vl": load_qwen2_vl,
     "qwen2_5_vl": load_qwen2_5_vl,
+    "smolvlm": load_smolvlm,
 }
 
 
@@ -586,15 +661,8 @@ def run_generate(model, question: str, image_urls: list[str],
     engine_args = asdict(req_data.engine_args) | {"seed": args.seed}
     llm = LLM(**engine_args)
 
-    # To maintain code compatibility in this script, we add LoRA here.
-    # You can also add LoRA using:
-    # llm.generate(prompts, lora_request=lora_request,...)
-    if req_data.lora_requests:
-        for lora_request in req_data.lora_requests:
-            llm.llm_engine.add_lora(lora_request=lora_request)
-
     sampling_params = SamplingParams(temperature=0.0,
-                                     max_tokens=128,
+                                     max_tokens=256,
                                      stop_token_ids=req_data.stop_token_ids)
 
     outputs = llm.generate(
@@ -604,11 +672,15 @@ def run_generate(model, question: str, image_urls: list[str],
                 "image": req_data.image_data
             },
         },
-        sampling_params=sampling_params)
+        sampling_params=sampling_params,
+        lora_request=req_data.lora_requests,
+    )
 
+    print("-" * 50)
     for o in outputs:
         generated_text = o.outputs[0].text
         print(generated_text)
+        print("-" * 50)
 
 
 def run_chat(model: str, question: str, image_urls: list[str],
@@ -626,7 +698,7 @@ def run_chat(model: str, question: str, image_urls: list[str],
             llm.llm_engine.add_lora(lora_request=lora_request)
 
     sampling_params = SamplingParams(temperature=0.0,
-                                     max_tokens=128,
+                                     max_tokens=256,
                                      stop_token_ids=req_data.stop_token_ids)
     outputs = llm.chat(
         [{
@@ -647,11 +719,14 @@ def run_chat(model: str, question: str, image_urls: list[str],
         }],
         sampling_params=sampling_params,
         chat_template=req_data.chat_template,
+        lora_request=req_data.lora_requests,
     )
 
+    print("-" * 50)
     for o in outputs:
         generated_text = o.outputs[0].text
         print(generated_text)
+        print("-" * 50)
 
 
 def main(args: Namespace):
@@ -659,10 +734,12 @@ def main(args: Namespace):
     method = args.method
     seed = args.seed
 
+    image_urls = IMAGE_URLS[:args.num_images]
+
     if method == "generate":
-        run_generate(model, QUESTION, IMAGE_URLS, seed)
+        run_generate(model, QUESTION, image_urls, seed)
     elif method == "chat":
-        run_chat(model, QUESTION, IMAGE_URLS, seed)
+        run_chat(model, QUESTION, image_urls, seed)
     else:
         raise ValueError(f"Invalid method: {method}")
 
@@ -687,6 +764,12 @@ if __name__ == "__main__":
                         type=int,
                         default=None,
                         help="Set the seed when initializing `vllm.LLM`.")
+    parser.add_argument(
+        "--num-images",
+        "-n",
+        choices=list(range(1, 13)),  # 12 is the max number of images
+        default=2,
+        help="Number of images to use for the demo.")
 
     args = parser.parse_args()
     main(args)

examples/online_serving/api_client.py

@@ -1,5 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 """Example Python client for `vllm.entrypoints.api_server`
+Start the demo server:
+    python -m vllm.entrypoints.api_server --model <model_name>
+
 NOTE: The API server is used only for demonstration and simple performance
 benchmarks. It is not intended for production use.
 For production use, we recommend `vllm serve` and the OpenAI client API.
@@ -7,6 +10,7 @@ For production use, we recommend `vllm serve` and the OpenAI client API.
 
 import argparse
 import json
+from argparse import Namespace
 from collections.abc import Iterable
 
 import requests
@@ -27,7 +31,6 @@ def post_http_request(prompt: str,
     pload = {
         "prompt": prompt,
         "n": n,
-        "use_beam_search": True,
         "temperature": 0.0,
         "max_tokens": 16,
         "stream": stream,
@@ -55,14 +58,7 @@ def get_response(response: requests.Response) -> list[str]:
     return output
 
 
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=8000)
-    parser.add_argument("--n", type=int, default=4)
-    parser.add_argument("--prompt", type=str, default="San Francisco is a")
-    parser.add_argument("--stream", action="store_true")
-    args = parser.parse_args()
+def main(args: Namespace):
     prompt = args.prompt
     api_url = f"http://{args.host}:{args.port}/generate"
     n = args.n
@@ -83,3 +79,14 @@ if __name__ == "__main__":
         output = get_response(response)
         for i, line in enumerate(output):
             print(f"Beam candidate {i}: {line!r}", flush=True)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", type=str, default="localhost")
+    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--n", type=int, default=1)
+    parser.add_argument("--prompt", type=str, default="San Francisco is a")
+    parser.add_argument("--stream", action="store_true")
+    args = parser.parse_args()
+    main(args)

examples/template_florence2.jinja

@@ -0,0 +1,7 @@
+{%- for message in messages -%}
+    {%- if message['role'] == 'user' -%}
+        {{- message['content'] -}}
+    {%- elif message['role'] == 'assistant' -%}
+        {{- message['content'] -}}
+    {%- endif -%}
+{%- endfor -%}

examples/tool_chat_template_llama3.2_pythonic.jinja

@@ -76,7 +76,7 @@
             {{- tool_call.name + '(' -}}
             {%- for param in tool_call.arguments %}
                 {{- param + '=' -}}
-                {{- "%sr" | format(tool_call.arguments[param]) -}}
+                {{- "%s" | format(tool_call.arguments[param]) -}}
                 {% if not loop.last %}, {% endif %}
             {%- endfor %}
             {{- ')' -}}

examples/tool_chat_template_toolace.jinja

@@ -44,7 +44,7 @@
             {{- tool_call.name + '(' -}}
             {%- for param in tool_call.arguments %}
                 {{- param + '=' -}}
-                {{- "%sr" | format(tool_call.arguments[param]) -}}
+                {{- "%s" | format(tool_call.arguments[param]) -}}
                 {% if not loop.last %}, {% endif %}
             {%- endfor %}
             {{- ')' -}}

pyproject.toml

@@ -30,7 +30,7 @@ classifiers = [
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Topic :: Scientific/Engineering :: Information Analysis",
 ]
-requires-python = ">=3.9"
+requires-python = ">=3.9,<3.13"
 dynamic = [ "version", "dependencies", "optional-dependencies"]
 
