merge commit 3

Signed-off-by: khluu <khluu000@gmail.com>

.buildkite/release-pipeline.yaml

@@ -8,6 +8,21 @@ steps:
   - group: "Build Python wheels"
     key: "build-wheels"
     steps:
+      - label: "Build wheel - aarch64 - CUDA 12.8"
+        depends_on: ~
+        id: build-wheel-arm64-cuda-12-8
+        agents:
+          queue: arm64_cpu_queue_postmerge
+        commands:
+          # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
+          # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --build-arg torch_cuda_arch_list='8.7 8.9 9.0 10.0+PTX' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "mkdir artifacts"
+          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh"
+        env:
+          DOCKER_BUILDKIT: "1"
+
       - label: "Build wheel - aarch64 - CUDA 12.9"
         depends_on: ~
         id: build-wheel-arm64-cuda-12-9
@@ -51,6 +66,19 @@ steps:
         env:
           DOCKER_BUILDKIT: "1"
 
+      - label: "Build wheel - x86_64 - CUDA 12.8"
+        depends_on: ~
+        id: build-wheel-x86-cuda-12-8
+        agents:
+          queue: cpu_queue_postmerge
+        commands:
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.8.1 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+          - "mkdir artifacts"
+          - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+          - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_31"
+        env:
+          DOCKER_BUILDKIT: "1"
+
       - label: "Build wheel - x86_64 - CUDA 12.9"
         depends_on: ~
         id: build-wheel-x86-cuda-12-9

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG c2afa9cb93e674d5a9120a170a6da57b89267208
+        GIT_TAG f46984b4caf19409bf4ab94a540341918dd6a1a3
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""

csrc/activation_kernels.cu

@@ -9,6 +9,116 @@
 
 namespace vllm {
 
+struct alignas(32) u32x8_t {
+  uint32_t u0, u1, u2, u3, u4, u5, u6, u7;
+};
+
+// 256-bit vector loads require both sm_100+ and CUDA 12.9+
+#define VLLM_SUPPORTS_256BIT_VECTORS                  \
+  (defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && \
+   __CUDACC_VER_MAJOR__ >= 12 && __CUDACC_VER_MINOR__ >= 9)
+
+__device__ __forceinline__ void ld256(u32x8_t& val, const u32x8_t* ptr) {
+#if VLLM_SUPPORTS_256BIT_VECTORS
+  asm volatile("ld.global.nc.v8.u32 {%0,%1,%2,%3,%4,%5,%6,%7}, [%8];\n"
+               : "=r"(val.u0), "=r"(val.u1), "=r"(val.u2), "=r"(val.u3),
+                 "=r"(val.u4), "=r"(val.u5), "=r"(val.u6), "=r"(val.u7)
+               : "l"(ptr));
+#else
+  const uint4* uint_ptr = reinterpret_cast<const uint4*>(ptr);
+  uint4 top_half = __ldg(&uint_ptr[0]);
+  uint4 bottom_half = __ldg(&uint_ptr[1]);
+  val.u0 = top_half.x;
+  val.u1 = top_half.y;
+  val.u2 = top_half.z;
+  val.u3 = top_half.w;
+  val.u4 = bottom_half.x;
+  val.u5 = bottom_half.y;
+  val.u6 = bottom_half.z;
+  val.u7 = bottom_half.w;
+#endif
+}
+
+__device__ __forceinline__ void st256(u32x8_t& val, u32x8_t* ptr) {
+#if VLLM_SUPPORTS_256BIT_VECTORS
+  asm volatile("st.global.v8.u32 [%0], {%1,%2,%3,%4,%5,%6,%7,%8};\n"
+               :
+               : "l"(ptr), "r"(val.u0), "r"(val.u1), "r"(val.u2), "r"(val.u3),
+                 "r"(val.u4), "r"(val.u5), "r"(val.u6), "r"(val.u7)
+               : "memory");
+#else
+  uint4* uint_ptr = reinterpret_cast<uint4*>(ptr);
+  uint_ptr[0] = make_uint4(val.u0, val.u1, val.u2, val.u3);
+  uint_ptr[1] = make_uint4(val.u4, val.u5, val.u6, val.u7);
+#endif
+}
+
+template <bool support_256>
+struct VecTraits;
+
+template <>
+struct VecTraits<true> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 32;
+  using vec_t = u32x8_t;
+};
+
+template <>
+struct VecTraits<false> {
+  static constexpr int ARCH_MAX_VEC_SIZE = 16;
+  using vec_t = int4;
+};
+
+template <typename T>
+struct PackedTraits;
+
+template <>
+struct PackedTraits<c10::BFloat16> {
+  using packed_t = __nv_bfloat162;
+};
+
+template <>
+struct PackedTraits<c10::Half> {
+  using packed_t = __half2;
+};
+
+template <>
+struct PackedTraits<float> {
+  using packed_t = float2;
+};
+
+template <typename packed_t>
+__device__ __forceinline__ float2 cast_to_float2(const packed_t& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __bfloat1622float2(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __half22float2(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t cast_to_packed(const float2& val) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162>) {
+    return __float22bfloat162_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, __half2>) {
+    return __float22half2_rn(val);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return float2(val);
+  }
+}
+
+template <typename packed_t>
+__device__ __forceinline__ packed_t packed_mul(const packed_t& x,
+                                               const packed_t& y) {
+  if constexpr (std::is_same_v<packed_t, __nv_bfloat162> ||
+                std::is_same_v<packed_t, __half2>) {
+    return __hmul2(x, y);
+  } else if constexpr (std::is_same_v<packed_t, float2>) {
+    return make_float2(x.x * y.x, x.y * y.y);
+  }
+}
+
 template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
           bool act_first>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,

