[bakana39]

What this changes

Reference archive only — no review or merge requested. I will close this PR immediately after submitting it.

This adds one isolated documentation/receipt directory with sanitized patches and experimental evidence from a two-GB10 GLM-5.3-Flash-FP4-Local deployment. It changes no active runtime code, defaults, build targets or support claims. Please feel free to ignore it; there is no request to reproduce, maintain or adopt these experiments.

The implementation, test orchestration, analysis and report were performed by OpenAI ChatGPT/Codex at my direction. I tested the service but did not perform a line-by-line code review. This is explicitly AI-generated experimental work, not independently reviewed production code or an OpenAI/upstream endorsement. AI disclosure is included at my request; there are no tool co-author trailers.

The initial FP4-Local loader integration is borrowed from Bizuayeu's #1050, pinned at b960da830a2722e39513f354211efe2f4531cd02, and is separately attributed and packaged. This archive does not claim that contribution, compete with the evolving loader PR, close an issue, or request an implementation claim.

Start at the reference README. The directory contains six sequential patches, reports, public synthetic fixtures, CPU comparison tools, sanitized receipts, license notices and a SHA-256 manifest. Patches target the historical pinned base, not current main.

Receipt

• Environment: two GB10s, 128 GB unified memory each, SM121a, TP=2/RoCE, CUDA compiler 13.0.88, Zig 0.17.0; historical source base f8d254c5e0f9ab91eb325a2e7c28d2e5cb07c94d. Checkpoint/drafter revisions and image/binary hashes are recorded. Detailed OS/driver inventory is not supplied; clocks were variable.
• Retained result: A4 prompt-kernel N64→N128 tiles reduced prefill time by approximately 14% on controlled 2K–66K fixtures. Compared runs had zero reused prompt tokens after kernel warmup. This is not a fresh-process, prefix-cache-disabled full upstream prefill qualification. Decode performance was essentially unchanged.
• Memory: DFlash-only loading omitted the unused MTP transformer, reducing the recorded model allocation from 94.83 to 92.78 GiB; this is not total RSS.
• Decode findings: sampled windows were about 95% GPU-busy; routed experts accounted for roughly 34–44% of sampled time and KDA projections 17–20%. These are bounded profiling windows, not complete request traces or proof of bandwidth saturation.
• Negative results included: V6 copy/tail/gather fusion, V8 low-row expert variants, and V9 KDA projection scheduling did not produce a reliable served improvement. They were not promoted. Frozen V5 remains the selected baseline.
• Exactness: V9 fixed-token 8K/32K/128K comparisons required matching recorded prompt/output hashes, sampling settings and speculative counters. All 36 new requests matched their recorded V5 reference. Hashes are retained; generated text/output IDs are not redistributed, so this is not an independent recomputation of those hashes. Wider model qualification is explicitly incomplete.
• Publication checks: all six patches replayed in order; 3,383 selected source files matched the frozen V9 snapshot byte-for-byte. Scoped diff whitespace checks passed. CPU comparison self-tests passed; 48 sanitized served records were recomputed in six comparison groups. All 68 manifest entries verified; Python sources and JSON/JSONL parsed successfully. No GPU service was restarted for publication.
• Historical tests: V5 CUDA-host subset: 1057 passed, 4 skipped. Reports distinguish completed checks from unrun checks. A single 521113-token capacity/checksum request and separate four-way approximately 41K requests do not establish four simultaneous full-context requests or general long-context quality.
• Not run / not claimed: full unified kernel-contract SUMMARY, official-backend teacher-forced fidelity, full current-main CI, Metal/non-GB10 qualification, complete zero-new-finding lean baseline, or all concurrency/cancellation/vision checks on rejected V9. Full-tree historical lean output had 439 findings, not a clean result.

Checklist / scope caveats

• One archive-only change; no active engine, deployment or model-weight change.
• Original dependency authorship and license notices preserved; no model weights or compiled binaries uploaded.
• Private addresses, host names, user paths, credentials and user conversations excluded or replaced by documentation examples.
• AI involvement and lack of human line-by-line review disclosed explicitly.
• Passed, failed, rejected and unrun experiments distinguished; no claimed decode speedup.
• Full runtime/kernel/precision/performance contribution gates: not claimed. This archive is not a candidate for merging or release qualification.

The main report explains precision boundaries, including that adding A4 prefill is not bit-equivalent to A16-only prefill. Anyone reusing a patch must review and requalify it independently. No reviewer or assignee is requested.

[bakana39]

Closing immediately as intended. This is an AI-produced reference archive only; no review, merge, reproduction or maintenance is requested. The patches and positive/negative experimental receipts remain available for anyone who independently finds them useful. The existing loader contribution is credited separately. No active runtime change is proposed.