[EXL3] preserve mixed-prefill numerics with paired M8 FC2

[yatesdr]

What

Preserves block-64 mixed-K3/K4 prefill route packing and bounded workspace
while restoring deterministic block-8 FC2 arithmetic inside the same
cooperative grid.

The kernel maps each FC2 compute block to its block-64 metadata parent,
computes exact M8 subtiles, and reuses each decoded weight fragment across a
pair of independent M8 outputs. The mixed-Trellis ABI is bumped so stale
cubins cannot be reused across the launch-contract change.

Why

Captured-input checks localized the mixed-prefill quality regression to FC2:
FC1 and activation outputs were bit exact, while the original block-64 FC2
path changed reduction order and was nondeterministic at the production tile
geometry. On the prescribed 2,047-position KLD gate, that path measured
0.07888687230031491 versus the 0.07776700789368204 reference.

The paired-M8 implementation restores reference arithmetic without returning
to serial homogeneous K3/K4 launches or unbounded workspace.

Validation

Component-specific gates:

• clean branch files are byte-identical to the qualified Gilded Gnosis r19
  overlay;
• production-shape operator proof: FC1, activation, FC2, and final output bit
  exact, with zero changed elements;
• paired-M8 operator time: 5,162.58 us versus 8,873.07 us for the naive exact
  M8 implementation (1.7187x);
• focused mixed-kernel capture/mapping tests passed;
• prescribed r17 KLD: 0.07776700789368204 over 2,047 positions, exact
  reference parity.

The combined r19 qualification (TR3-3.36, TP4/DCP4, dynamic NVFP4 MLA KV,
exact selector, i8_hier) measured:

• one-grid prefill planned at the production 3,072-row capacity;
• 609,024 GPU KV tokens at MTP0/480k;
• 38.9868 tok/s at MTP0/C1;
• 1,705 server tok/s on randomized cold 64k prefill after one-time JIT,
  cached_tokens=0;
• cold 350k retrieval returned exact 738216, cached_tokens=0.

No separate r19 KLD run is claimed; r19 re-ran the production-width exact
operator and long-context behavioral gates against source bytes identical to
the KLD-qualified implementation.

Dependency

This PR is stacked on #110 so its review diff
contains only the paired-M8 exact-FC2 implementation. Retarget to master
after #110 merges.

Companion vLLM planner PR:
local-inference-lab/vllm#226

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 7ed9a439-ef54-4245-b578-2c36e9649dd7

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

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.}

[voipmonitor]

Superseded by #112. The successor preserves the paired-M8 numerical contract without the stacked feature-branch base and is the PR used by the r20 release manifest.