Go to file
Kazeia Team 0a72654093 refonte(llm): moteur LLM unifié GGUF + .pte NPU (UnifiedLlmAdapter)
Remplace la couche LLM par la dernière version de kazeia-engine
(cf dist/LLM_INTEGRATION.md). Chargement par magic-byte :
- GGUF ("GGUF"@0)  -> EngineLlmEngine (libllama CPU i8mm, session cache-préfixe KV)
- .pte  ("ET12"@4) -> PteLlmEngine (ExecuTorch+QNN, NPU V79, sans root)

- LlmLoader.kt copié de l'engine + fix encode DJL (encode(p,false,false))
- UnifiedLlmAdapter : pont com.kazeia.core.LlmEngine, generateWithSystem par tour
- ModelRegistry : 5 modèles (Qwen3.5-4B GGUF défaut + 4 .pte en sous-dossiers)
- KazeiaService recâblé ; DEFAULT_SYSTEM_PROMPT -> ConfigStore.DEFAULT_SPEAKER_PROMPT
- DJL tokenizer arm64 Android : tokenizers + tokenizer-native alignés 0.33.0
  (natif libdjl_tokenizer.so dans l'AAR ; natifs desktop exclus, -23 Mo APK)
- jniLibs : +libkazeia_pte.so, +libQnnHtpNetRunExtensions.so,
  libqnn_executorch_backend.so -> QAIRT 2.42 ; manifest re-figé (28 libs)
- fichiers morts archivés (ExecuTorchLlmEngine, GenieLlmEngine)

Phase 1 (#287) : build debug green. Phase 2 = validation device.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-17 22:47:32 +02:00
docs feat(rag): corpus psychoéducation FR étendu à 18 fiches (corpus-seed-r2) 2026-06-11 22:49:36 +02:00
executorch-custom TTS tremor investigation: identify cross-arch numerical floor, gate diag flags 2026-04-13 00:15:14 +02:00
executorch-patches LLM: enable hybrid-mode export via num_sharding=1 — TTFT 2.9s → 113ms 2026-04-14 15:08:31 +02:00
kazeia-android refonte(llm): moteur LLM unifié GGUF + .pte NPU (UnifiedLlmAdapter) 2026-06-17 22:47:32 +02:00
keystore build: versionner le keystore release dans le dépôt privé (sauvegarde) 2026-06-11 08:37:34 +02:00
scripts chore: dégraissage du dépôt — suivi restreint à l'app mobile (liste blanche) 2026-06-09 21:21:00 +02:00
.gitignore docs: RELEASE_SIGNING à jour — keystore versionné dans le dépôt privé comme sauvegarde 2026-06-11 08:38:16 +02:00
AI_HUB_QUALCOMM.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
ARCHITECTURE_PIPELINE.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
AVATAR_3D_RAPPORT.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
BENCHMARK_RAPPORT.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
BENCHMARK_ROOT_VS_NONROOT.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
DEPLOY_EXECUTORCH_NPU.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
DOCUMENTATION_KAZEIA.txt Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
FROZEN.md chore: dégraissage du dépôt — suivi restreint à l'app mobile (liste blanche) 2026-06-09 21:21:00 +02:00
GUIDE_ROOT_ONEPLUS_PAD3.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
KAZEIA-CLAUDE.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
RAPPORT_KAZEIA_2026-05-04.md chore: dégraissage du dépôt — suivi restreint à l'app mobile (liste blanche) 2026-06-09 21:21:00 +02:00
RAPPORT_TTS.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
TTS_CALIBRATION_GUIDE.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
TTS_GPU_GUIDE.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
TTS_HEXAGON_NPU_GUIDE.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
TTS_RAPPORT_COMPLET.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
TTS_REPORT.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
kazeia-architecture.md Initial commit: Kazeia TTS pipeline on NPU via ExecuTorch 2026-04-09 08:42:11 +02:00
kazeia-no-root-report.md docs: add before/after performance comparison to no-root report 2026-04-14 11:37:15 +02:00