# FastH3 Preview on four H100 80 GB GPUs (sm_90).
#
# Hopper has no FA4 and no sm_100a VSA kernel. This profile keeps the checkpoint,
# the four forwards and the 90% VSA sparsity of the GB200 profile and changes
# three things: the 64-token VSA path runs on the ThunderKittens sm_90a kernel
# (build fastvideo-kernel with CMAKE_ARGS='-DFASTVIDEO_KERNEL_BUILD_TK=ON'), the
# DiT is FSDP-sharded across the four GPUs instead of replicated, and the text
# encoder stays resident as block-FP8 (35.5 GB instead of 66.7 GB; produce it
# with scripts/checkpoint_conversion/quantize_minimax_h3_text_encoder_fp8.py),
# so nothing is offloaded between requests. Both deferral flags below
# (`lazy_module_load`, `h3_sequential_load`) are explicitly off for that reason:
# with either one on, the pipeline frees the text encoder after conditioning and
# reloads it on the next request. Keep the env prefix below, or pass
# the equivalent CLI flags on examples/inference/basic/basic_fasth3.py
# (`--num-gpus 4 --no-replicated-dit --vsa-kernel tk --no-fa4
# --text-encoder-weights ./FastH3-TextEncoder-FP8 --no-offload-text-encoder
# --no-offload-vae`).
#
# request.sampling below is the measured 4xH100 shape (960x544, 345 frames =
# 14.375 s at 24 FPS). Legal H3 frame counts are 17n+5, max 345.
#
#   FASTVIDEO_VSA_PTX_BLACKWELL=0 FASTVIDEO_FA4=0 FASTVIDEO_VSA_TK=1 \
#     FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 FASTVIDEO_STAGE_LOGGING=1 \
#     fastvideo generate --config examples/inference/basic/basic_fasth3_h100.yaml
generator:
  model_path: FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree
  pipeline:
    components:
      text_encoder_weights: ./FastH3-TextEncoder-FP8
    experimental:
      attention_backend: VIDEO_SPARSE_ATTN_H3
      VSA_sparsity: 0.9
      VSA_tile_size: 64
      inference_torch_compile: true
      vae_parallel_decode: true
      vae_parallel_decode_strategy: gather
      h3_sequential_load: false
  engine:
    num_gpus: 4
    use_fsdp_inference: true
    parallelism:
      tp_size: 1
      sp_size: 4
    offload:
      dit: false
      dit_layerwise: false
      text_encoder: false
      vae: false
      pin_cpu_memory: true
      lazy_module_load: false
    compile:
      enabled: false
      vae_enabled: true
request:
  prompt: >-
    A wide cinematic shot of an alpine meadow at sunrise, pale pink mountain
    peaks above a blue valley filled with thin morning mist.
  negative_prompt: ""
  sampling:
    seed: 1000
    height: 544
    width: 960
    num_frames: 345
    fps: 24
    num_inference_steps: 5
    guidance_scale: 1.0
    batch_cfg: false
  output:
    output_path: outputs/fasth3_h100/
    save_video: true
    return_frames: false