Dom/rpa_vision_v3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs(dette): création DETTE-010 vérif preprocessor_config Qwen3-VL

Browse Source 
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Dom
2026-05-09 12:08:08 +02:00
parent e0b47e4518
commit e9702b4df9
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
docs/DETTE_TECHNIQUE.md
View File

@@ -25,6 +25,7 @@ P0 / P1 / P2 / P3 (alignées sur convention handoffs)
| DETTE-007 | 2026-05-09 | 2026-05-23 | P3 | OPEN | Trois implémentations smart_resize coexistent (server.py, infigui_worker.py, nouveau module officiel). Unification post-démo Kerella. | commit feat(grounding): module smart_resize officiel |
| DETTE-008 | 2026-05-09 | 2026-05-23 | P2 | OPEN | Pre-check VLM par-clic désactivé via `if False:` (observe_reason_act.py:1704-1713) | docs/AUDIT_CONTROLES_DEBRANCHES_2026-05-08.md (F6.1.1) |
| DETTE-009 | 2026-05-09 | 2026-05-23 | P3 | OPEN | `ShadowLearningHook` (core/grounding/shadow_learning_hook.py) défini mais jamais instancié — Phase 6 du pipeline FAST→SMART→THINK non câblée à l'observation Shadow | docs/INVESTIGATION_MEMOIRE_VISUELLE_ORPHELINE_2026-05-09.md |
| DETTE-010 | 2026-05-09 | 2026-05-09 | P1 | OPEN | Vérifier que le preprocessor_config.json du checkpoint Qwen3-VL-8B effectivement déployé via vLLM utilise bien max_pixels = 1_003_520. Si override dans le checkpoint, ajuster MAX_PIXELS_DEFAULT du module ou passer en paramètre explicite côté call sites. Bloquant pour l'Étape 2 (validation grounding isolée). | docs/MIGRATION_VLM_PLAN_2026-05-09.md + commit feat(grounding): module smart_resize officiel |
## Convention de référencement