From d88ea9842b7182b3c026b7ba080e448cfe29c748 Mon Sep 17 00:00:00 2001 From: biondizzle Date: Tue, 12 May 2026 21:21:24 +0000 Subject: [PATCH] =?UTF-8?q?fix:=20add=20missing=20staging=5Fkernel.py=20to?= =?UTF-8?q?=20Dockerfile=20=E2=80=94=20BF16=E2=86=92E2M1+UE4M3=20quantizat?= =?UTF-8?q?ion=20was=20never=20in=20container?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- Dockerfile | 2 ++ 1 file changed, 2 insertions(+) diff --git a/Dockerfile b/Dockerfile index 6eda6d7..2f32a5b 100644 --- a/Dockerfile +++ b/Dockerfile @@ -22,6 +22,8 @@ RUN cd /root/DeepGEMM && python3 setup.py build_ext --inplace && PATCH_CACHE_BUS ARG PATCH_CACHE_BUSTER=69 # Copy our DeepSeek V4 patch over vLLM's model file COPY patches/deepseek_v4.py /usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/deepseek_v4.py +# Copy the NVFP4 staging kernel (BF16→E2M1+UE4M3 quantization for activations) +COPY patches/staging_kernel.py /usr/local/lib/python3.12/dist-packages/vllm/model_executor/models/staging_kernel.py # Verify everything imports RUN python3 -c "import deep_gemm; print('DeepGEMM NVFP4 OK')" && \