Fetching metadata from the HF Docker repository...
Sep 2 01:56 fix: +torchvision (Gemma4Processor do text encoder LTX-2.5 exige) 01:45 fix imports: LTX2Guidance do top-level (guider path do card nao existe no main) 01:35 v2: LTX-2.5 22B distilled via Modular Diffusers (T2V+I2V, audio 48kHz); fix deps (unpin transformers); remove TierRouter 01:29 fix runtime: gradio 6 + unpins (requirements.txt) 01:29 fix runtime: gradio 6 + unpins (README.md) Jun 12 03:45 fix(mux): normaliza audio LTX-2 p/ libsndfile + SR via runtime probe 03:19 fix(mux): audio joint LTX-2 .cpu() antes do numpy + sample rate vocoder 02:09 A-diffusers fix: README.md (LTX2Pipeline T2V, sem variant/cache_examples) 02:09 A-diffusers fix: tier_router.py (LTX2Pipeline T2V, sem variant/cache_examples) 02:09 A-diffusers fix: app.py (LTX2Pipeline T2V, sem variant/cache_examples) May 29 19:48 fix(build): unpin spaces (HF builder forca spaces==0.50.2; ==0.50.0 quebrava ResolutionImpossible) 19:36 fix(LTX-2): LTX2Pipeline + variant fp8 + frame_rate kwarg + 16:9 1280x704 + show_error (conserta AppError na geracao) May 15 04:19 feat: TierRouter R1-R5 blindagem (lock + timeout + try-except + token + VRAM) Pauloeduspbr11 +1: Claude Opus 4.7 (1M context) R1 - threading.RLock em _execute_local protege _loaded_pipe contra race condition
R2 - ThreadPoolExecutor + future.result(timeout=600s) por strategy (StrategyTimeoutError)
R3 - try/except (StrategyTimeoutError, RuntimeError, Exception) em app.py generate(),
sempre retorna gr.Error estruturado pro caller (nunca crasha o Space)
R4 - validate_token() opcional: token presente mas 401 -> InvalidTokenError -> gr.Error
(em vez de cair pra FREE silencioso)
R5 - _free_vram_gb() via torch.cuda.mem_get_info() antes de LOCAL_PRO; se VRAM
livre < pro_min_free_vram_gb (default 30GB), cai pra LOCAL_FREE com motivo
Garantia: transicao entre strategies (FREE/PRO/PROVIDER) nao quebra Space.
Cada strategy isolada em thread com timeout. Falhas registradas em CircuitBreaker.
Exception sempre virou gr.Error formatado pro caller.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> 03:59 fix: enable_model_cpu_offload em vez de pipe.to(cuda) (NVML INTERNAL ASSERT em ZeroGPU) + docs TierRouter no README 03:46 fix: transformers==4.55.0 (diffusers LTX2 exige Gemma3ForConditionalGeneration >=4.50) May 14 21:43 fix: api_name 'generate' sem barra (Gradio gerava //generate com prefixo duplo) 21:28 feat: integrate TierRouter (dual-tier routing: LOCAL_FREE/LOCAL_PRO/REMOTE_PROVIDER) Pauloeduspbr11 +1: Claude Opus 4.7 (1M context) - tier_router.py: detect_tier(token) -> FREE|PRO|PAID via HfApi.whoami
- choose_strategy(tier, size, premium) -> RouteDecision com fallback cascade
- execute() roda strategy + cai para fallback se falhar (Circuit Breaker, Fowler 2014)
- cleanup_after_request(keep_warm) libera VRAM via del + empty_cache + gc
- app.py: gradio Request injetado, caller token extraido, premium_provider checkbox
- caller paga: HF debita cota do token do caller (decisao usuario 2026-05-14)
- design completo em audit/_tier_router_pattern.md (5 papers + HF docs + Fowler/Netflix)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> 21:07 fix: bump huggingface_hub==0.35.0 + accelerate==1.4.0 (diffusers main exige hub>=0.34.0) 21:03 fix: align spaces==0.50.0 + gradio[oauth] (HF builder injects these) 20:43 fix: torch==2.8.0 (ZeroGPU exige 2.8.0+; 2.5.1 rejeitado em CONFIG_ERROR) 20:42 audit FASE 1 Caminho A: pin deps, add packages.txt ffmpeg, remove cpu_offload (H200 80GB cabe fp8 ~27GB), add 3 gr.Examples, fix api_name /generate Pauloeduspbr11 +1: Claude Opus 4.7 (1M context) Validacoes 2026-05-14:
- Lightricks/LTX-2 mantido (LTX-2.3 ainda nao tem diffusers integration; LTX-2.3-fp8 sem model_index.json)
- diffusers @ git+main mantido com SHA livre (tech debt: pinar quando LTX-2 release oficial)
- pip index versions usado para escolher pins estaveis (gradio 5.27.0, torch 2.5.1, transformers 4.48.0, accelerate 1.1.1)
- packages.txt adicionado para garantir ffmpeg disponivel no runtime ZeroGPU (mux audio em _mux_video_audio)
- enable_model_cpu_offload removido + pipe.to(cuda): fp8 distilled (~27GB) cabe folgado em H200 80GB
- 3 examples T2V joint / I2V motion / T2A puro
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> May 8 23:15 Migrate to LTX-2 19B distilled fp8 (joint video+audio, T2V/I2V/V2V/A2V/V2A/T2A all-in-one, multi-lang incl PT) 22:51 Recreate video-generator: Wan2.2-I2V-A14B Apache-2.0 (out of canon v6, pontual use) 22:51 initial commit 
