A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 01:00 UTC
[ROCm][CI] Build fastsafetensors from source so it links against libamdhip64 (#39978)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Apr 17, 2026 at 14:30 UTC
1ae11e2bfcf948876487bf7319f5bbb049768ba4
4 files changed
+3
-27
docker/Dockerfile.rocm
-24
@@ -262,30 +262,6 @@ RUN --mount=type=bind,source=.git,target=vllm/.git \
262
&& echo "Detected vLLM version: ${VLLM_VERSION}" \
263
&& echo "${VLLM_VERSION}" > /tmp/vllm_version.txt
264
265
-# Fail if git-based package dependencies are found in requirements files
266
-# (uv doesn't handle git+ URLs well, and packages should be distributed on PyPI)
267
-# Extra notes: pip install is able to handle git+ URLs, but uv doesn't.
268
-RUN echo "Checking for git-based packages in requirements files..." \
269
- && echo "Checking common.txt for git-based packages:" \
270
- && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; then \
271
- echo "ERROR: Git-based packages found in common.txt:"; \
272
- grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; \
273
- echo "Please publish these packages to PyPI instead of using git dependencies."; \
274
- exit 1; \
275
- else \
276
- echo " ✓ No git-based packages found in common.txt"; \
277
- fi \
278
- && echo "Checking rocm.txt for git-based packages:" \
279
- && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; then \
280
- echo "ERROR: Git-based packages found in rocm.txt:"; \
281
- grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; \
282
- echo "Please publish these packages to PyPI instead of using git dependencies."; \
283
- exit 1; \
284
- else \
285
- echo " ✓ No git-based packages found in rocm.txt"; \
286
- fi \
287
- && echo "All requirements files are clean - no git-based packages found"
288
-
265
# Pin vLLM dependencies to exact versions of custom ROCm wheels
266
# This ensures 'pip install vllm' automatically installs correct torch/triton/torchvision/amdsmi
267
COPY tools/vllm-rocm/pin_rocm_dependencies.py /tmp/pin_rocm_dependencies.py
requirements/rocm.txt
+1
-1
@@ -22,4 +22,4 @@ timm>=1.0.17
22
# To be consistent with test_quark.py
23
amd-quark>=0.8.99
24
# Required for faster safetensors model loading
25
-fastsafetensors >= 0.2.2
\ No newline at end of file
25
+fastsafetensors @ git+https://github.com/foundation-model-stack/fastsafetensors.git@0.2.2
\ No newline at end of file
requirements/test/rocm.in
+1
-1
@@ -55,7 +55,7 @@ arctic-inference==0.1.1 # Required for suffix decoding test
55
numba==0.61.2 # Required for N-gram speculative decoding
56
numpy
57
runai-model-streamer[s3,gcs,azure]==0.15.7
58
-fastsafetensors>=0.2.2 # 0.2.2 contains important fixes for multi-GPU mem usage
58
+fastsafetensors @ git+https://github.com/foundation-model-stack/fastsafetensors.git@0.2.2 # PyPI only ships CUDA wheels
59
instanttensor>=0.1.5
60
pydantic>=2.12 # 2.11 leads to error on python 3.13
61
decord==0.6.0
requirements/test/rocm.txt
+1
-1
@@ -277,7 +277,7 @@ fastar==0.10.0
277
# via fastapi-cloud-cli
278
fastparquet==2026.3.0
279
# via genai-perf
280
-fastsafetensors==0.2.2
280
+fastsafetensors @ git+https://github.com/foundation-model-stack/fastsafetensors.git@65d80088fca7a8f567fba30415fbcc80f7d2259c
281
# via
282
# -c requirements/rocm.txt
283
# -r requirements/test/rocm.in