# FastH3 OmniRef (Ref2VA) across two DGX Sparks over QSFP RoCE, eight fused PDD
# forwards per clip. Start Ray on both Sparks and source
# examples/inference/optimizations/spark_pair_env.sh first; see
# docs/getting_started/installation/spark_pair.md.
#
# Point model_path at a composed OmniRef model directory (see
# basic_fasth3_omniref_pdd.py): the distilled transformer_ref export, with its
# packed nvfp4_weights.safetensors, symlinked with base MiniMax-H3's encoder,
# tokenizer, processor and VAEs.
#
# GB10 uses Triton VSA. The two nodes share the sequence (sp_size 2), which is
# the only split this pipeline has: H3 generates a clip at a time, so there is
# no depth to pipeline. The FP8 prefix split runs on the sequence-parallel path
# too: prefix rows stay dense and video rows stay on the FP4 kernel.
#
# 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_NVFP4_MM_BACKEND=cudnn \
# FASTVIDEO_H3_VAE_TILE_BATCH=12 FASTVIDEO_H3_VAE_INT8_OVERLAY=0 \
# 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 generate --config examples/inference/basic/basic_fasth3_omniref_spark_pair.yaml
#
# See docs/inference/fasth3_omniref.md.
generator:
  model_path: ./FastH3-OmniRef-Composed
  engine:
    num_gpus: 2
    execution_backend: ray
    use_fsdp_inference: false
    quantization:
      transformer_quant: NVFP4
      layer_profile: h3_dit_vsa
    parallelism:
      tp_size: 1
      sp_size: 2
    offload:
      dit: false
      dit_layerwise: false
      text_encoder: false
      image_encoder: false
      vae: false
      pin_cpu_memory: false
      lazy_module_load: false
    compile:
      enabled: false
      vae_enabled: true
  pipeline:
    workload_type: i2v
    components:
      override_pipeline_cls_name: MiniMaxH3Ref2VAModularPipeline
    vae_tiling: true
    experimental:
      attention_backend: VIDEO_SPARSE_ATTN_H3
      h3_vsa_fp4:
        enabled: true
        prefix_precision: fp8
        two_level_p: true
        smooth_k: true
      vae_parallel_decode: true
      vae_parallel_encode: true
      video_decode_backend: h3-vae
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/fasth3_omniref_spark_pair/
    save_video: true
    return_frames: false