# Restore InfiniteTalk + FP8 TensorRT on a fresh H100

## 1. Code
    tar xzf code.tar.gz -C /workspace

## 2. Weights (NOT in this backup - re-download, they are public)
    hf download Wan-AI/Wan2.1-I2V-14B-480P --local-dir /workspace/weights/Wan2.1-I2V-14B-480P
    hf download MeiGen-AI/InfiniteTalk --include "single/*" --local-dir /workspace/weights/InfiniteTalk
    hf download TencentGameMate/chinese-wav2vec2-base --local-dir /workspace/weights/chinese-wav2vec2-base
    hf download TencentGameMate/chinese-wav2vec2-base --revision refs/pr/1 --include "model.safetensors" --local-dir /workspace/weights/chinese-wav2vec2-base
    hf download Kijai/WanVideo_comfy Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank64_bf16.safetensors --local-dir /workspace/weights/loras

## 3. Python env (pin torch, or the CUDA stack gets swapped out from under you)
    uv pip install -c constraints.txt -r pip-freeze.txt
Key pins: torch==2.8.0+cu128, xformers==0.0.32.post2, transformers==4.51.3,
flash_attn 2.8.3.post1 (cu12torch2.8, cxx11abiTRUE), and
export XFORMERS_IGNORE_FLASH_VERSION_CHECK=1

## 4. TRT engines
The engines in engines/ are a CACHE, valid only on the same GPU model + TRT
version (see VERSIONS.txt). Try them first:

    python -c "
    import tensorrt as trt
    rt = trt.Runtime(trt.Logger(trt.Logger.ERROR))
    e = rt.deserialize_cuda_engine(open('/workspace/engines/block0.plan','rb').read())
    print('engines OK' if e else 'ENGINES INVALID -> rebuild')"

If invalid (different GPU/TRT), rebuild - ~18 min total, fully scripted:
    # a) capture calibration activations (~5 min)
    CAPTURE=1 <run generate_infinitetalk.py once, 81-frame clip>   # see notes below
    # b) build 40 FP8 engines (~13 min)
    python /workspace/trt/build_engines.py

## 5. Run
    /workspace/start_server.sh      # warm server, loads model once
    /workspace/generate.sh <image> <audio> "<prompt>" <name>

## Gotchas
- Engines are STATIC: 81-frame windows @ 448x832 (30,576 tokens) with the 4-step
  I2V LoRA baked in. Change resolution / frame count / LoRA / steps -> rebuild.
- Under WAN_TRT=1 the PyTorch blocks are DELETED (multitalk.py:689 would
  otherwise drag 27GB back onto the GPU).
- All 40 engine contexts share ONE scratch buffer (5.85GB); 40 private ones
  would need 234GB.