Files
club-3090/scripts/lib
noonghunnaandClaude Opus 4.8 0cfd099b37 dgemma: adopt official vllm/vllm-openai:gemma image (drop 120-file sideload)
vLLM publishes an OFFICIAL `vllm/vllm-openai:gemma` image (pushed 2026-06-10, a
stock build of the dgemma branch commit 74b5964f) with DiffusionGemma baked in —
`DiffusionGemmaForBlockDiffusion` registers natively, transformers 5.10.2. So we
pin that image (BY DIGEST, purge-resistant) and drop the bespoke sideload from
PR #358 (stock nightly + 123-file branch overlay + install_script).

3 fixes are NOT upstream (vLLM tests H100/B200 + TP=1) so they're not in :gemma —
the compose now bind-mounts them (site_package_overlay) from the new lean dir
models/diffusiongemma-26b-a4b/vllm/patches/gemma-image-fixes/:
  - marlin.py + marlin_utils_fp8.py — sm_86 fp8 Marlin sub-tile-K pad. :gemma
    clean dies in warmup ("Invalid thread config ... num_bits=8 ...
    max_shared_mem=101376", K=352/1056) without it.
  - diffusion_gemma.py — TP-vocab soft-embed + dtype fix (TP=2; their recipe is TP=1).

Changes:
  - base.yml: image -> :gemma@sha256:9c719fc0...; default `vllm serve` entrypoint
    + 3 file mounts (was: overlay-dir mount + install_script bash entrypoint).
    Status 🧪 experimental (was upstream-gated; supersedes PR #359 too).
  - engine vllm-diffusion-gemma: install.spec -> :gemma@digest; vendored_overlays
    -> the 3 fixes (delivery site_package_overlay).
  - registry status -> experimental; note rewritten.
  - patches.yml: dgemma-gemma-image-fixes (site_package_overlay, 3 overlay_files).
  - diagnose_profile_cli OVERLAY_PATH_HINTS + docs/UPSTREAM.md #45163 row.
  - DELETE the 123-file dgemma-overlay/ + the regeneration Dockerfile.

Validated live on 2x RTX 3090 (2026-06-11): :gemma clean dies on the Marlin wall;
:gemma + 3 mounts (via `docker compose -f base.yml up`) boots, serves coherent
output, 262K, ~177/180 TPS typical / ~1100 peak, 23.1 GB/card. verify-full: gen +
tool-call + reasoning + output-quality pass (streaming-SSE "1 chunk" is the
expected block-diffusion artifact). Full gate green (test-compose-registry-disk
local-only red = untracked nex-n2-mini WIP; CI-clean validated: disk 46 / reg 45).

Co-Authored-By: Claude Opus 4.8 <[email protected]>
2026-06-11 05:31:36 +00:00
..