17 lines
604 B
Docker
17 lines
604 B
Docker
|
ARG BASE_IMAGE
|
||
|
|
FROM ${BASE_IMAGE}
|
||
|
|
|
||
|
|
COPY manifest.json /opt/ppu-deepseek-v4-pro-patch/manifest.json
|
||
|
|
COPY scripts/apply.py /opt/ppu-deepseek-v4-pro-patch/scripts/apply.py
|
||
|
|
COPY src/ /opt/ppu-deepseek-v4-pro-patch/src/
|
||
|
|
COPY upstream/ /opt/ppu-deepseek-v4-pro-patch/upstream/
|
||
|
|
COPY examples/ /opt/ppu-deepseek-v4-pro-patch/examples/
|
||
|
|
|
||
|
|
RUN python /opt/ppu-deepseek-v4-pro-patch/scripts/apply.py --apply
|
||
|
|
|
||
|
|
ENV VLLM_DISABLED_KERNELS=PPUCutlassFp8BlockScaledMMKernel \
|
||
|
|
VLLM_PPU_BF16_O_PROJ=1 \
|
||
|
|
VLLM_PPU_SERIAL_ATTENTION=1 \
|
||
|
|
VLLM_MULTI_STREAM_GEMM_TOKEN_THRESHOLD=0 \
|
||
|
|
VLLM_ENGINE_READY_TIMEOUT_S=3600
|