more fixes.

dockerfiles/Dockerfile.ray-worker-nvidia

@@ -45,6 +45,7 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
         'librosa>=0.10.0,<1.0' \
         'torch>=2.0.0,<3.0' \
         'torchaudio>=2.0.0,<3.0' \
+        'torchcodec' \
         'transformers>=4.46.0,<5.0' \
         'fastapi>=0.100.0,<1.0' \
         'uvicorn>=0.23.0,<1.0' \

dockerfiles/Dockerfile.ray-worker-strixhalo

@@ -86,6 +86,11 @@ RUN --mount=type=cache,target=/home/ray/.cache/uv,uid=1000,gid=1000 \
         --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.
 # Pin numpy <2.3 because numba (required by vLLM for speculative decoding)