minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix(vlm): preserve Kimi-K3 GPU JPEG accuracy (#34163)

Browse Source
This commit is contained in:
Mick
2026-08-10 09:42:52 +08:00
committed by GitHub
parent 553dc0f936
commit c20e99bd22
8 changed files with 258 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docker/kimi_k3/kimi_k3_cu12.Dockerfile
+7
View File
@@ -26,6 +26,7 @@
FROM lmsysorg/sglang:v0.5.16-cu129 AS base
ARG SGL_DEEP_GEMM_VERSION="0.1.5.post2"
ARG NVIMGCODEC_VERSION="0.9.0.20"
# Current Kimi-K3 source auto-discovers and builds its PyO3 extensions.
ARG RUST_VERSION="1.90.0"
@@ -89,6 +90,12 @@ RUN TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" \
RUN python3 -m pip install --no-deps --force-reinstall \
"https://github.com/sgl-project/whl/releases/download/v${SGL_DEEP_GEMM_VERSION}/sgl_deep_gemm-${SGL_DEEP_GEMM_VERSION}+cu129-py3-none-manylinux2014_x86_64.whl"
# High-fidelity GPU JPEG decode. The K3 processor enables nvJPEG interpolated
# chroma upsampling through nvImageCodec and zero-copy DLPack handoff to Torch.
RUN python3 -m pip install \
"nvidia-nvimgcodec-cu12[all]==${NVIMGCODEC_VERSION}" && \
rm -rf /root/.cache/pip
# Install the pinned FlashInfer MXFP4 MoE runner cubin pool.
ARG TRTLLM_GEN_MOE_CUBIN_URL="https://github.com/sgl-project/whl/releases/download/trtllm_gen_moe_cubin_20260617/trtllm_gen_moe_cubin_pool_20260617_v0613rc1.zip"
ARG TRTLLM_GEN_MOE_CUBIN_SHA256="4900501cbe782a76b08a5858f9f07152287b97cb68114466dac286366b66c192"
docker/kimi_k3/kimi_k3_cu13.Dockerfile
+7
View File
@@ -27,6 +27,7 @@
FROM lmsysorg/sglang:v0.5.16 AS base
ARG SGL_DEEP_GEMM_VERSION="0.1.5.post2"
ARG NVIMGCODEC_VERSION="0.9.0.20"
# Current Kimi-K3 source auto-discovers and builds its PyO3 extensions.
ARG RUST_VERSION="1.90.0"
@@ -79,6 +80,12 @@ RUN TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" \
RUN python3 -m pip install --no-deps --force-reinstall \
"sgl-deep-gemm==${SGL_DEEP_GEMM_VERSION}"
# High-fidelity GPU JPEG decode. The K3 processor enables nvJPEG interpolated
# chroma upsampling through nvImageCodec and zero-copy DLPack handoff to Torch.
RUN python3 -m pip install \
"nvidia-nvimgcodec-cu13[all]==${NVIMGCODEC_VERSION}" && \
rm -rf /root/.cache/pip
# Install the pinned FlashInfer MXFP4 MoE runner cubin pool.
ARG TRTLLM_GEN_MOE_CUBIN_URL="https://github.com/sgl-project/whl/releases/download/trtllm_gen_moe_cubin_20260617/trtllm_gen_moe_cubin_pool_20260617_v0613rc1.zip"
ARG TRTLLM_GEN_MOE_CUBIN_SHA256="4900501cbe782a76b08a5858f9f07152287b97cb68114466dac286366b66c192"