cuda: the staged NVFP4 prompt entries leave an expert's items alone (skip), as the decode form does

[Bizuayeu]

What this changes

A family that computes one expert itself (GLM-5.3-Flash stacks its BF16 shared expert as the last expert id) plans
every pair together and has the grouped kernels skip that id. The decode form already takes skip;
Grouped.prompt refused it with OptionNotBuilt for NVFP4. The prompt entries now take skip and return before
staging any rows of that expert's items. Nemotron passes -1, so its bits do not change.

Receipt

• Environment: TensorFold 1.0.10 (6c5e5192) plus this change; one GB10 (DGX Spark, sm_121, CUDA 13.3).
• Exactness: with skip = -1 the kernels run as before. tf-cuda-test quant-ref now runs the prompt gate-up and
  down twice on one routed plan (300 rows, 4 slots, the shared expert last), without and with the shared expert
  skipped: its pairs' rows keep what the buffer held, and every other pair's row equals the run without skip, byte
  for byte, for fp8 and nvfp4.
• Decode and prompt speed: not applicable (one compare at an item's start).
• Tests run:
  • tf-cuda-test quant-ref: 4 PASS, 0 FAIL; the nvfp4 skip check fails on main with OptionNotBuilt.
  • zig build test-cuda-host: 451 passed, 0 failed. zig build test: 553 passed, 5 skipped, 0 failed.
  • zig fmt --check: clean. lean_check: 523, as main.
• Not run, and why: Metal (CUDA only).

Checklist

• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off.
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo.
• No precision traded for speed. If the bits change, the description says which sums change and why.
• Prompt processing is no slower than the last release.
• The description says which platforms I ran, Metal M1 to M5 and CUDA, and which I could not.
• New tests fail before the change, pass after it, and skip cleanly without their dependency.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

[Bizuayeu]

Rebased onto 1.0.9 (c36b536d), the commit's lines unchanged. Host tests 429/429, zig build test 484 passed, 5 skipped.

[Bizuayeu]

Rebased onto 1.0.10 (3a576e32). #805 split the prompt rows into prompt_tiles and the prompt_body that dispatches to it, so the skip check now sits in prompt_body, before any tile is staged; the changes to experts.zig and the quant-ref test are unchanged. quant-ref 4 PASS, and the description's counts are from this head.

[ashhart]

Thank you, @Bizuayeu. Letting the NVFP4 prompt entries skip an expert, as the decode entries already do, removes a real blocker for a family that computes its shared expert itself.

It ships in TensorFold 1.0.12 under your name (a90017a14), with one commit of ours after it (b50341f6c) that wraps quant-ref's new lines at 120 columns. Your head is merged after them with no change to the tree, so this pull request shows as merged.

On a GB10, tf-cuda-test quant-ref passes the skip case for FP8 and NVFP4: the skipped expert's prompt rows keep what the buffer held, and every other row equals the run without skip. Nemotron passes -1 on every call, so its NVFP4 replies are the same tokens: drafted equals "draft": false in all 17 release cells, chats equal the cache-off server's, and 2, 4 and 8 concurrent streams equal their solo runs. Cold prompts of 2k to 64k tokens ran within 0.3% of the build before.

No caller passes a real expert id yet, so quant-ref alone covers the skip path until a family such as GLM-5.3-Flash on CUDA starts passing its shared expert.

Posted by Ash's AI Assistant Synthena

[Bizuayeu]

@ashhart, for the record: GLM-5.3-Flash on CUDA already passes a real id here. cuda_moe.zig passes its BF16 shared expert's id (l.count) as skip to both the gate_up and down entries, on prompt chunks and on decode, so #804's lane checks and served runs go through the skip path with a real expert.