docs(readme): TurboKV precision known issue on M5 (not reproduced on M6) - #174
Merged
Merged
Conversation
Reproduced by the M5 session on main 318f712 with Qwen3.8-27B-4bit --vision at temperature 0. Prompt: n numbered lines ("Line {i}: the quick brown fox jumps over the lazy dog."), then "How many lines are there? Answer briefly." With --turbo-kv on M5: 125 lines → "125", 315 → "1,000", 500 → "14", 600 → "1,000", 700 → "1,000" (all repeated 2–3×). Without --turbo-kv it answers 315 / 500 / 700 correctly. On the M6 the same prompts are 24/24 correct in both modes, 2K–11.8K tokens. Points to a GPU-family-dependent TurboKV kernel path. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eam-experts - TurboKV-on-M5 note → tracking issue #175. - --gpu-layers GPU timeout: reproduced on M5 as well as M6; tracked in #176 with the fix in SharpAI/mlx-swift#17. CPU-resident MoE layers stay very slow after the fix (~0.4 tok/s prefill), so the note points 32 GB users at --stream-experts first. Co-Authored-By: Claude Opus 5.5 <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
README-only follow-up to #173. It adds a known-issue note for a
--turbo-kvprecision bug that shows up on the Apple M5 but not on the M6.What happens
Qwen3.8-27B-4bit (VLM path), temperature 0. The prompt is
nnumbered lines (Line {i}: the quick brown fox jumps over the lazy dog.), then "How many lines are there? Answer briefly." The model only has to read the last line's index, so the test measures exact long-range recall.--turbo-kv--turbo-kvM5 data: the Mac M5 session, main
318f712. M6 data: this session, main318f712+ #173. Identical prompt strings and token counts. On the M5 the error starts somewhere between ~2K and ~5K tokens. The likely cause is a GPU-family-dependent path in TurboKV's dequant or attention kernels. The note tells M5 users to avoid--turbo-kvwhen exact recall matters. The Mac M5 session will file the tracking issue.🤖 Generated with Claude Code