 [project.urls]

requirements/common.txt

@@ -6,7 +6,7 @@ requests >= 2.26.0
 tqdm
 blake3
 py-cpuinfo
-transformers >= 4.50.3
+transformers >= 4.51.1
 huggingface-hub[hf_xet] >= 0.30.0  # Required for Xet downloads.
 tokenizers >= 0.19.1  # Required for Llama 3.
 protobuf # Required by LlamaTokenizer.
@@ -22,13 +22,13 @@ lm-format-enforcer >= 0.10.11, < 0.11
 llguidance >= 0.7.9, < 0.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64"
 outlines == 0.1.11
 lark == 1.2.2
-xgrammar == 0.1.17; platform_machine == "x86_64" or platform_machine == "aarch64"
+xgrammar == 0.1.18; platform_machine == "x86_64" or platform_machine == "aarch64"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs
 pyzmq
 msgspec
-gguf == 0.10.0
+gguf >= 0.13.0
 importlib_metadata
 mistral_common[opencv] >= 1.5.4
 opencv-python-headless >= 4.11.0    # required for video IO
@@ -36,10 +36,14 @@ pyyaml
 six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that needs to be the latest version for python 3.12
 setuptools>=74.1.1; python_version > '3.11' # Setuptools is used by triton, we need to ensure a modern version is installed for 3.12+ so that it does not try to import distutils, which was removed in 3.12
 einops # Required for Qwen2-VL.
-compressed-tensors == 0.9.2 # required for compressed-tensors
+compressed-tensors == 0.9.3 # required for compressed-tensors
 depyf==0.18.0 # required for profiling and debugging with compilation config
 cloudpickle # allows pickling lambda functions in model_executor/models/registry.py
 watchfiles # required for http server to monitor the updates of TLS files
 python-json-logger # Used by logging as per examples/other/logging_configuration.md
 scipy # Required for phi-4-multimodal-instruct
 ninja # Required for xgrammar, rocm, tpu, xpu
+opentelemetry-sdk>=1.26.0,<1.27.0  # vllm.tracing
+opentelemetry-api>=1.26.0,<1.27.0  # vllm.tracing
+opentelemetry-exporter-otlp>=1.26.0,<1.27.0  # vllm.tracing
+opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0  # vllm.tracing

requirements/cpu.txt

@@ -15,3 +15,6 @@ torchaudio==2.6.0; platform_machine == "ppc64le"
 torchvision; platform_machine != "ppc64le"  and platform_machine != "s390x"
 torchvision==0.21.0; platform_machine == "ppc64le"
 datasets # for benchmark scripts
+
+# cpu cannot use triton 3.3.0
+triton==3.2.0; platform_machine != "ppc64le"

requirements/cuda.txt

@@ -2,7 +2,7 @@
 -r common.txt
 
 numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
-numba == 0.61; python_version > '3.9'
+numba == 0.61.2; python_version > '3.9'
 
 # Dependencies for NVIDIA GPUs
 ray[cgraph]>=2.43.0, !=2.44.* # Ray Compiled Graph, required for pipeline parallelism in V1.

requirements/hpu.txt

@@ -5,6 +5,7 @@
 ray
 triton==3.1.0
 pandas
+numpy==1.26.4
 tabulate
 setuptools>=61
 setuptools-scm>=8

requirements/rocm.txt

@@ -2,7 +2,7 @@
 -r common.txt
 
 numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
-numba == 0.61; python_version > '3.9'
+numba == 0.61.2; python_version > '3.9'
 
 # Dependencies for AMD GPUs
 awscli

requirements/test.in

@@ -5,6 +5,7 @@ pytest-forked
 pytest-asyncio
 pytest-rerunfailures
 pytest-shard
+pytest-timeout
 
 # testing utils
 awscli
@@ -27,10 +28,11 @@ torchvision==0.21.0
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 mistral_common[opencv] >= 1.5.4 # required for pixtral test
+num2words # required for smolvlm test
 opencv-python-headless >= 4.11.0 # required for video test
 datamodel_code_generator # required for minicpm3 test
 lm-eval[api]==0.4.8 # required for model evaluation test
-transformers==4.50.3
+transformers==4.51.1
 huggingface-hub[hf_xet]>=0.30.0  # Required for Xet downloads.
 # quantization
 bitsandbytes>=0.45.3
@@ -40,7 +42,7 @@ genai_perf==0.0.8
 tritonclient==2.51.0
 
 numba == 0.60.0; python_version == '3.9' # v0.61 doesn't support Python 3.9. Required for N-gram speculative decoding
-numba == 0.61; python_version > '3.9'
+numba == 0.61.2; python_version > '3.9'
 numpy
 runai-model-streamer==0.11.0
 runai-model-streamer-s3==0.11.0

requirements/test.txt

@@ -101,6 +101,8 @@ dill==0.3.8
     #   multiprocess
 dnspython==2.7.0
     # via email-validator
+docopt==0.6.2
+    # via num2words
 docutils==0.16
     # via awscli
 einops==0.8.0
@@ -263,7 +265,9 @@ networkx==3.2.1
     # via torch
 nltk==3.9.1
     # via rouge-score
-numba==0.61.0
+num2words==0.5.14
+    # via -r requirements/test.in
+numba==0.61.2
     # via
     #   -r requirements/test.in
     #   librosa
@@ -444,6 +448,7 @@ pytest==8.3.3
     #   pytest-mock
     #   pytest-rerunfailures
     #   pytest-shard
+    #   pytest-timeout
 pytest-asyncio==0.24.0
     # via -r requirements/test.in
 pytest-forked==1.6.0
@@ -454,6 +459,8 @@ pytest-rerunfailures==14.0
     # via -r requirements/test.in
 pytest-shard==0.1.2
     # via -r requirements/test.in
+pytest-timeout==2.3.1
+    # via -r requirements/test.in
 python-dateutil==2.9.0.post0
     # via
     #   botocore
@@ -645,7 +652,7 @@ tqdm==4.66.6
     #   transformers
 tqdm-multiprocess==0.0.11
     # via lm-eval
-transformers==4.50.3
+transformers==4.51.1
     # via
     #   -r requirements/test.in
     #   genai-perf

requirements/tpu.txt

@@ -17,10 +17,10 @@ ray[data]
 --find-links https://storage.googleapis.com/libtpu-releases/index.html
 --find-links https://storage.googleapis.com/jax-releases/jax_nightly_releases.html
 --find-links https://storage.googleapis.com/jax-releases/jaxlib_nightly_releases.html
-torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250403-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
-torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250403-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
-torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250403-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250403-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250403-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
-torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250403-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
+torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250408-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
+torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250408-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
+torch @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch-2.8.0.dev20250408-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp39-cp39-linux_x86_64.whl ; python_version == "3.9"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp310-cp310-linux_x86_64.whl ; python_version == "3.10"
+torch_xla[tpu, pallas] @ https://storage.googleapis.com/pytorch-xla-releases/wheels/tpuvm/torch_xla-2.8.0.dev20250408-cp311-cp311-linux_x86_64.whl ; python_version == "3.11"
 

tests/compile/test_full_graph.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-from typing import Any, Union
+from typing import Any, Optional, Union
 
 import pytest
 import torch
@@ -15,7 +15,7 @@ from vllm.platforms import current_platform
 from ..utils import create_new_process_for_each_test
 
