Skip to content

CI builds Vulkan files from the XNNPACK run's GPTQ codes; the 2026-11-09 RTN-to-GPTQ rollout and the checks that proved it - #29

Merged
alpharomercoma merged 3 commits into
mainfrom
gptq-rollout
Nov 9, 2026
Merged

alpharomercoma merged 3 commits into
mainfrom
gptq-rollout

Conversation

@alpharomercoma

@alpharomercoma alpharomercoma commented Nov 9, 2026 •

Copy link
Copy Markdown
Member

Why: export-vulkan.yml never passed --codes, so every Vulkan file CI built was rounded to nearest. On 2026-11-09 every such file on the Hub was replaced by hand. This PR makes CI build them the same way, so new models stop getting round-to-nearest GPU files.

CI: each window's Vulkan file is built in its XNNPACK job

  • Workflow: export-xnnpack.yml publishes the window's gated XNNPACK file. It then builds the Vulkan file in its own Modal Sandbox from the same codes.pt, with that report as --gate-proxy, and publishes it once its twin is on the Hub.
    • A GPU failure or timeout cannot cost the CPU file, but it fails the job.
    • Only GPTQ 8da4w runs of models Vulkan can take get a GPU file (export-matrix --with-vulkan).
  • export-vulkan.yml is deleted. A workflow of its own could only solve again, giving the CPU and GPU files different codes, or round to nearest.
  • Watcher: there's no Vulkan stage now. watch.BUILT_WITH marks Vulkan dispatched with the XNNPACK run.
    • It records Vulkan on the 26 state entries that predate it.
    • It skips a Vulkan line that its host run will never build.
  • Publishing keeps every GPU file beside the gate it cites:
    • An XNNPACK commit withdraws a Vulkan window whose proxy names a twin that commit removes or replaces, judged on the tree the commit produces.
    • A Vulkan publish requires that twin at the head it commits onto, re-checked on every retry.
    • GitHub releases follow the Hub.
    • --allow-unpublished-twin is the explicit way past these for a twin deliberately not published (SmolLM2-360M).

Tools and docs

  • tools/codecheck/: reads a .pte's int4 codes back and proves them against codes.pt or the fp32 source, plus the Hub verifier. Standalone, not wired into CI.
  • docs/PLAN.md: what is published now.
    • No round-to-nearest CPU or GPU file remains.
    • Qwen3-0.6B ships fp32 CPU files and has no Vulkan folder.
    • SmolLM2-360M's GPU files are gated through an unpublished twin.
    • Qwen3 gates from before 2026-10-31 measured nothing.
  • Research findings 43–45:
    • 43: Qwen3-0.6B loses its tool decision at int4.
    • 44: solving and gate references are unreliable on macOS arm64.
    • 45: how int4 codes read back out of .pte files.

Verification

  • Tests: ruff and pytest (327) pass.
  • Modal scripts: both, extracted verbatim from the workflow, ran on SmolLM2-135M at 2k with real codes. The XNNPACK twin passed its gate (4/4, 0.585 against 0.583), and the Vulkan file was built with a proxy naming the twin by sha256. Both passed require_publishable and require_vouched, and both carry the codes exactly (211/211, in execution order).
  • Codex: four adversarial review rounds; every high-severity finding is fixed. The remaining release-layer limits are written down in PLAN.

Not exercised: a live GitHub Actions run on Modal. The first model the watcher dispatches after merge will be one; export-xnnpack.yml on a small model with publish_hf: false is a cheap way to try it first.

🤖 Generated with Claude Code


Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need. Autofix is enabled.

alpharomercoma and others added 3 commits November 9, 2026 19:09
…s to be shown to carry the codes its report names

ptecodes reads every int4 linear out of an ExecuTorch 1.5.1 file in execution order: XNNPACK
qbint4 filters through each fully-connected node (delegate calls in program order), Vulkan
linear_dq8ca_q4gsw weights, scales and per-group code sums. codesmatch requires a file to carry
exactly a solve's codes.pt (codes, scales and order); check classifies a file's codes as rounded
or solved against its fp32 source with no codes needed; scan and make_tasks run check over an
org; verify_hub ties each published file to its byte check, codes hash, gate reference and gate
proxy. Every check of the 2026-11-09 GPTQ rollout ran through these. Standalone for now: not
wired into the pipeline or CI.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…codes, because a Vulkan workflow of its own could only solve again or round to nearest

export-vulkan.yml never passed --codes, so every Vulkan file CI built was rounded to nearest.
It is gone. export-xnnpack.yml now publishes each window's gated XNNPACK file, then builds the
Vulkan file in a Modal Sandbox of its own from the same codes.pt, with that report as
--gate-proxy, and publishes it once its twin is on the Hub. A GPU failure or timeout cannot
cost the CPU file, but it fails the job. Only GPTQ 8da4w runs of models Vulkan can take get
one (export-matrix --with-vulkan).

The watcher has no Vulkan stage: watch.BUILT_WITH marks Vulkan dispatched with the XNNPACK
run, records it on entries that predate it, and skips a Vulkan line its host run will never
build.

Publishing keeps every GPU file beside the gate it cites. An XNNPACK commit withdraws a Vulkan
window whose proxy names a twin that commit removes or replaces, judged on the tree the commit
produces. A Vulkan publish requires that twin at the head it commits onto, on every retry.
GitHub releases follow the Hub. --allow-unpublished-twin is the explicit way past these for a
twin deliberately not published (SmolLM2-360M).

Four Codex review rounds. Both Modal scripts, extracted verbatim, ran on SmolLM2-135M at 2k
with real codes: the twin was gated, the Vulkan file built with a proxy naming it, and both
passed publish's checks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ates before 2026-10-31 measured nothing

The 2026-11-09 rollout rebuilt the 13 repos with Vulkan folders on EC2: twins gated, every
file byte-checked against its codes, and every Hub file tied to that evidence. An independent
scan matched all 165 CPU and GPU files in the org to their reports. Decided exceptions:
- Qwen3-0.6B's XNNPACK files are fp32 linears and its Vulkan folder is withdrawn.
- SmolLM2-360M's Vulkan files are gated through an unpublished twin.

Research findings:
- 43: Qwen3-0.6B loses its tool decision at int4 whether rounded or solved; fp32 linears
  keep it.
- 44: solving on macOS arm64 turns hidden states into NaN, and the Mac's fp32 gate reference
  is off as well.
- 45: how int4 codes read back out of XNNPACK and Vulkan files, and what separates rounded
  codes from solved ones.

The Vulkan section describes the CI change of the previous commit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@alpharomercoma
alpharomercoma merged commit 4278711 into main Nov 9, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant