Repository navigation
docs(bench): reclassify GB10 0.3.1 results; minicpm-v-mxfp4 passes, drop qwen2.5-vl-3b - #335
Merged
Merged
Conversation
…rop qwen2.5-vl-3b Follows the loader fix (#334). Updates the GB10 0.3.1 benchmark CSVs and docs: - minicpm-v-4.6-mxfp4: now passes after #334. Text 347.59 prefill / 142.14 decode tok/s, image input 625.47 / 138.70 (48 tok). Added to the image-input table and flipped to a pass in the text table. - qwen2.5-vl-3b: removed. It was a broken duplicate of qwen2.5-vl-3b-4bit (the model dir was deleted); the working 4bit variant stays. Both CSVs go 148 -> 147 rows. - Reclassified six by-design non-failures from fail (red x) to N.A. (white circle): docling-layout-heron and granite-speech (not standalone text generators) and the four MTP/DFlash drafters (gemma-4-12b/31b assistant, qwen3.5-27b/4b-dflash) that need a target model. The raw CSV FAIL:bench markers stay (they genuinely produce no text-gen output); the N.A. tag is a doc-layer classification. GB10 0.3.1 summary is now 147 total: 136 pass, 0 code failures, 9 N.A., 2 too-large (qwen3-next-480b, deepseek-v3, shown red x, OOM), 54 VLM image rows. Legend and the by-cause section were rewritten to match; index counts and the benchmark-CSV table updated in step.
inureyes
added
status:review
5 of 11 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Follows the loader fix (#334). Brings the GB10 0.3.1 benchmark CSVs and docs in line with the now-resolved results.
Changes
qwen2.5-vl-3b-4bit(the model dir was deleted upstream); the working 4bit variant stays. Both CSVs go 148 → 147 rows.docling-layout-heron-mlx-bf16andgranite-speech-4.1-2b-nar-mlx(not standalone text generators), and the four MTP/DFlash draftersgemma-4-12b-it-assistant-4bit,gemma-4-31b-it-assistant-bf16,qwen3.5-27b-dflash,qwen3.5-4b-dflash(need a target model). The raw CSVFAIL:benchmarkers are unchanged (they genuinely emit no text-gen output); ⚪ is a doc-layer classification, matching how paligemma2 is already handled.GB10 0.3.1 summary (now)
136 + 0 + 9 + 2 = 147. Verified against the table rows (text-family 136 ✅ / 2 ❌ / 9 ⚪; image table 54 ✅). Legend and the by-cause section rewritten to match; index counts and the benchmark-CSV table updated in step.