 
-def models_list(all: bool):
+def models_list(*, all: bool = True, keywords: Optional[list[str]] = None):
     TEST_MODELS: list[tuple[str, dict[str, Any]]] = [
         ("facebook/opt-125m", {}),
         ("nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change", {
@@ -32,47 +32,50 @@ def models_list(all: bool):
         ("meta-llama/Llama-3.2-1B-Instruct", {}),
     ]
 
-    if not all:
-        return TEST_MODELS
-
-    if is_quant_method_supported("aqlm"):
-        TEST_MODELS.append(("ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf", {
-            "quantization": "aqlm"
-        }))
-
-    # TODO: figure out why this fails.
-    if False and is_quant_method_supported("gguf"):  # noqa: SIM223
-        TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", {
-            "quantization": "gguf"
-        }))
-
-    if is_quant_method_supported("gptq"):
-        TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", {
-            "quantization": "gptq"
-        }))
-
-    if is_quant_method_supported("gptq_marlin"):
-        TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", {
-            "quantization": "gptq_marlin"
-        }))
-
-    if is_quant_method_supported("gptq_marlin_24"):
-        TEST_MODELS.append(("alexm-nm/tinyllama-24-marlin24-4bit-g128", {
-            "quantization": "gptq_marlin_24"
-        }))
-
-    if is_quant_method_supported("marlin"):
-        TEST_MODELS.append(
-            ("robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin", {
-                "quantization": "marlin"
+    if all:
+        if is_quant_method_supported("aqlm"):
+            TEST_MODELS.append(("ISTA-DASLab/Llama-2-7b-AQLM-2Bit-1x16-hf", {
+                "quantization": "aqlm"
             }))
 
-    if not current_platform.is_rocm() and is_quant_method_supported("awq"):
-        TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", {
-            "quantization": "AWQ"
-        }))
+        # TODO: figure out why this fails.
+        if False and is_quant_method_supported("gguf"):  # noqa: SIM223
+            TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", {
+                "quantization": "gguf"
+            }))
 
-    return TEST_MODELS
+        if is_quant_method_supported("gptq"):
+            TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-GPTQ", {
+                "quantization": "gptq"
+            }))
+
+        if is_quant_method_supported("gptq_marlin"):
+            TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", {
+                "quantization": "gptq_marlin"
+            }))
+
+        if is_quant_method_supported("gptq_marlin_24"):
+            TEST_MODELS.append(("alexm-nm/tinyllama-24-marlin24-4bit-g128", {
+                "quantization": "gptq_marlin_24"
+            }))
+
+        if is_quant_method_supported("marlin"):
+            TEST_MODELS.append(
+                ("robertgshaw2/TinyLlama-1.1B-Chat-v1.0-g128-marlin", {
+                    "quantization": "marlin"
+                }))
+
+        if not current_platform.is_rocm() and is_quant_method_supported("awq"):
+            TEST_MODELS.append(("TheBloke/TinyLlama-1.1B-Chat-v0.3-AWQ", {
+                "quantization": "AWQ"
+            }))
+
+    if keywords is None:
+        return TEST_MODELS
+
+    # filter by keywords
+    pred = lambda model: any(keyword in model[0] for keyword in keywords)
+    return list(filter(pred, TEST_MODELS))
 
 
 @pytest.mark.parametrize(
@@ -96,20 +99,30 @@ def test_full_graph(
         run_model(optimization_level, model, model_kwargs)
 
 
+PassConfig = CompilationConfig.PassConfig
+
+
 # TODO(luka) add other supported compilation config scenarios here
 @pytest.mark.parametrize(
-    "compilation_config",
-    # additional compile sizes
+    "compilation_config, model_info",
     [
-        CompilationConfig(level=CompilationLevel.PIECEWISE,
-                          compile_sizes=[1, 2])
+        # additional compile sizes, only some of the models
+        (CompilationConfig(level=CompilationLevel.PIECEWISE,
+                           compile_sizes=[1, 2]), model)
+        for model in models_list(all=False)
+    ] + [
+        # RMSNorm + quant fusion, only 8-bit quant models
+        (CompilationConfig(level=CompilationLevel.PIECEWISE,
+                           custom_ops=["+rms_norm"],
+                           pass_config=PassConfig(enable_fusion=True,
+                                                  enable_noop=True)), model)
+        for model in models_list(keywords=["FP8-dynamic", "quantized.w8a8"])
     ])
 # only test some of the models
-@pytest.mark.parametrize("model_info", models_list(all=False))
 @create_new_process_for_each_test()
 def test_custom_compile_config(
-    model_info: tuple[str, dict[str, Any]],
     compilation_config: CompilationConfig,
+    model_info: tuple[str, dict[str, Any]],
 ):
     model, model_kwargs = model_info
     print(f"MODEL={model}")

tests/compile/test_fusion.py

@@ -44,12 +44,17 @@ class TestModel(torch.nn.Module):
         resid = torch.sqrt(x)
         y = self.norm[0](x)
 
-        x2 = self.fp8_linear.apply(y, self.w[0], self.wscale[0], self.scale[0])
+        x2 = self.fp8_linear.apply(y,
+                                   self.w[0],
+                                   self.wscale[0],
+                                   input_scale=self.scale[0])
         # make sure resid is used for replacement to work
         y2, resid = self.norm[1](x2, resid)
 
-        x3 = self.fp8_linear.apply(y2, self.w[1], self.wscale[1],
-                                   self.scale[1])
+        x3 = self.fp8_linear.apply(y2,
+                                   self.w[1],
+                                   self.wscale[1],
+                                   input_scale=self.scale[1])
         y3, resid = self.norm[2](x3, resid)  # use resid here
         return y3
 

tests/conftest.py

@@ -671,8 +671,9 @@ class HfRunner:
         return [(output_ids, output_str, output_logprobs)
                 for output_ids, output_str, output_logprobs in outputs]
 
-    def encode(self, prompts: list[str]) -> list[list[torch.Tensor]]:
-        return self.model.encode(prompts)
+    def encode(self, prompts: list[str], *args,
+               **kwargs) -> list[list[torch.Tensor]]:
+        return self.model.encode(prompts, *args, **kwargs)
 
     def predict(self, prompts: list[list[str]]) -> torch.Tensor:
         return self.model.predict(prompts, convert_to_tensor=True)

tests/engine/test_arg_utils.py

@@ -1,6 +1,6 @@
 # SPDX-License-Identifier: Apache-2.0
 
-from argparse import ArgumentTypeError
+from argparse import ArgumentError, ArgumentTypeError
 
