Add DeepGEMM to Dockerfile in vllm-base image (#21533)

Signed-off-by: Matthew Bonanni <mbonanni001@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>

docker/Dockerfile

@@ -1,4 +1,3 @@
-
 # The vLLM Dockerfile is used to construct vLLM image that can be directly used
 # to run the OpenAI compatible server.
 
@@ -16,6 +15,7 @@ ARG PYTHON_VERSION=3.12
 # Example:
 # docker build --build-arg BUILD_BASE_IMAGE=registry.acme.org/mirror/nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
 ARG BUILD_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04
+# TODO: Restore to base image after FlashInfer AOT wheel fixed
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
 # By parameterizing the Deadsnakes repository URL, we allow third-party to use
@@ -289,7 +289,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
-# TODO: Restore to base image after FlashInfer AOT wheel fixed
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 ARG CUDA_VERSION
 ARG PYTHON_VERSION
@@ -435,6 +434,33 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/build.txt \
         --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
+# Install DeepGEMM from source
+ARG DEEPGEMM_GIT_REPO="https://github.com/deepseek-ai/DeepGEMM.git"
+ARG DEEPGEMM_GIT_REF="187656694f7f69e3e7975617a68bc3387680a7e1"
+RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
+  . /etc/environment
+    CUDA_MAJOR="${CUDA_VERSION%%.*}"
+    CUDA_MINOR="${CUDA_VERSION#${CUDA_MAJOR}.}"
+    CUDA_MINOR="${CUDA_MINOR%%.*}"
+    if [ "$CUDA_MAJOR" -ge 12 ] && [ "$CUDA_MINOR" -ge 8 ]; then
+        git clone --recursive --shallow-submodules \
+            ${DEEPGEMM_GIT_REPO} deepgemm
+        echo "🏗️  Building DeepGEMM"
+        pushd deepgemm
+            git checkout ${DEEPGEMM_GIT_REF}
+            # Build DeepGEMM
+            # (Based on https://github.com/deepseek-ai/DeepGEMM/blob/main/install.sh)
+            rm -rf build dist
+            rm -rf *.egg-info
+            python3 setup.py bdist_wheel
+            uv pip install --system dist/*.whl
+        popd
+        rm -rf deepgemm
+    else
+        echo "Skipping DeepGEMM installation (requires CUDA 12.8+ but got ${CUDA_VERSION})"
+    fi
+BASH
+
 #################### vLLM installation IMAGE ####################
 
 #################### TEST IMAGE ####################

tests/kernels/moe/test_deepep_deepgemm_moe.py

@@ -20,7 +20,8 @@ from vllm.model_executor.layers.fused_moe.modular_kernel import (
     FusedMoEModularKernel)
 from vllm.platforms import current_platform
 from vllm.utils import has_deep_ep, has_deep_gemm
-from vllm.utils.deep_gemm import is_blackwell_deep_gemm_used
+from vllm.utils.deep_gemm import (is_blackwell_deep_gemm_used,
+                                  is_deep_gemm_supported)
 
 from .parallel_utils import ProcessGroupInfo, parallel_launch
 from .utils import make_test_weights
@@ -46,7 +47,7 @@ requires_deep_ep = pytest.mark.skipif(
 )
 
 requires_deep_gemm = pytest.mark.skipif(
-    not has_deep_gemm(),
+    not is_deep_gemm_supported(),
     reason="Requires deep_gemm kernels",
 )
 

tests/kernels/moe/test_deepgemm.py

@@ -15,13 +15,13 @@ import torch
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     per_token_group_quant_fp8)
-from vllm.utils import has_deep_gemm
-from vllm.utils.deep_gemm import calc_diff, per_block_cast_to_fp8
+from vllm.utils.deep_gemm import (calc_diff, is_deep_gemm_supported,
+                                  per_block_cast_to_fp8)
 
 BLOCK_SIZE = [128, 128]
 
 requires_deep_gemm = pytest.mark.skipif(
-    not has_deep_gemm(),
+    not is_deep_gemm_supported(),
     reason="Requires deep_gemm kernels",
 )
 

vllm/utils/deep_gemm.py

@@ -17,6 +17,17 @@ from vllm.platforms import current_platform
 from vllm.utils import has_deep_gemm
 
 
+@functools.cache
+def is_deep_gemm_supported() -> bool:
+    """Return ``True`` if DeepGEMM is supported on the current platform.
+    Currently, only Hopper and Blackwell GPUs are supported.
+    """
+    supported_arch = current_platform.is_cuda() and (
+        current_platform.is_device_capability(90)
+        or current_platform.is_device_capability(100))
+    return has_deep_gemm() and supported_arch
+
+
 @functools.cache
 def is_blackwell_deep_gemm_used() -> bool:
     """Return ``True`` if vLLM is configured to use DeepGEMM on a
@@ -142,4 +153,5 @@ __all__ = [
     "fp8_m_grouped_gemm_nt_masked",
     "per_block_cast_to_fp8",
     "is_blackwell_deep_gemm_used",
+    "is_deep_gemm_supported",
 ]