Skip to content

[II] Serve projection-mixed EXL3 K3/K4/K5 weights - #148

Closed
voipmonitor wants to merge 2 commits into
masterfrom
feat/ii-r7-mixed-trellis-k345-20260812
Closed

voipmonitor wants to merge 2 commits into
masterfrom
feat/ii-r7-mixed-trellis-k345-20260812

Conversation

@voipmonitor

Copy link
Copy Markdown
Contributor

Status

  • Implementation: implemented
  • Unit and kernel qualification: qualified
  • GLM-5.2 checkpoint E2E qualification: pending

Purpose

Serve EXL3 checkpoints that assign MCG K3, K4, and K5 independently to routed-expert gate, up, and down projections. The B12X runtime reads packed checkpoint storage directly and executes all projection tiers through one CUDA-graph-capturable mixed-Trellis launch plan.

Resulting behavior

  • Represent tier membership per expert and per projection with exact gate, up, and down bounds.
  • Validate descriptor counts, packed extents, route identifiers, and projection counts before launch.
  • Keep expert counts as runtime data so layer-dependent partitions reuse compiled kernels.
  • Prewarm scheduler capacity, rank-sliced draft, route-identifier, and scalar-alignment specializations before graph capture.
  • Preserve graph-stable scratch ownership for decode and prefill.
  • Preserve QSRT atoms-v2 compact-offset pair modes and the qualified FC2 32x512 tile contract.

Compatibility

The MCG codebook accepts K3, K4, and K5 for this mixed runtime. SQG-XOR-Cheb-T12 retains its K2/K3/K4 contract. Existing QSRT paths and the removed route-major W4A8 interface are unchanged. Malformed or incomplete projection maps fail closed.

The vLLM checkpoint loader is provided by a companion pull request against dev/infernal-invocation.

Validation

Source base: ac07f6746edce4852f01cc11f34dbe7abae1fdaa.

  • 58 host-runnable focused tests passed; 40 GPU-only cases skipped.
  • 6 SM120 tests passed for K3/K4/K5 numerical parity, projection-tight payloads, CUDA graph replay, K6 small-M dispatch, one-grid prefill planning, and GLM-5.2 large-M execution.
  • Ruff check and format: passed.
  • git diff --check: passed.

Implement one graph-capturable mixed-Trellis launch plan for EXL3 checkpoints that assign MCG K3, K4, and K5 independently to gate, up, and down projections. Validate descriptor counts and packed extents before launch, retain graph-stable scratch ownership, and prewarm route, rank-slice, and scalar-alignment specializations. Preserve the current QSRT atoms-v2 pair contract and the removal of the route-major W4A8 interface.
@coderabbitai

coderabbitai Bot commented Aug 12, 2026

Copy link
Copy Markdown
Contributor

Important

Review skipped

Draft detected.

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: 3fd7a793-4955-485b-b42e-a85dbb114b09

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

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

Remove the duplicate 32x512 FC2 tile exception from the mixed-Trellis change. Pull request #146 owns tile validation and FC2 scoping; applying #146 before this branch provides the qualified geometry without overlapping kernel policy.
@lukealonso

Copy link
Copy Markdown
Collaborator

Closing without merge because this branch is incompatible with the master execution contract. It adds an entirely new projection-mixed Trellis kernel path, has unresolved conflicts in the shared W4A16 kernels, and adds dense K6 dispatch through b12x.gemm.trellis_linear._small_m, which is absent from master. The branch also depends on unmerged PR #146 and does not include checkpoint-level end-to-end qualification.

The host-only route-pack specialization prewarm idea is separable and should be proposed independently from master with focused capture evidence.

@lukealonso lukealonso closed this Aug 16, 2026
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.

2 participants