different rocm

;
;

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -60,39 +60,26 @@ COPY --from=ghcr.io/astral-sh/uv:latest /uv /usr/local/bin/uv
 
 USER ray
 
-# Install vLLM ROCm build and inference dependencies.
+# Install vLLM ROCm build, torch ROCm 7.0, and inference dependencies.
-# The vLLM ROCm wheel from wheels.vllm.ai includes HIP-compiled C-extensions
+# IMPORTANT: vLLM ROCm wheel (0.15.1+rocm700) has C-extensions compiled against
-# (vllm._C, vllm._rocm_C) that are ABI-compatible with ROCm PyTorch.
+# the official PyTorch ROCm 7.0 ABI. We MUST use torch from the same ROCm 7.0
-# PyPI vLLM is CUDA-only and crashes with: libcudart.so.12 not found.
+# index — TheRock nightlies have an incompatible c10::hip ABI that causes
-# uv gives --extra-index-url higher priority than PyPI, so the ROCm wheel
+# undefined symbol errors in vllm._C and vllm._rocm_C.
-# is selected over the CUDA wheel.
+# HSA_OVERRIDE_GFX_VERSION=11.0.0 makes gfx1151 appear as gfx1100 to this torch.
 RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
     uv pip install --system \
         --extra-index-url https://wheels.vllm.ai/rocm/ \
+        --extra-index-url https://download.pytorch.org/whl/rocm7.0 \
         vllm \
+        torch torchaudio torchvision \
         'transformers>=4.35.0,<5.0' \
         'accelerate>=0.25.0,<1.0' \
         'sentence-transformers>=2.3.0,<3.0' \
         'httpx>=0.27.0,<1.0' \
         'scipy>=1.11.0,<2.0'
 
-# WORKAROUND: ROCm/ROCm#5853 - Standard PyTorch ROCm wheels cause segfault
-# in libhsa-runtime64.so during VRAM allocation on gfx1151 (Strix Halo).
-# TheRock gfx110X-all packages provide compatible Python 3.12 wheels.
-# Reinstall AFTER vLLM to override the standard ROCm torch it pulled in.
-# vLLM's ROCm C-extensions remain compatible (same HIP ABI, torch 2.10.x).
-RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
-    uv pip install --system --reinstall \
-        --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/ \
-        torch torchaudio torchvision
-
-# FIX: Uninstall flash_attn — it was compiled against the vLLM ROCm wheel's
-# PyTorch, but the TheRock nightly above has a different c10::hip ABI.
-# vLLM ROCm uses its own Triton/CK attention backends, so flash_attn is not needed.
-RUN pip uninstall -y flash-attn 2>/dev/null || true
-
 # FIX: Ray base image has pandas 1.5.3 which is incompatible with numpy 2.x
-# The TheRock PyTorch wheels require numpy 2.x, so upgrade pandas to match.
+# PyTorch ROCm 7.0 requires numpy 2.x, so upgrade pandas to match.
 # Pin numpy <2.3 because numba (required by vLLM for speculative decoding)
 # does not yet support numpy 2.3+.
 RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \