# FastH3 OmniRef (Ref2VA) on one RTX 4090 (24 GB, no FP4): the FP8/bf16 equivalent, one clip per prompt.
#
# Point model_path at a composed OmniRef model directory: the distilled
# dense or FP8 transformer_ref export (without nvfp4_weights.safetensors) symlinked
# together with base MiniMax-H3's text_encoder, tokenizer, processor, vae and
# audio_vae. examples/inference/basic/basic_fasth3_omniref_pdd.py builds that
# directory and runs it with --profile rtx-4090.
#
# The contract in the export's fastvideo_inference.json fixes the attention
# backend, the VSA tile size and the eight-forward PDD partition; the run only
# picks the device's offload policy and its environment:
#
# sm_89 has no FP4 hardware, so this profile is the closest equivalent rather than
# the FP4 stack: FP8 linears, bf16 VSA attention, the dense light VAE decoder (the
# NVFP4 one needs FP4 hardware too), a non-resident encoder and expandable segments
# to reduce memory fragmentation. Layerwise offload is mandatory: the 47 GB FP8
# DiT cannot be resident on 24 GB. Run one process per clip; repeated requests in
# one process exceed the 24 GB GPU and the host cgroup when the encoder reloads.
#
# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 \
# FASTVIDEO_H3_VSA_FP4=0 \
# FASTVIDEO_H3_VSA_FINE_ATTENTION=emulator-bf16 \
# FASTVIDEO_LAYERWISE_OFFLOAD_BUFFERS=1 \
# FASTVIDEO_LAYERWISE_RESIDENT_BLOCKS=16 \
# FASTVIDEO_H3_FFN_CHUNK_TOKENS=16384 \
# FASTVIDEO_H3_PINNED_SWAP=0 \
# PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
# FASTVIDEO_H3_ADALN_TABLE=/path/to/adaln_table.pt \
# fastvideo generate --config examples/inference/basic/basic_fasth3_omniref_rtx4090.yaml
#
# See docs/inference/fasth3_omniref.md.
generator:
  model_path: ./FastH3-OmniRef-Dense-Composed
  engine:
    num_gpus: 1
    use_fsdp_inference: false
    quantization:
      transformer_quant: FP8
    parallelism:
      tp_size: 1
      sp_size: 1
    offload:
      dit: false
      dit_layerwise: true
      text_encoder: true
      vae: false
      pin_cpu_memory: true
      lazy_module_load: false
  pipeline:
    workload_type: i2v
    components:
      override_pipeline_cls_name: MiniMaxH3Ref2VAModularPipeline
    experimental:
      attention_backend: VIDEO_SPARSE_ATTN_H3
      video_decode_backend: h3-vae
      h3_sequential_load: true
request:
  prompt: >-
    A person on a sofa, lit by a warm lamp, speaking to camera.
  negative_prompt: ""
  inputs:
    references:
      # Ordered: the clip is conditioned on these. Replace with your own media.
      - source: assets/images/couple.jpg
        media_type: image
  sampling:
    seed: 1234
    height: 480
    width: 832
    num_frames: 124
    fps: 24
    num_inference_steps: 8  # eight fused PDD blocks; the contract sets it
    guidance_scale: 1.0
    batch_cfg: false
  output:
    output_path: outputs/basic_fasth3_omniref_rtx4090/
    save_video: true
    return_frames: false