record(minimax-h3): ALL THREE modalities work — the ref2va grid was the QUANTISATION, §8.9 REFUTED - #251
Merged
Merged
Conversation
…he QUANTISATION, §8.9 REFUTED Three campaigns fought the ref2va patch grid (#86 text-only, §8.9 vision-enriched, §8.12 load-path). §8.9 re-attributed it to the ref2va-specific reference-row conditioning ASSEMBLY, having ruled out the DiT forward and the encoder. That attribution is REFUTED. Every prior ref2va attempt ran on NVFP4. The cell nobody filled in was ref2va on the quantisation verified coherent everywhere else in this project. MiniMax-H3-REF2VA-Q4_K_M.gguf renders COHERENTLY with the IDENTICAL assembly code: period-16 seam 1.13 (clean 1.15-1.19, broken 2.28) and VAE-input latent adjacent-cell cosine 0.8526 (white 0.06, real encode 0.789, coherent 0.89+). The decoded frame is a coherent office scene with no trace of the patch grid. §8.12's quieter attribution — the per-weight NVFP4 quantisation error — was right. All three modalities are now verified on Q4_K_M with rendered artifacts: t2va (§8.16, 1344x768, ASR-verified speech), fl2va (first-frame conditioning load-bearing, keyframe->frame0 correlation 0.7881, evolving to 0.7653 by frame 100), ref2va (this row). Why it survived three investigations: the FL2VA/Q4_K_M cell was verified and the Ref2VA/NVFP4 cell was verified-broken, but PARTITION and QUANTISATION were never crossed. Each campaign held quantisation fixed at NVFP4 and searched the ref2va code path, which is where the bug was not. docs/USAGE.md now steers users to the REF2VA GGUF with a worked command and warns off the NVFP4 arm, instead of saying image-conditioned renders "are not clean even on the matching partition" — which this row disproves. Also corrects the Q8 row: Q8_0/Q6_K/Q5_0 ARE published (unsloth/MiniMax-H3-GGUF) but PRUNED only, and the restructured AdaLN is not drop-in for this loader. The previous "no published GGUF is known" was wrong. Records + docs only; no code change — the ref2va path needed none. FOLLOWING_AGENTS_PROTOCOL Assisted-by: Claude Code:claude-opus-5 [ClaudeCode]
localai-bot
pushed a commit
that referenced
this pull request
Aug 10, 2026
…sweep summary catches up (#243, #251) Two stale records, landed together because the second is only landable with the first. MiniMax-H3. 1c1749c (#251) established that ref2va renders COHERENTLY from MiniMax-H3-REF2VA-Q4_K_M.gguf with the IDENTICAL assembly code: period-16 seam 1.13 against a 1.15-1.19 clean band and 2.28 broken, VAE-input latent adjacent- cell cosine 0.8526. So the patch grid that three campaigns chased was the NVFP4 per-weight quantisation error, §8.12's quieter attribution, and §8.9's re-attribution to the ref2va conditioning ASSEMBLY is REFUTED. FEATURES still told the opposite story: "ref2va ckpt-limited (§8.12)" and "ref2va NVFP4 render blocked on checkpoint quant (needs official modelopt NVFP4)". That reads backwards now. The NVFP4 arm is the broken one and Q4_K_M is the arm that works, which is the reverse of what a reader would have done with the old text. Both H3 rows now say all three modalities render coherent on Q4_K_M, anchored at §8.20, and the speed cell steers to the Q4_K_M GGUF. feature-matrix.md gap summary. It still named xgrammar and reasoning parsers among the "top HIGH misses" after 3fd1fc1 closed both. This edit was written during #243 and DROPPED there: check-doc-checkpoint classifies feature-matrix.md as a feature_surface file demanding a matching docs/FEATURES.md update, that change moved no capability, and editing FEATURES to satisfy a gate would have manufactured a claim. Here FEATURES genuinely changes for the H3 reason above, so the gate's requirement is met on its own terms rather than worked around. The summary now carries the reconciled counts and the post-reconcile HIGH list. Gates: agent-preflight.sh --staged, All gates green. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Claude-Code:claude-opus-5 [ClaudeCode]
localai-bot
pushed a commit
that referenced
this pull request
Aug 10, 2026
…ies; gap summary catches up (#243, #251) Authorized local merge (developer-directed, 2026-08-10). Cherry-picked onto current main rather than forward-merged, as with #242 and #243, to avoid inheriting an unrelated red from check-role-discipline (see #250). Closes the feature-matrix.md residual named in #243: the edit is landable here because FEATURES genuinely changes for the MiniMax-H3 reason, so check-doc-checkpoint's feature_surface requirement is satisfied on its own terms instead of being worked around. FOLLOWING_AGENTS_PROTOCOL Following-Agents-Protocol: true AI-Assisted: true Assisted-by: Claude-Code:claude-opus-5 [ClaudeCode]
This was referenced Aug 10, 2026
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
2 participants
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.
The result
ref2va works. Three campaigns fought its patch grid (#86, §8.9, §8.12). §8.9 re-attributed it to the ref2va-specific reference-row conditioning assembly, having ruled out the DiT forward and the encoder path. That attribution is refuted.
Every prior ref2va attempt ran on NVFP4. The cell nobody filled in was ref2va on the quantisation verified coherent everywhere else in this project.
Scales: seam 1.15-1.19 clean / 2.28 broken; cosine 0.06 white / 0.789 real encode / 0.89+ coherent. Same assembly code, same packed-block layout, same denoise loop — only the checkpoint changed. §8.12's quieter attribution (the per-weight NVFP4 quantisation error) was the correct one.
All three modalities verified on Q4_K_M
Why it survived three investigations
The FL2VA/Q4_K_M cell was verified good and the Ref2VA/NVFP4 cell verified broken — but partition and quantisation were never crossed. Each campaign held quantisation fixed at NVFP4 and searched the ref2va code path, which is where the bug was not.
Generalises to: when a mode fails on exactly one checkpoint, vary the checkpoint before searching the mode's code.
Docs
docs/USAGE.mdnow points atMiniMax-H3-REF2VA-Q4_K_M.ggufwith a worked command, and warns off the NVFP4 Ref2VA arm — replacing the claim that image-conditioned renders "are not clean even on the matching partition", which this row disproves.Risk
Records and docs only. No code change — the ref2va path needed none.