hao-ai-lab-FastVideo/tests/local_tests/minimax_h3

MiniMax H3 validation

Local tests keep checks that need the pinned Diffusers source, published weights, or the public registry surface. FastVideo-owned unit contracts belong under fastvideo/tests/.

Reference

• Diffusers implementation: https://github.com/huggingface/diffusers/pull/14355
• Source checkout: ${MINIMAX_H3_OFFICIAL_REF_DIR:-$PWD/DiffusersMiniMaxH3}
• Checkpoint: MiniMaxAI/MiniMax-H3

The reference helper verifies the pinned source and import origin. A missing checkout may skip a source-parity module; that skip is not parity evidence.

FastVideo unit contracts

pytest \
  fastvideo/tests/vaes/test_minimax_h3_video_vae_streaming.py \
  fastvideo/tests/stages/test_minimax_h3_vae_streaming.py -q

Registry smoke

pytest tests/local_tests/pipelines/test_minimax_h3_pipeline_smoke.py -q

Pinned implementation parity

PYTHONPATH="${MINIMAX_H3_OFFICIAL_REF_DIR:-$PWD/DiffusersMiniMaxH3}/src:$PWD" pytest \
  tests/local_tests/minimax_h3/test_minimax_h3_scheduler_parity.py \
  tests/local_tests/minimax_h3/test_minimax_h3_packing.py \
  tests/local_tests/minimax_h3/test_minimax_h3_ref2va_packing.py \
  tests/local_tests/minimax_h3/test_minimax_h3_ref2va_media.py -v -s

Checkpoint component parity

export MINIMAX_H3_MODEL_ROOT=/path/to/MiniMax-H3
export MINIMAX_H3_OFFICIAL_REF_DIR=/path/to/DiffusersMiniMaxH3

PYTHONPATH="$MINIMAX_H3_OFFICIAL_REF_DIR/src:$PWD" \
MINIMAX_H3_RUN_ENCODER_PARITY=1 \
pytest tests/local_tests/encoders/test_minimax_h3_qwen3_vl_parity.py -v -s

PYTHONPATH="$MINIMAX_H3_OFFICIAL_REF_DIR/src:$PWD" \
MINIMAX_H3_RUN_DIT_PARITY=1 \
MINIMAX_H3_RUN_VIDEO_VAE_PARITY=1 \
MINIMAX_H3_RUN_AUDIO_VAE_PARITY=1 \
pytest \
  tests/local_tests/transformers/test_minimax_h3_transformer_parity.py \
  tests/local_tests/vaes/test_minimax_h3_video_vae_parity.py \
  tests/local_tests/vaes/test_minimax_h3_audio_vae_parity.py -v -s

With a gate enabled, missing CUDA, source, or weights is a failure. Recorded component evidence is exact for both DiT partitions and the video VAE; audio decode has maximum absolute drift 2.4e-7. The encoder gate compares the slim forward's selected layer-50 hidden state bit-exactly against the same state from the official full stack across text, image, and video inputs.

The video VAE test verifies the reference checkout at commit abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc and compares the production CPU uint8 encode_pixels() path against the official posterior element by element.

Video VAE memory benchmark

The benchmark uses one warmup and three measured runs with vae_cpu_offload=True. It reports absolute and stage-incremental allocated/reserved CUDA peaks for every rank. For SP runs, the reported aggregate is explicitly the sum of rank-local maxima, not a simultaneous node peak.

python tests/local_tests/vaes/benchmark_minimax_h3_video_vae_memory.py \
  --source-root "$PWD" --model-root "$MINIMAX_H3_MODEL_ROOT" \
  --revision-label candidate --operation encode

python -m torch.distributed.run --nproc_per_node=4 \
  tests/local_tests/vaes/benchmark_minimax_h3_video_vae_memory.py \
  --source-root "$PWD" --model-root "$MINIMAX_H3_MODEL_ROOT" \
  --revision-label candidate-sp4 --operation decode

Run the same script with --source-root pointed at the base checkout for a comparable baseline. The default workload is deterministic 124 x 768 x 1344 video geometry with seed 20260803; the JSON record includes source/model revisions, software/allocator metadata, exact measurement boundaries, per-repetition values, and output shapes.

FastVideo joint audio/video generation and SP=1/SP=4 latent consistency have been validated. T2VA, FL2VA, and Ref2VA video/audio latents match the pinned Diffusers pipeline exactly.