[IGNORE][CI] Prevent vllm-base docker build stage from depending on entire vLLM wheel

docker/Dockerfile

@@ -112,6 +112,8 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 # Use copy mode to avoid hardlink failures with Docker cache mounts
 ENV UV_LINK_MODE=copy
 
+ENV SCCACHE_LOG="debug"
+
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
 RUN apt-get install -y gcc-10 g++-10
@@ -134,7 +136,9 @@ COPY requirements/cuda.txt requirements/cuda.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
     # TODO: remove apache-tvm-ffi once FlashInfer is fixed https://github.com/flashinfer-ai/flashinfer/issues/1962
     uv pip install --python /opt/venv/bin/python3 --pre apache-tvm-ffi==0.1.0b15 \
+    --prerelease=allow \
     && uv pip install --python /opt/venv/bin/python3 -r requirements/cuda.txt \
+    --prerelease=allow \
     --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
 # cuda arch list used by torch
@@ -153,6 +157,12 @@ ARG PIP_INDEX_URL UV_INDEX_URL
 ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
 ARG PYTORCH_CUDA_INDEX_BASE_URL
 
+# Ensure build caching is enabled
+ENV CMAKE_C_COMPILER_LAUNCHER=sccache
+ENV CMAKE_CXX_COMPILER_LAUNCHER=sccache
+ENV CMAKE_CUDA_COMPILER_LAUNCHER=sccache
+
+
 # install build dependencies
 COPY requirements/build.txt requirements/build.txt
 
@@ -173,10 +183,10 @@ RUN --mount=type=bind,source=.git,target=.git \
     if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
 
 # max jobs used by Ninja to build extensions
-ARG max_jobs=2
+ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
 # number of threads used by nvcc
-ARG nvcc_threads=8
+ARG nvcc_threads=32
 ENV NVCC_THREADS=$nvcc_threads
 
 ARG USE_SCCACHE
@@ -353,13 +363,13 @@ RUN --mount=type=cache,target=/root/.cache/uv \
             --pre pytorch_triton==3.3.0+gitab727c40 ; \
     fi
 
-# Install vllm wheel first, so that torch etc will be installed.
-RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
-    --mount=type=cache,target=/root/.cache/uv \
-    # TODO: remove apache-tvm-ffi once FlashInfer is fixed https://github.com/flashinfer-ai/flashinfer/issues/1962
-    uv pip install --system --pre apache-tvm-ffi==0.1.0b15 \
-    && uv pip install --system dist/*.whl --verbose \
-        --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
+# install build and runtime dependencies
+COPY requirements/common.txt requirements/common.txt
+COPY requirements/cuda.txt requirements/cuda.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r requirements/cuda.txt \
+    --prerelease=allow \
+    --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
 
 # TODO (huydhn): Remove this once xformers is released for 2.9.0
 RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
@@ -428,8 +438,6 @@ ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}
 # note that this uses vllm installed by `pip`
 FROM vllm-base AS test
 
-ADD . /vllm-workspace/
-
 ARG PYTHON_VERSION
 
 ARG PIP_INDEX_URL UV_INDEX_URL
@@ -444,13 +452,19 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
 ENV UV_LINK_MODE=copy
 
 # install development dependencies (for testing)
+COPY requirements/lint.txt requirements/lint.txt
+COPY requirements/test.txt requirements/test.txt
+COPY requirements/dev.txt requirements/dev.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
     CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
     if [ "$CUDA_MAJOR" -ge 12 ]; then \
         uv pip install --system -r requirements/dev.txt \
         --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
     fi
 
+# Move host vLLM workspace into image
+ADD . /vllm-workspace/
+
 # install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -e tests/vllm_test_utils
@@ -463,6 +477,14 @@ ENV HF_HUB_ENABLE_HF_TRANSFER 1
 # Copy in the v1 package for testing (it isn't distributed yet)
 COPY vllm/v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1
 
+# Install vllm wheel
+RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
+    --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system --pre apache-tvm-ffi==0.1.0b15 \
+        --prerelease=allow \
+        dist/*.whl --verbose \
+        --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
+
 # Source code is used in the `python_only_compile.sh` test
 # We hide it inside `src/` so that this source code
 # will not be imported by other tests

vllm/v1/attention/backends/flash_attn.py

@@ -303,7 +303,7 @@ def build(
         max_num_splits = 0  # 0 means use FA3's heuristics, not CG compatible
         if self.use_full_cuda_graph and num_actual_tokens <= self.max_cudagraph_size:
             # NOTE(woosuk): Setting num_splits > 1 may increase the memory
-            # usage, because the intermediate buffers of size [num_splits,
+            # usage, because theintermediate bffers of size [num_splits,
             # num_heads, num_tokens, head_size] are allocated. Therefore,
             # we only set num_splits when using cuda graphs.
             max_num_splits = self.max_num_splits

vllm/v1/cudagraph_dispatcher.py

@@ -11,9 +11,9 @@ class CudagraphDispatcher:
     Runtime cudagraph dispatcher to dispatch keys for multiple set of
     cudagraphs.
 
-    The dispatcher stores two sets of dispatch keys, one for PIECEWISE and one
+    The dispatcherstores two sets of dispatch keys, one for PIECEWISE and one
     for FULL cudagraph runtime mode. The keys are initialized depending on
-    attention support and what cudagraph mode is set in CompilationConfig. The
+    attention support and what cudagraph modeis set in CompilationConfig. The
     keys stored in dispatcher are the only source of truth for valid
     cudagraphs that can be dispatched at runtime.