 import pytest
 
@@ -142,3 +142,39 @@ def test_composite_arg_parser(arg, expected, option):
     else:
         args = parser.parse_args([f"--{option}", arg])
     assert getattr(args, option.replace("-", "_")) == expected
+
+
+def test_human_readable_model_len():
+    # `exit_on_error` disabled to test invalid values below
+    parser = EngineArgs.add_cli_args(
+        FlexibleArgumentParser(exit_on_error=False))
+
+    args = parser.parse_args([])
+    assert args.max_model_len is None
+
+    args = parser.parse_args(["--max-model-len", "1024"])
+    assert args.max_model_len == 1024
+
+    # Lower
+    args = parser.parse_args(["--max-model-len", "1m"])
+    assert args.max_model_len == 1_000_000
+    args = parser.parse_args(["--max-model-len", "10k"])
+    assert args.max_model_len == 10_000
+
+    # Capital
+    args = parser.parse_args(["--max-model-len", "3K"])
+    assert args.max_model_len == 1024 * 3
+    args = parser.parse_args(["--max-model-len", "10M"])
+    assert args.max_model_len == 2**20 * 10
+
+    # Decimal values
+    args = parser.parse_args(["--max-model-len", "10.2k"])
+    assert args.max_model_len == 10200
+    # ..truncated to the nearest int
+    args = parser.parse_args(["--max-model-len", "10.212345k"])
+    assert args.max_model_len == 10212
+
+    # Invalid (do not allow decimals with binary multipliers)
+    for invalid in ["1a", "pwd", "10.24", "1.23M"]:
+        with pytest.raises(ArgumentError):
+            args = parser.parse_args(["--max-model-len", invalid])

tests/engine/test_short_mm_context.py

@@ -18,7 +18,8 @@ models = ["llava-hf/llava-1.5-7b-hf"]
 def test_context_length_too_short(vllm_runner, image_assets, model):
     images = [asset.pil_image for asset in image_assets]
 
-    with pytest.raises(ValueError, match="too long to fit into the model"):
+    with pytest.raises(ValueError,
+                       match="longer than the maximum model length"):
         vllm_model = vllm_runner(
             model,
             max_model_len=128,  # LLaVA has a feature size of 576

tests/entrypoints/llm/test_guided_generate.py

@@ -3,9 +3,11 @@
 import json
 import re
 import weakref
+from enum import Enum
 
 import jsonschema
 import pytest
+from pydantic import BaseModel
 
 from vllm.distributed import cleanup_dist_env_and_memory
 from vllm.entrypoints.llm import LLM
@@ -330,3 +332,44 @@ def test_guided_json_object(llm, guided_decoding_backend: str):
             # Parse to verify it is valid JSON
             parsed_json = json.loads(generated_text)
             assert isinstance(parsed_json, dict)
+
+
+class CarType(str, Enum):
+    sedan = "sedan"
+    suv = "SUV"
+    truck = "Truck"
+    coupe = "Coupe"
+
+
+class CarDescription(BaseModel):
+    brand: str
+    model: str
+    car_type: CarType
+
+
+@pytest.mark.skip_global_cleanup
+@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
+def test_guided_json_completion_with_enum(llm, guided_decoding_backend: str):
+    json_schema = CarDescription.model_json_schema()
+    sampling_params = SamplingParams(temperature=1.0,
+                                     max_tokens=1000,
+                                     guided_decoding=GuidedDecodingParams(
+                                         json=json_schema,
+                                         backend=guided_decoding_backend))
+    outputs = llm.generate(
+        prompts="Generate a JSON with the brand, model and car_type of"
+        "the most iconic car from the 90's",
+        sampling_params=sampling_params,
+        use_tqdm=True)
+
+    assert outputs is not None
+    for output in outputs:
+        assert output is not None
+        assert isinstance(output, RequestOutput)
+        prompt = output.prompt
+
+        generated_text = output.outputs[0].text
+        assert generated_text is not None
+        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+        output_json = json.loads(generated_text)
+        jsonschema.validate(instance=output_json, schema=json_schema)

tests/entrypoints/llm/test_prompt_validation.py

@@ -15,7 +15,7 @@ def v1(run_with_both_engines):
 
 def test_empty_prompt():
     llm = LLM(model="openai-community/gpt2", enforce_eager=True)
-    with pytest.raises(ValueError, match='Prompt cannot be empty'):
+    with pytest.raises(ValueError, match='decoder prompt cannot be empty'):
         llm.generate([""])
 
 

tests/entrypoints/openai/test_chat.py

@@ -20,8 +20,6 @@ from .test_completion import zephyr_lora_files  # noqa: F401
 # any model with a chat template should work here
 MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
 
-GUIDED_DECODING_BACKENDS = ["outlines", "lm-format-enforcer", "xgrammar"]
-
 
 @pytest.fixture(scope="module")
 def monkeypatch_module():
@@ -487,20 +485,9 @@ async def test_chat_completion_stream_options(client: openai.AsyncOpenAI,
     assert last_completion_tokens == 10
 
 
-# NOTE: Not sure why, but when I place this after `test_guided_regex_chat`
-# (i.e. using the same ordering as in the Completions API tests), the test
-# will fail on the second `guided_decoding_backend` even when I swap their order
-# (ref: https://github.com/vllm-project/vllm/pull/5526#issuecomment-2173772256)
 @pytest.mark.asyncio
-@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
 async def test_guided_choice_chat(client: openai.AsyncOpenAI,
-                                  is_v1_server: bool,
-                                  guided_decoding_backend: str,
                                   sample_guided_choice):
-
-    if is_v1_server and guided_decoding_backend != 'xgrammar':
-        pytest.skip("Only xgrammar backend is supported with V1")
-
     messages = [{
         "role": "system",
         "content": "you are a helpful assistant"
@@ -515,8 +502,7 @@ async def test_guided_choice_chat(client: openai.AsyncOpenAI,
         messages=messages,
         max_completion_tokens=10,
         temperature=0.7,
-        extra_body=dict(guided_choice=sample_guided_choice,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_choice=sample_guided_choice))
     choice1 = chat_completion.choices[0].message.content
     assert choice1 in sample_guided_choice
 
@@ -530,22 +516,16 @@ async def test_guided_choice_chat(client: openai.AsyncOpenAI,
         messages=messages,
         max_completion_tokens=10,
         temperature=0.7,
-        extra_body=dict(guided_choice=sample_guided_choice,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_choice=sample_guided_choice))
     choice2 = chat_completion.choices[0].message.content
     assert choice2 in sample_guided_choice
     assert choice1 != choice2
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
-async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
-                                guided_decoding_backend: str,
+async def test_guided_json_chat(client: openai.AsyncOpenAI,
                                 sample_json_schema):
 
