| 1234567891011121314151617181920212223242526272829303132333435363738 |
- # syntax=docker/dockerfile:1.7
- # GLM-5.3-Flash DFlash2 draft-model support for DGX Spark (SM121/GB10), TP=2.
- #
- # Stacks on glm53-flash-nvfp4-lab:local (day-0 base + NoPE-MLA + lab
- # checkpoint patches) and adds DFlash2 draft-model execution ported from
- # local-inference-lab/vllm `dev/jovian-judgement` (see UPSTREAM.md):
- # * qwen3_dflash.py — refreshed base (fork is a superset of the day-0
- # file; identical imports, +DFlash2 wiring incl.
- # native MXFP8 serialized projections)
- # * qwen3_dflash2.py — the DFlash2DraftModel itself (294 lines)
- # * dflash.py — v1 spec-decode proposer (+rope is_neox_style
- # handoff from the target, 1 hunk)
- # * registry.py — one line: DFlash2DraftModel -> qwen3_dflash2
- #
- # The day-0 base already ships W4A16_NVFP4, MIXED_PRECISION,
- # ModelOptMxFp8LinearMethod, CompressedTensorsW8A8Mxfp8 and the dflash
- # proposer — only the DFlash2 model class + registration were missing.
- ARG BASE_IMAGE
- FROM ${BASE_IMAGE}
- COPY qwen3_dflash.py qwen3_dflash2.py dflash.py /tmp/
- RUN set -eux; \
- B=/usr/local/lib/python3.12/dist-packages/vllm; \
- cp /tmp/qwen3_dflash.py /tmp/qwen3_dflash2.py "$B/model_executor/models/"; \
- cp /tmp/dflash.py "$B/v1/spec_decode/dflash.py"; \
- grep -q '"DFlashDraftModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),' \
- "$B/model_executor/models/registry.py"; \
- sed -i '/"DFlashDraftModel": ("qwen3_dflash", "DFlashQwen3ForCausalLM"),/a\ "DFlash2DraftModel": ("qwen3_dflash2", "DFlash2Qwen3ForCausalLM"),' \
- "$B/model_executor/models/registry.py"; \
- grep -q '"DFlash2DraftModel"' "$B/model_executor/models/registry.py"; \
- python3 -m py_compile \
- "$B/model_executor/models/qwen3_dflash.py" \
- "$B/model_executor/models/qwen3_dflash2.py" \
- "$B/v1/spec_decode/dflash.py" \
- "$B/model_executor/models/registry.py"; \
- rm -f /tmp/qwen3_dflash.py /tmp/qwen3_dflash2.py /tmp/dflash.py
|
