atharv-1447/META-TRIBE-v2-api

feat amke it t4 compaible

fix: marge remote

feat: made t4 compatibale

fix: avoid t4 vram contention during feature extraction

fix: build text events without tts roundtrip

fix: avoid per-request hf login hangs

fix: stabilize gpu inference pipelines

fix: pull origin

fix: error messgae .config file

fix: avoid cuda probe during web startup

fix: decouple gpu warmup from model readiness

fix: attempt fix build #5

fix: attempt fix build #4

fix: docker and requirements txt

fix: attempt fix build #2

fix: upgrade to Python 3.12, in-process whisper, fp16 autocast, T4 hardening

fix: docker git fetch error

fix: implemented support for T4

chore: hidden probe

fix: image processing for odd dims, whisper language support, base64 text route

fix: cache-bust Dockerfile to force app.py re-copy on HF build

fix: global DataLoader patch + warmup prediction to fix container hang

feat: production hardening — gunicorn, pre-load, hard timeout, GPU cleanup, persistent cache

fix: detect site-packages via python and verify tribev2 import in build

fix: copy tribev2 to site-packages not dist-packages — python:3.11-slim uses site-packages

fix: pin matched torch 2.5.1 + torchvision 0.20.1 + torchaudio 2.5.1 (cu121) — resolves torchvision::nms ABI mismatch

fix: remove torch force-reinstall — neuralset pulls torch 2.11.0+cu130 which has register_fake API

fix: install tribev2 via git clone+copy — pip wheel builds as UNKNOWN with no module files

fix: use python:3.11-slim base + cu121 torch — avoid tribev2 path mismatch

feat: GPU support via nvidia/cuda:12.1 base image — remove CPU-only patches, MAX_WORDS=75

fix: replace whisperx with openai-whisper to bypass ctranslate2 executable stack block

fix: use whisperx==3.3.0 to resolve av==11.* build failure

fix: pre-install av>=12 binary wheel to unblock whisperx build

fix: force-reinstall torch family after tribev2 to fix torchaudio binary mismatch

fix: resolve whisperx dependency and runtime bugs in tribe-space

fix: fix wisper vulnerability

fix: serialize full pipeline to fix concurrent 503

fix: word cap, busy 503, and token format validation

fix: shorten short_description to meet HF 60-char limit

docs: comprehensive README with API reference, build history, and CPU patches

fix: remove transformers version pin — use latest to get AutoVideoProcessor

fix: loosen huggingface_hub to >=0.24.0 to satisfy transformers==4.47.0 dependency

fix: pin transformers==4.47.0 to get AutoVideoProcessor (forces Docker cache invalidation)

fix: patch all neuralset extractor device fields to cpu via model_post_init; upgrade transformers>=4.46.0

fix: upgrade transformers>=4.46.0 for AutoVideoProcessor support

fix: patch BatchEncoding.to instead of Tensor.to (C ext not patchable), set CUDA_VISIBLE_DEVICES in Dockerfile

fix: redirect all .to(cuda) → .to(cpu) at PyTorch level, hide CUDA entirely

feat: per-request hf_token auth — no Space secret needed, users bring their own token

fix: replace float16 with int8 in WhisperX cmd and patch ExtractWordsFromAudio directly

fix: patch subprocess.run to pass --compute_type int8 to WhisperX (CPU has no float16)