-    if is_v1_server:
-        pytest.skip("sample_json_schema has features unsupported in V1")
-
     messages = [{
         "role": "system",
         "content": "you are a helpful assistant"
@@ -560,8 +540,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
         model=MODEL_NAME,
         messages=messages,
         max_completion_tokens=1000,
-        extra_body=dict(guided_json=sample_json_schema,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_json=sample_json_schema))
     message = chat_completion.choices[0].message
     assert message.content is not None
     json1 = json.loads(message.content)
@@ -578,8 +557,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
         model=MODEL_NAME,
         messages=messages,
         max_completion_tokens=1000,
-        extra_body=dict(guided_json=sample_json_schema,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_json=sample_json_schema))
     message = chat_completion.choices[0].message
     assert message.content is not None
     json2 = json.loads(message.content)
@@ -589,13 +567,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
-async def test_guided_regex_chat(client: openai.AsyncOpenAI,
-                                 is_v1_server: bool,
-                                 guided_decoding_backend: str, sample_regex):
-
-    if is_v1_server and guided_decoding_backend != 'xgrammar':
-        pytest.skip("Only xgrammar backend is supported with V1")
+async def test_guided_regex_chat(client: openai.AsyncOpenAI, sample_regex):
 
     messages = [{
         "role": "system",
@@ -610,8 +582,7 @@ async def test_guided_regex_chat(client: openai.AsyncOpenAI,
         model=MODEL_NAME,
         messages=messages,
         max_completion_tokens=20,
-        extra_body=dict(guided_regex=sample_regex,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_regex=sample_regex))
     ip1 = chat_completion.choices[0].message.content
     assert ip1 is not None
     assert re.fullmatch(sample_regex, ip1) is not None
@@ -622,8 +593,7 @@ async def test_guided_regex_chat(client: openai.AsyncOpenAI,
         model=MODEL_NAME,
         messages=messages,
         max_completion_tokens=20,
-        extra_body=dict(guided_regex=sample_regex,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_regex=sample_regex))
     ip2 = chat_completion.choices[0].message.content
     assert ip2 is not None
     assert re.fullmatch(sample_regex, ip2) is not None
@@ -652,15 +622,9 @@ async def test_guided_decoding_type_error(client: openai.AsyncOpenAI):
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
 async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
-                                           is_v1_server: bool,
-                                           guided_decoding_backend: str,
                                            sample_guided_choice):
 
-    if is_v1_server and guided_decoding_backend != 'xgrammar':
-        pytest.skip("Only xgrammar backend is supported with V1")
-
     messages = [{
         "role": "system",
         "content": "you are a helpful assistant"
@@ -676,8 +640,7 @@ async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
         max_completion_tokens=10,
         logprobs=True,
         top_logprobs=5,
-        extra_body=dict(guided_choice=sample_guided_choice,
-                        guided_decoding_backend=guided_decoding_backend))
+        extra_body=dict(guided_choice=sample_guided_choice))
 
     assert chat_completion.choices[0].logprobs is not None
     assert chat_completion.choices[0].logprobs.content is not None
@@ -689,14 +652,7 @@ async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
 
 
 @pytest.mark.asyncio
-@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
-async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
-                              guided_decoding_backend: str,
-                              sample_json_schema):
-
-    if is_v1_server:
-        pytest.skip("sample_json_schema has features unsupported on V1")
-
+async def test_named_tool_use(client: openai.AsyncOpenAI, sample_json_schema):
     messages = [{
         "role": "system",
         "content": "you are a helpful assistant"
@@ -728,7 +684,7 @@ async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
                 "name": "dummy_function_name"
             }
         },
-        extra_body=dict(guided_decoding_backend=guided_decoding_backend))
+    )
     message = chat_completion.choices[0].message
     assert len(message.content) == 0
     json_string = message.tool_calls[0].function.arguments
@@ -763,7 +719,6 @@ async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
                 "name": "dummy_function_name"
             }
         },
-        extra_body=dict(guided_decoding_backend=guided_decoding_backend),
         stream=True)
 
     output = []
@@ -888,7 +843,6 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
         model=model_name,
         tools=tools,
         tool_choice="required",
-        extra_body=dict(guided_decoding_backend="outlines"),
     )
 
     assert chat_completion.choices[0].message.tool_calls is not None
@@ -900,7 +854,6 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
         model=model_name,
         tools=tools,
         tool_choice="required",
-        extra_body=dict(guided_decoding_backend="outlines"),
         stream=True,
     )
 
@@ -914,12 +867,7 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
 
 @pytest.mark.asyncio
 async def test_inconsistent_tool_choice_and_tools(client: openai.AsyncOpenAI,
-                                                  is_v1_server: bool,
                                                   sample_json_schema):
-
-    if is_v1_server:
-        pytest.skip("sample_json_schema has features unsupported on V1")
-
     messages = [{
         "role": "system",
         "content": "you are a helpful assistant"

tests/entrypoints/openai/test_embedding.py

@@ -11,6 +11,7 @@ import requests
 from vllm.entrypoints.openai.protocol import EmbeddingResponse
 from vllm.transformers_utils.tokenizer import get_tokenizer
 
+from ...models.embedding.utils import check_embeddings_close
 from ...utils import RemoteOpenAIServer
 
 MODEL_NAME = "intfloat/multilingual-e5-small"
@@ -190,30 +191,35 @@ async def test_batch_base64_embedding(client: openai.AsyncOpenAI,
     responses_float = await client.embeddings.create(input=input_texts,
                                                      model=model_name,
                                                      encoding_format="float")
+    float_data = [d.embedding for d in responses_float.data]
 
     responses_base64 = await client.embeddings.create(input=input_texts,
                                                       model=model_name,
                                                       encoding_format="base64")
-
-    decoded_responses_base64_data = []
+    base64_data = []
     for data in responses_base64.data:
-        decoded_responses_base64_data.append(
+        base64_data.append(
             np.frombuffer(base64.b64decode(data.embedding),
                           dtype="float32").tolist())
 
-    assert responses_float.data[0].embedding == decoded_responses_base64_data[
-        0]
-    assert responses_float.data[1].embedding == decoded_responses_base64_data[
-        1]
+    check_embeddings_close(
+        embeddings_0_lst=float_data,
+        embeddings_1_lst=base64_data,
+        name_0="float",
+        name_1="base64",
+    )
 
     # Default response is float32 decoded from base64 by OpenAI Client
     responses_default = await client.embeddings.create(input=input_texts,
                                                        model=model_name)
+    default_data = [d.embedding for d in responses_default.data]
 
-    assert responses_float.data[0].embedding == responses_default.data[
-        0].embedding
-    assert responses_float.data[1].embedding == responses_default.data[
-        1].embedding
+    check_embeddings_close(
+        embeddings_0_lst=float_data,
+        embeddings_1_lst=default_data,
+        name_0="float",
+        name_1="default",
+    )
 
 
 @pytest.mark.asyncio

tests/entrypoints/openai/test_prompt_validation.py

@@ -17,7 +17,7 @@ async def test_empty_prompt():
         client = remote_server.get_async_client()
 
         with pytest.raises(openai.BadRequestError,
-                           match=re.compile('.+Prompt cannot be empty.+')):
+                           match="decoder prompt cannot be empty"):
             await client.completions.create(model=model_name,
                                             prompt="",
                                             max_tokens=5,

tests/entrypoints/test_chat_utils.py

@@ -30,6 +30,7 @@ QWEN25VL_MODEL_ID = "Qwen/Qwen2.5-VL-3B-Instruct"
 MLLAMA_MODEL_ID = "meta-llama/Llama-3.2-11B-Vision-Instruct"
 LLAMA_GUARD_MODEL_ID = "meta-llama/Llama-Guard-3-1B"
 HERMES_MODEL_ID = "NousResearch/Hermes-3-Llama-3.1-8B"
+MISTRAL_MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
 
