[CPU] Fix torch version in x86 CPU backend (#19258)

Signed-off-by: jiang1.li <jiang1.li@intel.com>

csrc/cpu/torch_bindings.cpp

@@ -131,16 +131,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   // Quantization
 #ifdef __AVX512F__
+  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
       "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
-      "Tensor? azp) -> ()");
+      "Tensor? azp) -> ()",
+      {stride_tag});
   ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
 
   // Compute int8 quantized tensor and scaling factor
   ops.def(
       "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
-      "Tensor!? azp) -> ()");
+      "Tensor!? azp) -> ()",
+      {stride_tag});
   ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
            &dynamic_scaled_int8_quant);
   // W8A8 GEMM, supporting symmetric per-tensor or per-row/column
@@ -148,7 +151,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "cutlass_scaled_mm(Tensor! out, Tensor a,"
       "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()");
+      "                  Tensor b_scales, Tensor? bias) -> ()",
+      {stride_tag});
   ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm);
   // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
   // quantization.
@@ -156,7 +160,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
       "                  Tensor b, Tensor a_scales,"
       "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()");
+      "                  Tensor? azp, Tensor? bias) -> ()",
+      {stride_tag});
   ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #elif defined(__powerpc64__)
   // Compute int8 quantized tensor for given scaling factor.

docker/Dockerfile.cpu

@@ -66,7 +66,7 @@ ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
 WORKDIR /workspace/vllm
 
 RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,src=requirements/build.txt,target=requirements/build.txt \
+    --mount=type=bind,src=requirements/cpu-build.txt,target=requirements/build.txt \
     uv pip install -r requirements/build.txt
 
 COPY . .
@@ -79,6 +79,22 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel 
 
+######################### TEST DEPS #########################
+FROM base AS vllm-test-deps
+
+WORKDIR /workspace/vllm
+
+RUN --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
+    cp requirements/test.in requirements/cpu-test.in && \
+    sed -i '/mamba_ssm/d' requirements/cpu-test.in && \
+    sed -i 's/torch==.*/torch==2.6.0/g' requirements/cpu-test.in && \
+    sed -i 's/torchaudio.*/torchaudio/g' requirements/cpu-test.in && \
+    sed -i 's/torchvision.*/torchvision/g' requirements/cpu-test.in && \
+    uv pip compile requirements/cpu-test.in -o requirements/cpu-test.txt --index-strategy unsafe-best-match --torch-backend cpu
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -r requirements/cpu-test.txt 
+
 ######################### DEV IMAGE #########################
 FROM vllm-build AS vllm-dev
 
@@ -97,28 +113,19 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     VLLM_TARGET_DEVICE=cpu python3 setup.py develop 
 
+COPY --from=vllm-test-deps /workspace/vllm/requirements/cpu-test.txt requirements/test.txt
+
 RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
-    cp requirements/test.in requirements/test-cpu.in && \
-    sed -i '/mamba_ssm/d' requirements/test-cpu.in && \
-    uv pip compile requirements/test-cpu.in -o requirements/test.txt && \
     uv pip install -r requirements/dev.txt && \
     pre-commit install --hook-type pre-commit --hook-type commit-msg
 
 ENTRYPOINT ["bash"]
 
 ######################### TEST IMAGE #########################
-FROM base AS vllm-test
+FROM vllm-test-deps AS vllm-test
 
 WORKDIR /workspace/
 
-RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,src=requirements/test.in,target=requirements/test.in \
-    cp requirements/test.in requirements/test-cpu.in && \
-    sed -i '/mamba_ssm/d' requirements/test-cpu.in && \
-    uv pip compile requirements/test-cpu.in -o requirements/cpu-test.txt && \
-    uv pip install -r requirements/cpu-test.txt
-
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,from=vllm-build,src=/workspace/vllm/dist,target=dist \
     uv pip install dist/*.whl

requirements/cpu-build.txt

@@ -0,0 +1,12 @@
+# Temporarily used for x86 CPU backend to avoid performance regression of torch>2.6.0+cpu,
+# see https://github.com/pytorch/pytorch/pull/151218
+cmake>=3.26.1
+ninja
+packaging>=24.2
+setuptools>=77.0.3,<80.0.0
+setuptools-scm>=8
+--extra-index-url https://download.pytorch.org/whl/cpu
+torch==2.6.0+cpu
+wheel
+jinja2>=3.1.6
+regex

requirements/cpu.txt

@@ -8,7 +8,7 @@ numba == 0.61.2; python_version > '3.9'
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 --extra-index-url https://download.pytorch.org/whl/cpu
-torch==2.7.0+cpu; platform_machine == "x86_64"
+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" or platform_machine == "aarch64"
 
@@ -23,6 +23,7 @@ datasets # for benchmark scripts
 
 # Intel Extension for PyTorch, only for x86_64 CPUs
 intel-openmp==2024.2.1; platform_machine == "x86_64"
-intel_extension_for_pytorch==2.7.0; platform_machine == "x86_64"
+intel_extension_for_pytorch==2.6.0; platform_machine == "x86_64" # torch>2.6.0+cpu has performance regression on x86 platform, see https://github.com/pytorch/pytorch/pull/151218
 py-libnuma; platform_system != "Darwin"
 psutil; platform_system != "Darwin"
+triton==3.2.0; platform_machine == "x86_64" # Triton is required for torch 2.6+cpu, as it is imported in torch.compile.

tests/models/multimodal/generation/test_common.py

@@ -107,6 +107,8 @@ VLM_TEST_SETTINGS = {
             ),
             limit_mm_per_prompt={"image": 4},
         )],
+        # TODO: Revert to "auto" when CPU backend can use torch > 2.6
+        dtype="bfloat16" if current_platform.is_cpu() else "auto",
         marks=[pytest.mark.core_model, pytest.mark.cpu_model],
     ),
     "paligemma": VLMTestInfo(

tests/models/multimodal/generation/vlm_utils/builders.py

@@ -203,6 +203,9 @@ def build_embedding_inputs_from_test_info(
 
     images = [asset.pil_image for asset in image_assets]
     embeds = test_info.convert_assets_to_embeddings(image_assets)
+    if test_info.dtype != "auto":
+        dtype = getattr(torch, test_info.dtype)  # type: ignore
+        embeds = [e.to(dtype=dtype) for e in embeds]
     assert len(images) == len(model_prompts)
 
     inputs = build_single_image_inputs(images, model_prompts, size_wrapper)

vllm/model_executor/layers/fused_moe/layer.py

@@ -54,6 +54,8 @@ else:
 if is_rocm_aiter_moe_enabled():
     from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa: E501
         rocm_aiter_grouped_topk as grouped_topk)
+elif current_platform.is_cpu():
+    pass
 else:
     from vllm.model_executor.layers.fused_moe.fused_moe import grouped_topk
 if current_platform.is_tpu():

vllm/model_executor/layers/quantization/ipex_quant.py

@@ -15,7 +15,7 @@ from vllm.model_executor.layers.quantization.base_config import (
 from vllm.model_executor.layers.quantization.gptq import GPTQLinearMethod
 from vllm.platforms import current_platform
 
-MIN_IPEX_VERSION = "2.7.0"
+MIN_IPEX_VERSION = "2.6.0"
 
 
 class IPEXConfig(QuantizationConfig):