vllm-project · youkaichao · Aug 29, 2025 · Jul 2, 2025 · Jul 2, 2025 · Jul 8, 2025
diff --git a/.buildkite/test-pipeline.yaml b/.buildkite/test-pipeline.yaml
@@ -449,8 +449,8 @@ steps:
   - tests/quantization
   commands:
   # temporary install here since we need nightly, will move to requirements/test.in
-  # after torchao 0.12 release
-  - pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
+  # after torchao 0.12 release, and pin a working version of torchao nightly here
+  - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min

@@ -45,8 +45,8 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.7.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.7.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.8.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.8.0")
 
 #
 # Try to find python package with an executable that exactly matches

diff --git a/pyproject.toml b/pyproject.toml
@@ -6,7 +6,7 @@ requires = [
     "packaging>=24.2",
     "setuptools>=77.0.3,<80.0.0",
     "setuptools-scm>=8.0",
-    "torch == 2.7.1",
+    "torch == 2.8.0",
     "wheel",
     "jinja2",
 ]

diff --git a/requirements/build.txt b/requirements/build.txt
@@ -4,7 +4,8 @@ ninja
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
-torch==2.7.1
+torch==2.8.0
 wheel
 jinja2>=3.1.6
 regex
+build
diff --git a/requirements/cpu.txt b/requirements/cpu.txt
@@ -9,17 +9,16 @@ packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 --extra-index-url https://download.pytorch.org/whl/cpu
 torch==2.6.0+cpu; platform_machine == "x86_64" # torch>2.6.0+cpu has performance regression on x86 platform, see https://github.com/pytorch/pytorch/pull/151218
-torch==2.7.0; platform_system == "Darwin"
-torch==2.7.0; platform_machine == "ppc64le"
-torch==2.6.0; platform_machine == "aarch64" # for arm64 CPUs, torch 2.7.0 has a issue: https://github.com/vllm-project/vllm/issues/17960
+torch==2.8.0; platform_system == "Darwin"
+torch==2.8.0; platform_machine == "ppc64le" or platform_machine == "aarch64"
 
 # required for the image processor of minicpm-o-2_6, this must be updated alongside torch
 torchaudio; platform_machine != "ppc64le" and platform_machine != "s390x"
-torchaudio==2.7.0; platform_machine == "ppc64le"
+torchaudio==2.8.0; platform_machine == "ppc64le"
 
 # required for the image processor of phi3v, this must be updated alongside torch
 torchvision; platform_machine != "ppc64le" and platform_machine != "s390x"
-torchvision==0.22.0; platform_machine == "ppc64le"
+torchvision==0.23.0; platform_machine == "ppc64le"
 datasets # for benchmark scripts
 
 # Intel Extension for PyTorch, only for x86_64 CPUs

diff --git a/requirements/cuda.txt b/requirements/cuda.txt
@@ -6,9 +6,9 @@ numba == 0.61.2; python_version > '3.9'
 
 # Dependencies for NVIDIA GPUs
 ray[cgraph]>=2.48.0 # Ray Compiled Graph, required for pipeline parallelism in V1.
-torch==2.7.1
-torchaudio==2.7.1
+torch==2.8.0
+torchaudio==2.8.0
 # These must be updated alongside torch
-torchvision==0.22.1 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
-# https://github.com/facebookresearch/xformers/releases/tag/v0.0.31
-xformers==0.0.31; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch >= 2.7
+torchvision==0.23.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
+# https://github.com/facebookresearch/xformers/releases/tag/v0.0.32.post1
+xformers==0.0.32.post1; platform_system == 'Linux' and platform_machine == 'x86_64'  # Requires PyTorch >= 2.8
diff --git a/requirements/rocm-build.txt b/requirements/rocm-build.txt
@@ -1,10 +1,10 @@
 # Common dependencies
 -r common.txt
 
---extra-index-url https://download.pytorch.org/whl/rocm6.2.4
-torch==2.7.0
-torchvision==0.22.0
-torchaudio==2.7.0
+--extra-index-url https://download.pytorch.org/whl/rocm6.3
+torch==2.8.0
+torchvision==0.23.0
+torchaudio==2.8.0
 
 triton==3.3.0
 cmake>=3.26.1,<4

diff --git a/requirements/test.in b/requirements/test.in
@@ -22,9 +22,9 @@ sentence-transformers # required for embedding tests
 soundfile # required for audio tests
 jiwer # required for audio tests
 timm >=1.0.17 # required for internvl and gemma3n-mm test
-torch==2.7.1
-torchaudio==2.7.1
-torchvision==0.22.1
+torch==2.8.0
+torchaudio==2.8.0
+torchvision==0.23.0
 transformers_stream_generator # required for qwen-vl test
 matplotlib # required for qwen-vl test
 mistral_common[image,audio] >= 1.8.2 # required for voxtral test

