Target only H200 (sm_90) for SageAttention build

Blackwell (sm_100) may not be fully supported yet.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Dockerfile

@@ -65,8 +65,9 @@ WORKDIR /tmp
 ENV EXT_PARALLEL=2
 ENV NVCC_APPEND_FLAGS="--threads 2"
 ENV MAX_JOBS=4
-# Target RunPod GPU architectures: H100/H200(9.0), RTX 5090/Blackwell(10.0)
-ENV TORCH_CUDA_ARCH_LIST="9.0;10.0"
+# Target RunPod GPU architectures: H100/H200(9.0)
+# Note: Blackwell (10.0) not yet supported by SageAttention
+ENV TORCH_CUDA_ARCH_LIST="9.0"
 RUN git clone https://github.com/thu-ml/SageAttention.git && \
     cd SageAttention && \
     pip install --no-build-isolation . && \