Merge remote-tracking branch 'upstream/main' into HEAD

.buildkite/run-amd-test.sh

@@ -107,11 +107,12 @@ fi
 PARALLEL_JOB_COUNT=8
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
+  # assign job count as the number of shards used   
+  commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
   for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
-    #replace shard arguments
-    commands=${commands//"--shard-id= "/"--shard-id=${GPU} "}
-    commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
-    echo "Shard ${GPU} commands:$commands"
+    # assign shard-id for each shard
+    commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
+    echo "Shard ${GPU} commands:$commands_gpu"
     docker run \
         --device /dev/kfd --device /dev/dri \
         --network host \
@@ -123,7 +124,7 @@ if [[ $commands == *"--shard-id="* ]]; then
         -e HF_HOME=${HF_MOUNT} \
         --name ${container_name}_${GPU}  \
         ${image_name} \
-        /bin/bash -c "${commands}" \
+        /bin/bash -c "${commands_gpu}" \
         |& while read -r line; do echo ">>Shard $GPU: $line"; done &
     PIDS+=($!)
   done

.buildkite/run-tpu-test.sh

@@ -12,4 +12,4 @@ remove_docker_container
 # For HF_TOKEN.
 source /etc/environment
 # Run a simple end-to-end example.
-docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 -m pip install pytest  && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"
+docker run --privileged --net host --shm-size=16G -it -e HF_TOKEN=$HF_TOKEN --name tpu-test vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git && python3 -m pip install pytest && python3 -m pip install lm_eval[api]==0.4.4 && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py && python3 /workspace/vllm/tests/tpu/test_compilation.py && python3 /workspace/vllm/examples/offline_inference_tpu.py"

CMakeLists.txt

@@ -49,8 +49,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx11
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.5.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.5.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.5.1")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.5.1")
 
 #
 # Try to find python package with an executable that exactly matches

benchmarks/benchmark_throughput.py

@@ -4,7 +4,7 @@
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import List, Optional
 
 import torch
 import uvloop
@@ -15,16 +15,35 @@
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
     build_async_engine_client_from_engine_args)
+from vllm.inputs import TextPrompt
+from vllm.multimodal import MultiModalDataDict
 from vllm.sampling_params import BeamSearchParams
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
 
 
+@dataclasses.dataclass
+class SampleRequest:
+    """A class representing a single inference request for benchmarking.
+
+    Attributes:
+        prompt: The input text prompt for the model.
+        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
+            images).
+        prompt_len: The length of the prompt in tokens.
+        expected_output_len: The expected length of the output in tokens.
+    """
+    prompt: str
+    prompt_len: int
+    expected_output_len: int
+    multi_modal_data: Optional[MultiModalDataDict] = None
+
+
 def sample_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     fixed_output_len: Optional[int],
-) -> List[Tuple[str, int, int]]:
+) -> List[SampleRequest]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
 
@@ -41,7 +60,7 @@ def sample_requests(
     random.shuffle(dataset)
 
     # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_dataset: List[SampleRequest] = []
     for i in range(len(dataset)):
         if len(filtered_dataset) == num_requests:
             break
@@ -60,31 +79,34 @@ def sample_requests(
         if prompt_len > 1024 or prompt_len + output_len > 2048:
             # Prune too long sequences.
             continue
-        filtered_dataset.append((prompt, prompt_len, output_len))
+        filtered_dataset.append(
+            SampleRequest(prompt=prompt,
+                          prompt_len=prompt_len,
+                          expected_output_len=output_len))
 
     return filtered_dataset
 
 
 def run_vllm(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
     n: int,
     engine_args: EngineArgs,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
 
     # Add the requests to the engine.
-    prompts: List[str] = []
+    prompts: List[TextPrompt] = []
     sampling_params: List[SamplingParams] = []
-    for prompt, _, output_len in requests:
-        prompts.append(prompt)
+    for request in requests:
+        prompts.append(TextPrompt(prompt=request.prompt))
         sampling_params.append(
             SamplingParams(
                 n=n,
                 temperature=1.0,
                 top_p=1.0,
                 ignore_eos=True,
-                max_tokens=output_len,
+                max_tokens=request.expected_output_len,
             ))
 
     use_beam_search = False
@@ -94,11 +116,11 @@ def run_vllm(
         llm.generate(prompts, sampling_params, use_tqdm=True)
         end = time.perf_counter()
     else:
-        prompts = [prompt for prompt, _, _ in requests]
+        prompts = [request.prompt for request in requests]
         # output_len should be the same for all requests.
         output_len = requests[0][2]
-        for prompt, input_len, _output_len in requests:
-            assert _output_len == output_len
+        for request in requests:
+            assert request.expected_output_len == output_len
         start = time.perf_counter()
         llm.beam_search(
             prompts,
@@ -112,7 +134,7 @@ def run_vllm(
 
 
 async def run_vllm_async(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
     n: int,
     engine_args: AsyncEngineArgs,
     disable_frontend_multiprocessing: bool = False,
@@ -123,17 +145,17 @@ async def run_vllm_async(
             engine_args, disable_frontend_multiprocessing) as llm:
 