diff --git a/requirements/test.txt b/requirements/test.txt
@@ -541,42 +541,42 @@ numpy==1.26.4
     #   tritonclient
     #   vocos
     #   xarray
-nvidia-cublas-cu12==12.8.3.14
+nvidia-cublas-cu12==12.8.4.1
     # via
     #   nvidia-cudnn-cu12
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cuda-cupti-cu12==12.8.57
+nvidia-cuda-cupti-cu12==12.8.90
     # via torch
-nvidia-cuda-nvrtc-cu12==12.8.61
+nvidia-cuda-nvrtc-cu12==12.8.93
     # via torch
-nvidia-cuda-runtime-cu12==12.8.57
+nvidia-cuda-runtime-cu12==12.8.90
     # via torch
-nvidia-cudnn-cu12==9.7.1.26
+nvidia-cudnn-cu12==9.10.2.21
     # via torch
-nvidia-cufft-cu12==11.3.3.41
+nvidia-cufft-cu12==11.3.3.83
     # via torch
-nvidia-cufile-cu12==1.13.0.11
+nvidia-cufile-cu12==1.13.1.3
     # via torch
-nvidia-curand-cu12==10.3.9.55
+nvidia-curand-cu12==10.3.9.90
     # via torch
-nvidia-cusolver-cu12==11.7.2.55
+nvidia-cusolver-cu12==11.7.3.90
     # via torch
-nvidia-cusparse-cu12==12.5.7.53
+nvidia-cusparse-cu12==12.5.8.93
     # via
     #   nvidia-cusolver-cu12
     #   torch
-nvidia-cusparselt-cu12==0.6.3
+nvidia-cusparselt-cu12==0.7.1
     # via torch
-nvidia-nccl-cu12==2.26.2
+nvidia-nccl-cu12==2.27.3
     # via torch
-nvidia-nvjitlink-cu12==12.8.61
+nvidia-nvjitlink-cu12==12.8.93
     # via
     #   nvidia-cufft-cu12
     #   nvidia-cusolver-cu12
     #   nvidia-cusparse-cu12
     #   torch
-nvidia-nvtx-cu12==12.8.55
+nvidia-nvtx-cu12==12.8.90
     # via torch
 omegaconf==2.3.0
     # via
@@ -1069,7 +1069,7 @@ tomli==2.2.1
     # via schemathesis
 tomli-w==1.2.0
     # via schemathesis
-torch==2.7.1+cu128
+torch==2.8.0+cu128
     # via
     #   -r requirements/test.in
     #   accelerate
@@ -1098,7 +1098,7 @@ torch==2.7.1+cu128
     #   torchvision
     #   vector-quantize-pytorch
     #   vocos
-torchaudio==2.7.1+cu128
+torchaudio==2.8.0+cu128
     # via
     #   -r requirements/test.in
     #   encodec
@@ -1111,7 +1111,7 @@ torchmetrics==1.7.4
     #   pytorch-lightning
     #   terratorch
     #   torchgeo
-torchvision==0.22.1+cu128
+torchvision==0.23.0+cu128
     # via
     #   -r requirements/test.in
     #   lightly
@@ -1152,7 +1152,7 @@ transformers==4.55.2
     #   transformers-stream-generator
 transformers-stream-generator==0.0.5
     # via -r requirements/test.in
-triton==3.3.1
+triton==3.4.0
     # via torch
 tritonclient==2.51.0
     # via

diff --git a/tests/distributed/test_sequence_parallel.py b/tests/distributed/test_sequence_parallel.py
@@ -292,7 +292,7 @@ def _compare_sp(
     # TODO support other models
     # [LANGUAGE GENERATION]
     "meta-llama/Llama-3.2-1B-Instruct",
-    "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8"
+    "RedHatAI/Meta-Llama-3.1-8B-Instruct-FP8",
 ]
 
 

@@ -87,6 +87,9 @@ def test_chatglm3_lora_tp4(chatglm3_lora_files):
 @multi_gpu_test(num_gpus=4)
 @create_new_process_for_each_test()
 def test_chatglm3_lora_tp4_fully_sharded_loras(chatglm3_lora_files):
+    # https://github.com/NVIDIA/nccl/issues/1790, set a lower value for
+    # gpu_memory_utilization here because NCCL >= 2.26.3 seems to use
+    # more GPU memory causing vLLM to OOM
     llm = vllm.LLM(MODEL_PATH,
                    max_model_len=1024,
                    enable_lora=True,
@@ -95,7 +98,8 @@ def test_chatglm3_lora_tp4_fully_sharded_loras(chatglm3_lora_files):
                    tensor_parallel_size=4,
                    trust_remote_code=True,
                    fully_sharded_loras=True,
-                   enable_chunked_prefill=True)
+                   enable_chunked_prefill=True,
+                   gpu_memory_utilization=0.85)
     output1 = do_sample(llm, chatglm3_lora_files, lora_id=1)
     for i in range(len(EXPECTED_LORA_OUTPUT)):
         assert output1[i] == EXPECTED_LORA_OUTPUT[i]

@@ -789,6 +789,7 @@ def get_kernel_options(query, block_m, block_n,
             device_props = torch.cuda.get_device_properties()
             max_shared_memory = device_props.shared_memory_per_block_optin
             if max_shared_memory < 144 * 1024:
-                kernel_options["BLOCK_M"] = 32
-                kernel_options["BLOCK_N"] = 32
+                kernel_options["BLOCK_M"] = kernel_options["BLOCK_M"] // 2
+                kernel_options["BLOCK_N"] = kernel_options["BLOCK_N"] // 2
+
     return kernel_options