 
 @pytest.fixture(scope="function")
@@ -80,6 +81,30 @@ def mllama_tokenizer():
     )
 
 
+@pytest.fixture(scope="function")
+def mistral_model_config():
+    return ModelConfig(MISTRAL_MODEL_ID,
+                       task="generate",
+                       tokenizer=MISTRAL_MODEL_ID,
+                       tokenizer_mode="auto",
+                       trust_remote_code=True,
+                       dtype="auto",
+                       seed=0,
+                       limit_mm_per_prompt={
+                           "image": 2,
+                       })
+
+
+@pytest.fixture(scope="module")
+def mistral_tokenizer():
+    return TokenizerGroup(
+        tokenizer_id=MISTRAL_MODEL_ID,
+        enable_lora=False,
+        max_num_seqs=5,
+        max_input_length=None,
+    )
+
+
 @pytest.fixture(scope="module")
 def image_url():
     image = ImageAsset('cherry_blossom')
@@ -131,6 +156,66 @@ def test_parse_chat_messages_single_image(
     _assert_mm_data_is_image_input(mm_data, 1)
 
 
+def test_parse_chat_messages_empty_system(
+    mistral_model_config,
+    mistral_tokenizer,
+):
+    # Test string format
+    conversation, _ = parse_chat_messages(
+        [{
+            "role": "system",
+            "content": ""
+        }, {
+            "role": "user",
+            "content": [{
+                "type": "text",
+                "text": "Who are you?"
+            }]
+        }],
+        mistral_model_config,
+        mistral_tokenizer,
+        content_format="string",
+    )
+    assert conversation == [{
+        "role": "system",
+        "content": ""
+    }, {
+        "role": "user",
+        "content": "Who are you?"
+    }]
+
+    # Test openai format
+    conversation, _ = parse_chat_messages(
+        [{
+            "role": "system",
+            "content": ""
+        }, {
+            "role": "user",
+            "content": [{
+                "type": "text",
+                "text": "Who are you?"
+            }]
+        }],
+        mistral_model_config,
+        mistral_tokenizer,
+        content_format="openai",
+    )
+    assert conversation == [{
+        "role": "system",
+        "content": [{
+            "type": "text",
+            "text": ""
+        }]
+    }, {
+        "role":
+        "user",
+        "content": [{
+            "type": "text",
+            "text": "Who are you?"
+        }]
+    }]
+
+
 @pytest.mark.asyncio
 async def test_parse_chat_messages_single_image_async(
     phi3v_model_config,
@@ -671,7 +756,7 @@ def test_multimodal_image_parsing_matches_hf(model, image_url):
     # Build a config for the model
     model_config = ModelConfig(model,
                                task="generate",
-                               tokenizer=MLLAMA_MODEL_ID,
+                               tokenizer=model,
                                tokenizer_mode="auto",
                                trust_remote_code=True,
                                dtype="auto",
@@ -682,7 +767,7 @@ def test_multimodal_image_parsing_matches_hf(model, image_url):
 
     # Build the tokenizer group and grab the underlying tokenizer
     tokenizer_group = TokenizerGroup(
-        MLLAMA_MODEL_ID,
+        model,
         enable_lora=False,
         max_num_seqs=5,
         max_input_length=None,

tests/kernels/test_flashmla.py

@@ -124,7 +124,7 @@ def test_flash_mla(b, s_q, mean_sk, h_q, h_kv, d, dv, block_size, causal,
     cal_diff(out_flash, out_torch, "out")
     cal_diff(lse_flash, lse_torch, "lse")
 
-    t = triton.testing.do_bench(flash_mla, fast_flush=False)
+    t = triton.testing.do_bench(flash_mla)
     FLOPS = s_q * total_seqlens * h_q * (d + dv) * 2
     bytes = (total_seqlens * h_kv * d + b * s_q * h_q * d +
              b * s_q * h_q * dv) * (torch.finfo(dtype).bits // 8)

