repro(glm53): publish qualified W4A4 DCP4 MTP5 source snapshot

[jackzampolin]

Purpose

Publish the exact vLLM source snapshot used by the qualified
ghcr.io/jackzampolin/glm53-flash-nvfp4-jovian community experimental image so
the runtime is reproducible and its branch/PR composition is explicit.

This is a source/reproduction snapshot, not a request to merge this combined
diff as-is. The individual changes should continue upstream feature-by-feature.

Exact source and artifact

• source branch: jackzampolin:repro/glm53-flash-nvfp4-w4a4-dcp4-mtp5-20260828
• source commit: 5c1702c572abad8672f66b6d5eab398583532ac5
• source tree: 5177283cf2b757bb75bc65b824cc0766b2324ec3
• recorded Jovian branch point: 3e63556c2b79de0b3a2f04648e575cb06e8ad276
• qualified OCI index: sha256:6a89cd18c1dc38753d7eb9f18dbcb3e937c28a30675cee61f1a1d0dc662a4ecc
• exact Linux/amd64 image manifest: sha256:6ebeaa39d468250e661a43e02201e7a01bc6aa420e3416aca5b47e545081d202
• pinned parent image: sha256:60f6e0c3e8f7f299b6f73374cc51c43e61180f53ec6c924dbd1f1a1d0024decb
• model revision: 520de24eabf507659eaef7c70f14fd584527facc

The prior image index sha256:eb9f5b9b... is superseded: its labels named this
commit, but its embedded source omitted the final CUDA-graph profiling
implementation and test. The replacement above was rebuilt directly from this
PR fetch ref and requalified.

PR composition

Included:

• feat(glm5next): support B12X C4 DCP #486 — merged GLM5Next B12X C4 DCP/MTP foundation.
• perf(glm5next): complete B12X CKV support for DCP and MTP #488 — CKV/DCP/MTP/prefix-cache work; its final workspace and preallocation
  commits are patch-equivalent to the snapshot.
• perf(glm5next): use direct single-token KDA decode #490 — direct single-token KDA decode, patch-equivalent to the snapshot.
• [II] Align hybrid recurrent checkpoints with cache-group boundaries #457 — recurrent checkpoint cadence fix, harvested from Infernal Invocation.
• perf(spec-decode): bound autoregressive draft prefill graph memory #502 lineage — earlier 746600d3 bounded draft-prefill implementation; not
  patch-identical to the current perf(spec-decode): bound autoregressive draft prefill graph memory #502 head.
• [Model] add GLM-5.3-Flash support vllm-project/vllm#53906 GLM-5.3 support lineage at recorded head
  142062f13d16bed254b5d97cc3d371fbd4f7790a.
• feat(gdn): add direct single-token KDA decode b12x#249 direct KDA lineage inherited from the parent.

Not included:

• feat(dflash2): execute serialized MXFP8 projections natively #499. This qualified profile uses MTP5, not DFlash2.

The remaining commits after the PR-derived work are the tested production
harvest: caller-owned KDA scratch, parallel C4 prefill, mHC dispatch, stable KDA
backend selection, CUDA-graph resource retention/accounting, disabled-capture
profiling policy, bounded CKV workspace, and pre-KV-profiling B12X scratch.

Qualified runtime profile

4x RTX PRO 6000 Blackwell at stock clocks; NVFP4 W4A4 target experts through
Humming; TP4/DCP4; FP8 KV; MTP5 with Humming MXFP8; prefix caching/CKV; B12X
attention and linear; full decode plus piecewise prefill CUDA graphs; 524288
model length; 32 sequences; 4096 batched tokens; 0.88 GPU memory utilization.

Clean-rebuild validation:

• five Estonia c30/high runs: 29, 28, 29, 29, 30; 145/150 pass (96.7%), zero
  runtime errors and zero max-token truncations
• boot: 10,835,285 usable KV tokens and 9.62 GiB graph pool
• matched prefill at 8K/64K/128K: -1.1%, +1.2%, -0.3% versus prior runtime
• sustained C8: +5.4%/+6.1% at zero/8K context
• sustained C32: +0.3%/-0.3% at zero/8K context
• burst: 410/410 completed, zero errors

Full build/run instructions, component locks, commands, and result tables:
https://github.com/jackzampolin/glm53-flash-nvfp4-jovian

AI assistance disclosure: Codex assembled, tested, and documented this
integration snapshot. Human review remains required for individual upstream
changes.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}