# FastH3 OmniRef (Ref2VA) on one RTX 5090 (32 GB): the full NVFP4 stack, one clip per prompt.
#
# Point model_path at a composed OmniRef model directory: the distilled
# transformer_ref export (with its packed 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 the same stack with --profile rtx-5090.
#
# 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:
#
# 32 GB: the text encoder stays off the GPU, the denoise modules load on demand and
# the AdaLN table replaces adaln_proj. The DiT itself stays whole - no layerwise streaming.
# FASTVIDEO_ATTENTION_BACKEND=VIDEO_SPARSE_ATTN_H3 \
# FASTVIDEO_H3_VSA_FP4=1 FASTVIDEO_H3_VSA_FP4_PREFIX=fp8 \
# FASTVIDEO_H3_VSA_FP4_TWO_LEVEL_P=1 FASTVIDEO_H3_VSA_FP4_K_SMOOTH=1 \
# FASTVIDEO_H3_GATE_CHUNK_ROWS=8192 \
# FASTVIDEO_ULYSSES_A2A=auto FASTVIDEO_MINIMAX_H3_EXACT_KERNELS=all \
# FASTVIDEO_H3_VSA_HEADS_FIRST_TILE=1 FASTVIDEO_H3_VAE_TILE_PARALLEL=1 \
# FASTVIDEO_H3_REF2VA_MEMO_ENTRIES=16 \
# FASTVIDEO_MINIMAX_H3_FUSIONS=all \
# FASTVIDEO_VSA_TRITON=1 FASTVIDEO_VSA_PTX_BLACKWELL=0 FASTVIDEO_FA4=0 \
# FASTVIDEO_H3_ADALN_TABLE=/path/to/adaln_table.pt \
# FASTVIDEO_H3_VAE_NVFP4_DECODER=/path/to/decoder-nvfp4.safetensors \
# FASTVIDEO_H3_VAE_TILE_BATCH=4 \
# fastvideo generate --config examples/inference/basic/basic_fasth3_omniref_rtx5090.yaml
#
# See docs/inference/fasth3_omniref.md.
generator:
  model_path: ./FastH3-OmniRef-Composed
  engine:
    num_gpus: 1
    use_fsdp_inference: false
    quantization:
      transformer_quant: NVFP4
      layer_profile: h3_dit_vsa
    parallelism:
      tp_size: 1
      sp_size: 1
    offload:
      dit: false
      dit_layerwise: false
      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
      h3_vsa_fp4:
        enabled: true
        prefix_precision: fp8
        two_level_p: true
        smooth_k: true
      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_rtx5090/
    save_video: true
    return_frames: false