tests/kernels/test_merge_attn_states.py

@@ -0,0 +1,265 @@
+# SPDX-License-Identifier: Apache-2.0
+from typing import Optional
+
+import pytest
+import torch
+
+from vllm._custom_ops import merge_attn_states as merge_attn_states_cuda
+from vllm.attention.ops.triton_merge_attn_states import (
+    merge_attn_states as merge_attn_states_triton)
+from vllm.platforms import current_platform
+
+
+# Naive PyTorch Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+# can be used to combine partial attention results (in the split-KV case)
+def merge_attn_states_torch(
+        output: torch.Tensor,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+        prefix_output: torch.Tensor,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+        prefix_lse: torch.Tensor,  # [NUM_HEADS, NUM_TOKENS]
+        suffix_output: torch.Tensor,  # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]
+        suffix_lse: torch.Tensor,  # [NUM_HEADS, NUM_TOKENS]
+        output_lse: Optional[torch.Tensor] = None,  # [NUM_HEADS, NUM_TOKENS]
+):
+    p_lse = prefix_lse
+    s_lse = suffix_lse
+    # inf -> -inf
+    p_lse[p_lse == torch.inf] = -torch.inf
+    s_lse[s_lse == torch.inf] = -torch.inf
+    # max_lse [NUM_HEADS, NUM_TOKENS]
+    max_lse = torch.maximum(p_lse, s_lse)
+    p_lse = p_lse - max_lse
+    s_lse = s_lse - max_lse
+    p_lse_exp = torch.exp(p_lse)
+    s_lse_exp = torch.exp(s_lse)
+    out_se = (p_lse_exp + s_lse_exp)
+    if output_lse is not None:
+        output_lse = torch.log(out_se) + max_lse
+    p_scale = p_lse_exp / out_se  # [NUM_HEADS, NUM_TOKENS]
+    s_scale = s_lse_exp / out_se  # [NUM_HEADS, NUM_TOKENS]
+    p_scale = torch.transpose(p_scale, 0,
+                              1).unsqueeze(2)  # [NUM_TOKENS, NUM_HEADS, 1]
+    s_scale = torch.transpose(s_scale, 0,
+                              1).unsqueeze(2)  # [NUM_TOKENS, NUM_HEADS, 1]
+    output = prefix_output * p_scale + suffix_output * s_scale
+    return output, output_lse
+
+
+NUM_BATCH_TOKENS = [256, 512, 613, 1024, 1536, 4096]
+NUM_QUERY_HEADS = [4, 8, 16, 32, 48, 64]
+HEAD_SIZES = [32, 48, 64, 96, 128, 256]
+DTYPES = [torch.float32, torch.half, torch.bfloat16]
+
+all_case_info: list[tuple] = []
+
+
+def generate_markdown_table():
+    global all_case_info
+    table_header = ("| tokens | heads | headsize | dtype "
+                    "| device | torch | triton | cuda | speedup |")
+    table_separator = "| --- | --- | --- | --- | --- | --- | --- | --- | --- |"
+
+    def shortly_dtype(dtype: torch.dtype) -> str:
+        return str(dtype).removeprefix("torch.")
+
+    def shortly_device(device: str) -> str:
+        return device.removeprefix("NVIDIA").strip()
+
+    print(table_header)
+    print(table_separator)
+    for info in all_case_info:
+        (num_tokens, num_heads, head_size, dtype, device,
+         avg_time_torch_kernel, avg_time_triton_kernel, avg_time_cuda_kernel,
+         performance_improved) = info
+        dtype = shortly_dtype(dtype)
+        device = shortly_device(device)
+        print(f"| {num_tokens} | {num_heads} | {head_size} "
+              f"| {dtype} | {device} | {avg_time_torch_kernel:.5f}ms "
+              f"| {avg_time_triton_kernel:.5f}ms "
+              f"| {avg_time_cuda_kernel:.5f}ms "
+              f"| {performance_improved:.4f}x |")
+
+
+@pytest.mark.parametrize("num_tokens", NUM_BATCH_TOKENS)
+@pytest.mark.parametrize("num_query_heads", NUM_QUERY_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("output_dtype", DTYPES)
+@torch.inference_mode()
+def test_merge_attn_states(num_tokens: int, num_query_heads: int,
+                           head_size: int, output_dtype: torch.dtype):
+    if not current_platform.is_cuda():
+        pytest.skip('Currently only support compare triton merge_attn_states '
+                    'with custom cuda merge_attn_states kernel')
+
+    NUM_TOKENS = num_tokens
+    NUM_HEADS = num_query_heads
+    HEAD_SIZE = head_size
+
+    print(f"\nNUM_TOKENS:{NUM_TOKENS}, NUM_HEADS:{NUM_HEADS}, "
+          f"HEAD_SIZE:{HEAD_SIZE}, DTYPE: {output_dtype}, "
+          f"Device: {current_platform.get_device_name()}")
+
+    # prefix_lse and suffix_lse contain inf and normal values
+    prefix_lse = torch.randn(NUM_HEADS,
+                             NUM_TOKENS,
+                             dtype=torch.float32,
+                             device="cuda")
+    suffix_lse = torch.randn(NUM_HEADS,
+                             NUM_TOKENS,
+                             dtype=torch.float32,
+                             device="cuda")
+
+    # Generate boolean masks
+    mask_prefix = torch.rand(NUM_HEADS, NUM_TOKENS) < 0.1
+    mask_suffix = torch.rand(NUM_HEADS, NUM_TOKENS) < 0.1
+    # Ensure that the same position is not True at the same time
+    combined_mask = torch.logical_and(mask_prefix, mask_suffix)
+    mask_prefix = torch.logical_and(mask_prefix, ~combined_mask)
+    mask_suffix = torch.logical_and(mask_suffix, ~combined_mask)
+
+    prefix_lse[mask_prefix] = float('inf')
+    suffix_lse[mask_suffix] = float('inf')
+
+    # Other input tensors (need to be initialized but
+    # no actual calculation needed)
+    output = torch.zeros((NUM_TOKENS, NUM_HEADS, HEAD_SIZE),
+                         dtype=output_dtype,
+                         device="cuda")
+    output_lse = torch.zeros((NUM_HEADS, NUM_TOKENS),
+                             dtype=torch.float32,
+                             device="cuda")
+    prefix_output = torch.randn((NUM_TOKENS, NUM_HEADS, HEAD_SIZE),
+                                dtype=output_dtype,
+                                device="cuda")
+    suffix_output = torch.randn((NUM_TOKENS, NUM_HEADS, HEAD_SIZE),
+                                dtype=output_dtype,
+                                device="cuda")
+
+    warmup_times = 2
+    repeat_times = 20
+
+    output_torch = output.clone()
+    output_lse_torch = output_lse.clone()
+    total_time_torch_kernel = 0
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+
+    # 0. Run the Torch kernel
+    prefix_lse_torch = prefix_lse.clone()
+    suffix_lse_torch = suffix_lse.clone()
+    for _ in range(warmup_times):
+        output_torch, output_lse_torch = merge_attn_states_torch(
+            output_torch, prefix_output, prefix_lse_torch, suffix_output,
+            suffix_lse_torch, output_lse_torch)
+    torch.cuda.synchronize()
+
+    for _ in range(repeat_times):
+        start.record()
+        output_torch, output_lse_torch = merge_attn_states_torch(
+            output_torch, prefix_output, prefix_lse_torch, suffix_output,
+            suffix_lse_torch, output_lse_torch)
+        end.record()
+        torch.cuda.synchronize()
+        total_time_torch_kernel += start.elapsed_time(end)
+
+    avg_time_torch_kernel = total_time_torch_kernel / repeat_times
+
+    # 1. Run the Triton kernel
+    output_ref_triton = output.clone()
+    output_lse_ref_triton = output_lse.clone()
+
+    total_time_triton_kernel = 0
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+
+    for _ in range(warmup_times):
+        merge_attn_states_triton(output_ref_triton, prefix_output, prefix_lse,
+                                 suffix_output, suffix_lse,
+                                 output_lse_ref_triton)
+    torch.cuda.synchronize()
+
+    for _ in range(repeat_times):
+        start.record()
+        merge_attn_states_triton(output_ref_triton, prefix_output, prefix_lse,
+                                 suffix_output, suffix_lse,
+                                 output_lse_ref_triton)
+        end.record()
+        torch.cuda.synchronize()
+        total_time_triton_kernel += start.elapsed_time(end)
+
+    avg_time_triton_kernel = total_time_triton_kernel / repeat_times
+
+    # 2. Run the CUDA kernel
+    total_time_cuda_kernel = 0
+    output_cuda = output.clone()
+    output_lse_cuda = output_lse.clone()
+
+    for _ in range(warmup_times):
+        merge_attn_states_cuda(output_cuda, prefix_output, prefix_lse,
+                               suffix_output, suffix_lse, output_lse_cuda)
+    torch.cuda.synchronize()
+
+    for _ in range(repeat_times):
+        start.record()
+        merge_attn_states_cuda(output_cuda, prefix_output, prefix_lse,
+                               suffix_output, suffix_lse, output_lse_cuda)
+        end.record()
+        torch.cuda.synchronize()
+        total_time_cuda_kernel += start.elapsed_time(end)
+
+    avg_time_cuda_kernel = total_time_cuda_kernel / repeat_times
+
+    # 3. Performance compare
+    performance_improved = avg_time_triton_kernel / avg_time_cuda_kernel
+    print(f" Torch time: {avg_time_torch_kernel:.6f}ms")
+    print(f"Triton time: {avg_time_triton_kernel:.6f}ms")
+    print(f"  CUDA time: {avg_time_cuda_kernel:.6f}ms, "
+          f"Performance: {performance_improved:.5f}x")
+    print("-" * 100)
+
+    # 4. Correctness compare
+    # Liger Kernel: Efficient Triton Kernels for LLM Training
+    # https://arxiv.org/pdf/2410.10989, 3.3 Correctness
+    # use rtol = 1e-2 for bfloat16.
+    rtol = 1e-2 if output_dtype == torch.bfloat16 else 1e-3
+
+    def diff(a: torch.Tensor, b: torch.Tensor):
+        max_diff = torch.max(torch.abs(a.float() - b.float()))
+        return max_diff
+
+    # Use Triton output as reference because we want to replace
+    # the Triton kernel with custom CUDA kernel for merge attn
+    # states operation.
+    output_ref = output_ref_triton
+    output_lse_ref = output_lse_ref_triton
+    torch.testing.assert_close(output_cuda.float(),
+                               output_ref.float(),
+                               atol=1e-3,
+                               rtol=rtol)
+    print("Output all match, max abs diff:")
+    print(f"(Triton vs Torch) : {diff(output_torch, output_ref)}")
+    print(f"  (CUDA vs Torch) : {diff(output_torch, output_cuda)}")
+    print(f"  (CUDA vs Triton): {diff(output_ref, output_cuda)}")
+    print("-" * 100)
+
+    torch.testing.assert_close(output_lse_cuda.float(),
+                               output_lse_ref.float(),
+                               atol=1e-3,
+                               rtol=rtol)
+    print("Output LSE all match, max abs diff:")
+    print(f"(Triton vs Torch) : {diff(output_lse_torch, output_lse_ref)}")
+    print(f"  (CUDA vs Torch) : {diff(output_lse_torch, output_lse_cuda)}")
+    print(f"  (CUDA vs Triton): {diff(output_lse_ref, output_lse_cuda)}")
+    print("-" * 100)
+
+    print("All output values test passed! All inf values "
+          "are correctly replaced with -inf.")
+    print("-" * 100)
+
+    device = current_platform.get_device_name()
+    all_case_info.append(
+        (NUM_TOKENS, NUM_HEADS, HEAD_SIZE, output_dtype, device,
+         avg_time_torch_kernel, avg_time_triton_kernel, avg_time_cuda_kernel,
+         performance_improved))
+    if len(all_case_info) == (len(NUM_BATCH_TOKENS) * len(HEAD_SIZES) *
+                              len(NUM_QUERY_HEADS) * len(DTYPES)):
+        generate_markdown_table()

