# Upstream provenance — vendored DFlash2 sources

All three `.py` files are copied verbatim (no local edits) from

- repo:   https://github.com/local-inference-lab/vllm
- branch: `dev/jovian-judgement`
- commit: `da4d7be6c97434f6942292ed8abbf4b32dc44355` (2026-08-28,
  "feat(dflash2): execute serialized MXFP8 projections natively")
- license: Apache-2.0 (SPDX headers retained)

| local file     | upstream path                                       |
|----------------|-----------------------------------------------------|
| qwen3_dflash.py  | vllm/model_executor/models/qwen3_dflash.py        |
| qwen3_dflash2.py | vllm/model_executor/models/qwen3_dflash2.py       |
| dflash.py        | vllm/v1/spec_decode/dflash.py                     |

Why vendored instead of "just use the fork image": every published
`local-inference-lab` / `voipmonitor` vLLM image is **amd64-only** (RTX PRO
6000 TP=4 world). The day-0 arm64 cu130 base our lab image uses is already a
superset of the fork for everything EXCEPT the DFlash2 model class — verified
by import-level diffing: the fork files import only symbols the day-0 image
ships (`ModelOptMxFp8LinearMethod`, `CompressedTensorsW8A8Mxfp8`, the dflash
proposer machinery). So the whole port is these 3 files + one registry line.

Refresh: re-fetch the same three paths from the branch tip, `diff` against
the vendored copies, rebuild.