Dockerfile.runtime

ARG BASE_IMAGE=verdictai/glm52-exl3-sparkinfer:v21-mtp78tr3-gg6722c1d-si1a88b38-cu132-sm120a@sha256:9b1befc116b080f4d1328ede8ebe6f0896e10959e816c5db5da2b5445ae8423b
FROM ${BASE_IMAGE}

# Register the Glm5v MoonViT/PatchMerger wrapper. The EXL3 loader and Trellis
# kernels remain those of the sha-pinned base runtime.
COPY runtime/glm52_vision \
  /opt/venv/lib/python3.12/site-packages/glm52_vision
COPY runtime/glm52_vision_vllm-0.1.0.dist-info \
  /opt/venv/lib/python3.12/site-packages/glm52_vision_vllm-0.1.0.dist-info

# Preserve the fast DCP project-before-merge path for partial prefill batches.
COPY patch_dcp_workspace_stride.py /tmp/patch_dcp_workspace_stride.py
RUN python3 /tmp/patch_dcp_workspace_stride.py \
    && rm /tmp/patch_dcp_workspace_stride.py