Skip to content

Add test manifest for PR #1410 (glm_moe_dsa)#4

Open
pcuenca wants to merge 3 commits into
mainfrom
manifest-pr-1410
Open

Add test manifest for PR #1410 (glm_moe_dsa)#4
pcuenca wants to merge 3 commits into
mainfrom
manifest-pr-1410

Conversation

@pcuenca

@pcuenca pcuenca commented Jun 17, 2026

Copy link
Copy Markdown
Owner

Test manifest for mlx-lm PR ml-explore/mlx-lm#1410 (glm_moe_dsa — GLM-5.2 DSA cross-layer indexer sharing).

cuda_reference: true (~750B; can't host the bf16 reference locally). See the manifest notes for the >2048-token requirement and the indexer-RoPE caveat about the stock-transformers reference.

pcuenca and others added 3 commits June 17, 2026 10:17
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The main result is that generation is coherent, but there's drift in the hidden states. The delta increase in the later layers could potentially be explained by different experts being selected as differences accumulate, but this is just speculation.

The transformers reference was obtained on B200x8, using "grouped_mm" experts implementation (Triton kernels) in tp mode.
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