docker/Dockerfile

@@ -418,10 +418,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
     fi && \
     python3 setup.py bdist_wheel --dist-dir=dist --py-limited-api=cp38
 
-# Copy extension wheels from extensions-build stage for later use
-COPY --from=extensions-build /tmp/deepgemm/dist /tmp/deepgemm/dist
-COPY --from=extensions-build /tmp/ep_kernels_workspace/dist /tmp/ep_kernels_workspace/dist
-
 # Check the size of the wheel if RUN_WHEEL_CHECK is true
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
 # sync the default value with .buildkite/check-wheel-size.py
@@ -660,9 +656,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 . /etc/environment && \
 uv pip list
 
-# Install deepgemm wheel that has been built in the `build` stage
+# Install deepgemm wheel that has been built in the `extensions-build` stage
 RUN --mount=type=cache,target=/root/.cache/uv \
-    --mount=type=bind,from=build,source=/tmp/deepgemm/dist,target=/tmp/deepgemm/dist,ro \
+    --mount=type=bind,from=extensions-build,source=/tmp/deepgemm/dist,target=/tmp/deepgemm/dist,ro \
     sh -c 'if ls /tmp/deepgemm/dist/*.whl >/dev/null 2>&1; then \
               uv pip install --system /tmp/deepgemm/dist/*.whl; \
            else \
@@ -672,8 +668,8 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Pytorch now installs NVSHMEM, setting LD_LIBRARY_PATH
 ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
 
-# Install EP kernels wheels (pplx-kernels and DeepEP) that have been built in the `build` stage
-RUN --mount=type=bind,from=build,src=/tmp/ep_kernels_workspace/dist,target=/vllm-workspace/ep_kernels/dist \
+# Install EP kernels wheels (pplx-kernels and DeepEP) that have been built in the `extensions-build` stage
+RUN --mount=type=bind,from=extensions-build,src=/tmp/ep_kernels_workspace/dist,target=/vllm-workspace/ep_kernels/dist \
     --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system ep_kernels/dist/*.whl --verbose \
         --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')

setup.py

@@ -977,9 +977,9 @@ if _is_cuda():
         # FA3 requires CUDA 12.3 or later
         ext_modules.append(CMakeExtension(name="vllm.vllm_flash_attn._vllm_fa3_C"))
     if envs.VLLM_USE_PRECOMPILED or (
-        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.9")
+        CUDA_HOME and get_nvcc_cuda_version() >= Version("12.8")
     ):
-        # FlashMLA requires CUDA 12.9 or later
+        # FlashMLA requires CUDA 12.8 or later
         # Optional since this doesn't get built (produce an .so file) when
         # not targeting a hopper system
         ext_modules.append(CMakeExtension(name="vllm._flashmla_C", optional=True))