Catalog: laguna-s earns auto placement — 4/4 on team-bench at 2bit - #134
Merged
Conversation
…bit, floors encode the wired-memory rule The user-requested laguna-s@2bit live test confirmed the serving hypothesis and passed the promotion gate: 35GB of 2bit weights fit inside the ~72%-of-RAM wired GPU limit on a 64GB Mac and score 4/4 on team-bench at 5.4 tok/s (codegen/bugfix/edit compiled + asserted, strict JSON), while 3bit (51.5GB) exceeds that limit and stalls to ~0 tok/s in the MoE expert path. laguna-s therefore returns to the head of the auto ladder with quant floors raised to keep every rung GPU-resident (2bit@64, 3/4bit@96, 6bit@192, 8bit@256): auto on a 64GB machine now provisions the bench-proven laguna-s@2bit, and no tier can resolve to a quant that would repeat the 3bit stall. nemotron-30b (1/4) and qwen3.6-35b (nvfp4 unsupported) stay explicit-only. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Follow-up to #133, completing the requested laguna-s@2bit live test — and it passed the promotion gate:
That makes laguna-s the second 4/4 model (alongside coder-32b) and confirms the serving hypothesis from #133's failure analysis: the MoE path works exactly when the weights fit wired GPU memory (~72% of RAM). 2bit = 35GB fits a 64GB Mac's ~46GB wired limit and generates cleanly; 3bit = 51.5GB exceeds it and stalls to ~0 tok/s.
Changes
/team delegate localnow provisions the bench-provenlaguna-s@2bit.laguna_floors_keep_weights_inside_wired_gpu_memoryasserts 3bit can never resolve on 64GB.Tests
1,573/1,574 across hi-agent / hi-tools / hi / hi-tui (the one failure is the documented load-flaky
dependent_task_waits_for_successin an untouched subsystem; passes in isolation).🤖 Generated with Claude Code