From 35f6b666781a118c4e491884619fb26afc2318fc Mon Sep 17 00:00:00 2001 From: biondizzle Date: Mon, 11 May 2026 19:40:08 +0000 Subject: [PATCH] fix: UE8M0 reinterpret in DeepGEMM fold_global_scale + bump cache --- Dockerfile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Dockerfile b/Dockerfile index 340583a..a9e45a4 100644 --- a/Dockerfile +++ b/Dockerfile @@ -8,7 +8,7 @@ RUN apt-get update && apt-get install -y git screen cmake && rm -rf /var/lib/apt # Clone and build DeepGEMM with NVFP4 mega_moe kernel # CACHE_BUSTER: increment to force fresh clone -RUN git clone -b nvfp4-mega-moe https://sweetapi.com/biondizzle/DeepGEMM.git /root/DeepGEMM && CACHE_BUSTER=24 +RUN git clone -b nvfp4-mega-moe https://sweetapi.com/biondizzle/DeepGEMM.git /root/DeepGEMM && CACHE_BUSTER=28 # Build DeepGEMM with proper CUDA/NVRTC paths ENV CPATH="/usr/local/lib/python3.12/dist-packages/flashinfer/data/cutlass/include:/usr/local/lib/python3.12/dist-packages/nvidia/cu13/include:/usr/local/cuda-13.0/include:${CPATH}" @@ -19,7 +19,7 @@ RUN ln -sf /usr/local/lib/python3.12/dist-packages/nvidia/cu13/lib/libnvrtc.so.1 RUN cd /root/DeepGEMM && python3 setup.py build_ext --inplace # Bust cache for patch changes — ARG before COPY ensures layer invalidation -ARG PATCH_CACHE_BUSTER=27 +ARG PATCH_CACHE_BUSTER=28 # 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