tests/lora/conftest.py

@@ -256,3 +256,15 @@ def run_with_both_engines_lora(request, monkeypatch):
         monkeypatch.setenv('VLLM_USE_V1', '0')
 
     yield
+
+
+@pytest.fixture
+def reset_default_device():
+    """
+    Some tests, such as `test_punica_ops.py`, explicitly set the 
+    default device, which can affect subsequent tests. Adding this fixture 
+    helps avoid this problem.
+    """
+    original_device = torch.get_default_device()
+    yield
+    torch.set_default_device(original_device)

tests/lora/test_baichuan.py

@@ -73,7 +73,6 @@ def test_baichuan_tensor_parallel_equality(baichuan_lora_files,
                        max_num_seqs=16,
                        max_loras=4,
                        max_lora_rank=64,
-                       tensor_parallel_size=1,
                        trust_remote_code=True,
                        fully_sharded_loras=fully_sharded)
     output_tp1 = do_sample(llm_tp1, baichuan_lora_files, lora_id=1)

tests/lora/test_chatglm3_tp.py

@@ -61,7 +61,6 @@ def test_chatglm3_lora(chatglm3_lora_files):
                    enable_lora=True,
                    max_loras=4,
                    max_lora_rank=64,
-                   tensor_parallel_size=1,
                    trust_remote_code=True,
                    enable_chunked_prefill=True)
 

tests/lora/test_layers.py

@@ -65,7 +65,7 @@ VOCAB_PARALLEL_EMBEDDING_TEST_NUM_RANDOM_SEEDS = 128
 
 
 @pytest.fixture(autouse=True)
-def clean_cache():
+def clean_cache_reset_device(reset_default_device):
     # Release any memory we might be holding on to. CI runs OOMs otherwise.
     from vllm.lora.ops.triton_ops.utils import (_LORA_A_PTR_DICT,
                                                 _LORA_B_PTR_DICT)

tests/lora/test_llama_tp.py

@@ -88,7 +88,6 @@ def test_llama_lora(sql_lora_files):
         # also test odd max_num_seqs
         max_num_seqs=13,
         max_loras=4,
-        tensor_parallel_size=1,
         enable_chunked_prefill=True)
     generate_and_test(llm, sql_lora_files)
 

tests/lora/test_punica_ops.py

@@ -13,6 +13,11 @@ from vllm.platforms import current_platform
 from .utils import PunicaTensors, assert_close, generate_data_for_nslices
 
 
+@pytest.fixture(autouse=True)
+def reset_device(reset_default_device):
+    pass
+
+
 # Utility shrink and expand operations used as reference implementations.
 def sgmv_shrink_for_nslices(
         nslices: int, inputs_tensor: torch.Tensor,

tests/lora/test_quant_model.py

@@ -78,12 +78,7 @@ def do_sample(llm: vllm.LLM,
 
 
 @pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("tp_size", [1])
-def test_quant_model_lora(tinyllama_lora_files, num_gpus_available, model,
-                          tp_size):
-    if num_gpus_available < tp_size and \
-        tp_size > 1 and current_platform.is_cuda_alike():
-        pytest.skip(f"Not enough GPUs for tensor parallelism {tp_size}")
+def test_quant_model_lora(tinyllama_lora_files, model):
 
     llm = vllm.LLM(
         model=model.model_path,
@@ -91,7 +86,6 @@ def test_quant_model_lora(tinyllama_lora_files, num_gpus_available, model,
         max_num_seqs=16,
         max_loras=4,
         max_model_len=400,
-        tensor_parallel_size=tp_size,
         gpu_memory_utilization=0.2,  #avoid OOM
         quantization=model.quantization,
         trust_remote_code=True,
@@ -185,7 +179,6 @@ def test_quant_model_tp_equality(tinyllama_lora_files, num_gpus_available,
         enable_lora=True,
         max_num_seqs=16,
         max_loras=4,
-        tensor_parallel_size=1,
         gpu_memory_utilization=0.2,  #avoid OOM
         quantization=model.quantization,
         trust_remote_code=True,

tests/lora/test_transfomers_model.py

@@ -53,7 +53,6 @@ def test_ilama_lora(ilama_lora_files):
                    enable_lora=True,
                    max_loras=4,
                    max_lora_rank=16,
-                   tensor_parallel_size=1,
                    trust_remote_code=True,
                    enable_chunked_prefill=True)