launcher/docker-images/glm52/Dockerfile

# GLM-5.2 Int4-Int8Mix overlay on sparkrun-vllm-ds4-gb10:glm52
#
# Unsets VLLM_PREFIX_CACHE_RETENTION_INTERVAL (baked at 4096 in the base
# deepseek image, breaks GLM-5.2) via a wrapper that sources the original
# nvidia_entrypoint.sh first, then hands off with exec after the unset.
#
# Build:
#   cd docker-images/glm52 && bash build.sh
#   Tags: sparkrun-vllm-ds4-gb10:glm52

FROM aidendle94/sparkrun-vllm-ds4-gb10:production-hybrid-1.4

# 1.4 patches quantization/__init__.py to `from . import hybrid_modelopt_ct`
# but never ships hybrid_modelopt_ct.py — a build bug in the base image that
# breaks all vLLM imports with a circular-import error. Remove the dangling
# import; no recipe uses hybrid_modelopt_ct (all use hybrid_mxfp4_ct).
RUN sed -i '/from . import hybrid_modelopt_ct/d' \
    /opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/__init__.py

# Patch vLLM structured-output reasoning-boundary bug under speculative
# decoding (MTP). When the reasoning parser emits the reasoning-end marker
# () under MTP, the grammar FSM (used by response_format / guided decoding)
# failed to advance, so the content channel was unconstrained — models
# emitted prose or markdown-fence prefixes before JSON (tool-eval-bench
# Structured Output: 50% instead of 100%).
#
# Ports upstream vLLM PRs:
#   #44297 — Constrain bitmask and trim grammar advance at the boundary
#   #44993 — Advance grammar across reasoning boundary (fixes #34650,
#            #43388, #48228)
#
# Changes: should_advance() accepts new_token_ids (authoritative step delta),
# records reasoning_end_token_index, returns True for all constraint types;
# scheduler trims reasoning content before accept_tokens; grammar_bitmask
# constrains draft tokens after the marker mid-window.
COPY patch-structured-output.py /tmp/patch-structured-output.py
RUN python3 /tmp/patch-structured-output.py && rm /tmp/patch-structured-output.py

RUN pip install --no-cache-dir --no-deps --force-reinstall nvidia-nccl-cu13==2.30.4 \
 && pip install --no-cache-dir --no-deps --force-reinstall instanttensor==0.1.9

# Wrapper: unset the baked var, then hand off to the original entrypoint.
# The original nvidia_entrypoint.sh does `exec "$@"` at the end, so this
# runs as PID 1 with conda/CUDA paths sourced and the baked var gone.
RUN printf '#!/bin/bash\nunset VLLM_PREFIX_CACHE_RETENTION_INTERVAL\nexec /opt/nvidia/nvidia_entrypoint.sh "$@"\n' > /entrypoint-wrapper.sh && chmod +x /entrypoint-wrapper.sh

ENTRYPOINT ["/entrypoint-wrapper.sh"]