         # Add the requests to the engine.
-        prompts: List[str] = []
+        prompts: List[TextPrompt] = []
         sampling_params: List[SamplingParams] = []
-        for prompt, _, output_len in requests:
-            prompts.append(prompt)
+        for request in requests:
+            prompts.append(TextPrompt(prompt=request.prompt))
             sampling_params.append(
                 SamplingParams(
                     n=n,
                     temperature=1.0,
                     top_p=1.0,
                     ignore_eos=True,
-                    max_tokens=output_len,
+                    max_tokens=request.expected_output_len,
                 ))
 
         generators = []
@@ -149,7 +171,7 @@ async def run_vllm_async(
 
 
 def run_hf(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
     model: str,
     tokenizer: PreTrainedTokenizerBase,
     n: int,
@@ -207,14 +229,14 @@ def run_hf(
 
 
 def run_mii(
-    requests: List[Tuple[str, int, int]],
+    requests: List[SampleRequest],
     model: str,
     tensor_parallel_size: int,
     output_len: int,
 ) -> float:
     from mii import client, serve
     llm = serve(model, tensor_parallel=tensor_parallel_size)
-    prompts = [prompt for prompt, _, _ in requests]
+    prompts = [request.prompt for request in requests]
 
     start = time.perf_counter()
     llm.generate(prompts, max_new_tokens=output_len)
@@ -243,8 +265,12 @@ def main(args: argparse.Namespace):
         else:
             raise ValueError(
                 f"Failed to synthesize a prompt with {args.input_len} tokens.")
-        requests = [(prompt, args.input_len, args.output_len)
-                    for _ in range(args.num_prompts)]
+        requests = [
+            SampleRequest(prompt=prompt,
+                          prompt_len=args.input_len,
+                          expected_output_len=args.output_len)
+            for _ in range(args.num_prompts)
+        ]
     else:
         requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
                                    args.output_len)
@@ -270,9 +296,10 @@ def main(args: argparse.Namespace):
                                args.output_len)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(prompt_len + output_len
-                           for _, prompt_len, output_len in requests)
-    total_output_tokens = sum(output_len for _, _, output_len in requests)
+    total_num_tokens = sum(request.prompt_len + request.expected_output_len
+                           for request in requests)
+    total_output_tokens = sum(request.expected_output_len
+                              for request in requests)
     print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
           f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
           f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
@@ -299,7 +326,9 @@ def main(args: argparse.Namespace):
     parser.add_argument("--dataset",
                         type=str,
                         default=None,
-                        help="Path to the dataset.")
+                        help="Path to the dataset. The dataset is expected to "
+                        "be a json in form of List[Dict[..., conversations: "
+                        "List[Dict[..., value: <prompt_or_response>]]]]")
     parser.add_argument("--input-len",
                         type=int,
                         default=None,

docs/source/models/vlm.rst

@@ -242,6 +242,10 @@ To consume the server, you can use the OpenAI client like in the example below:
 
 A full code example can be found in `examples/openai_chat_completion_client_for_multimodal.py <https://github.com/vllm-project/vllm/blob/main/examples/openai_chat_completion_client_for_multimodal.py>`_.
 
+.. tip::
+    Loading from local file paths is also supported on vLLM: You can specify the allowed local media path via ``--allowed-local-media-path`` when launching the API server/engine,
+    and pass the file path as ``url`` in the API request.
+
 .. tip::
     There is no need to place image placeholders in the text content of the API request - they are already represented by the image content.
     In fact, you can place image placeholders in the middle of the text by interleaving text and image content.

pyproject.toml

@@ -6,7 +6,7 @@ requires = [
     "packaging",
     "setuptools>=61",
     "setuptools-scm>=8.0",
-    "torch == 2.5.0",
+    "torch == 2.5.1",
     "wheel",
     "jinja2",
 ]

requirements-build.txt

@@ -4,6 +4,6 @@ ninja
 packaging
 setuptools>=61
 setuptools-scm>=8
-torch==2.5.0
+torch==2.5.1
 wheel
 jinja2

requirements-cuda.txt

@@ -4,7 +4,7 @@
 # Dependencies for NVIDIA GPUs
 ray >= 2.9
 nvidia-ml-py >= 12.560.30 # for pynvml package
-torch == 2.5.0
+torch == 2.5.1
 # These must be updated alongside torch
-torchvision == 0.20   # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
-xformers == 0.0.28.post2; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch 2.5.0
+torchvision == 0.20.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+xformers == 0.0.28.post3; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch 2.5.1

requirements-openvino.txt

@@ -1,7 +1,7 @@
 # Common dependencies
 -r requirements-common.txt
 
-torch == 2.5.0 #  should be aligned with "common" vLLM torch version
+torch == 2.5.1 #  should be aligned with "common" vLLM torch version
 openvino >= 2024.4.0 # since 2024.4.0 both CPU and GPU support Paged Attention
 
 optimum @ git+https://github.com/huggingface/optimum.git@main # latest optimum is used to support latest transformers version

requirements-test.in

@@ -18,7 +18,7 @@ ray[adag]==2.35
 sentence-transformers # required for embedding
 soundfile # required for audio test
 timm # required for internvl test
-torch==2.5.0
+torch==2.5.1
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 datamodel_code_generator # required for minicpm3 test

requirements-test.txt

@@ -492,7 +492,7 @@ timm==1.0.11
     # via -r requirements-test.in
 tokenizers==0.20.1
     # via transformers
-torch==2.5.0
+torch==2.5.1
     # via
     #   -r requirements-test.in
     #   accelerate
@@ -503,7 +503,7 @@ torch==2.5.0
     #   tensorizer
     #   timm
     #   torchvision
-torchvision==0.20.0
+torchvision==0.20.1
     # via timm
 tqdm==4.66.6
     # via