Port of vllm-project/vllm-omni#8602. PAN2 is a 29B mixture-of-transformers
video DiT (separate video and text streams joined in one attention per
block) with a Qwen3.5 text encoder and the Wan2.2 VAE. One pipeline serves
text-to-video and first-frame image-to-video.
- Native PAN2 DiT; per-stream Q/K/V fused at load time, Ulysses SP with
replicated text tokens through DistributedAttention.
- Native Qwen3.5 text encoder (Gated DeltaNet + gated attention) with a
PyTorch chunked delta rule.
- Reuses the Wan2.2 VAE and the flow-match Euler scheduler.
- PAN2 stages: chat-template prompt crop at hidden_states[-3], first-frame
conditioning, uniform sigmas with shift 7, fp32 latents. Adds a
DenoisingStage.prepare_scheduler_model_output hook for the last one.
- Registry entry for the Modular Diffusers PAN2ModularPipeline manifest,
pan2_t2v / pan2_i2v presets, example, docs, and local parity tests
against a diffusers/transformers reference on the public tiny checkpoint.
The unit lane fails on main in two contract tests. All three failures are
bookkeeping that drifted from the code, not runtime bugs:
- `test_env_policy.py::test_env_doc_table_matches_registry`: the generated
table in docs/contributing/env_vars.md is stale for
FASTVIDEO_MLX_QUANT_GEMM. Regenerated with the documented command
(`python fastvideo/tests/contract/test_env_policy.py`).
- `test_env_policy.py::test_env_access_follows_policy`: two new violations.
- fastvideo/tests/mlx/test_mlx_affine_dq_gemm.py wrote
FASTVIDEO_MLX_QUANT_GEMM through monkeypatch.setenv; now uses the
registered field's override() through the env_overrides fixture.
- fastvideo/tests/train/models/test_load_kandinsky5.py read
FASTVIDEO_ATTENTION_BACKEND with `not in os.environ`; now uses
envs.FASTVIDEO_ATTENTION_BACKEND.is_set() and override(). Its entry in
KNOWN_VIOLATIONS is deleted, as the shrinking-list rule requires.
- `test_schema_parity_inventory.py::..._extension_fields_are_classified`:
MiniMaxH3PipelineConfig declares video_scheduler_shift and
audio_scheduler_shift (training-time noise clocks rejected by
check_pipeline_config at inference); both are added to
pipeline_config_extensions.internal_only.
Test plan (CPU, no GPU required):
python -m pytest fastvideo/tests/contract/test_env_policy.py \
fastvideo/tests/api/test_schema_parity_inventory.py \
fastvideo/tests/mlx/test_mlx_affine_dq_gemm.py -q
# 45 passed
python -m pytest fastvideo/tests/contract/ fastvideo/tests/api/ -q -m "not gpu"
# 501 passed
VideoGenerator.from_config calls from_fastvideo_args with a log_queue keyword
argument. The StreamingVideoGenerator override did not accept it, so
StreamingVideoGenerator.from_pretrained raised TypeError. The override now
accepts log_queue and passes it to the executor through VideoGenerator.__init__.
A new test covers the call.
The HunyuanVideo batch test wrote MASTER_PORT, MASTER_ADDR and
FASTVIDEO_ATTENTION_BACKEND with monkeypatch.setenv, so the environment-variable
contract test failed. The fixture now uses envs.override_external and the
registry override. The values and the restore behavior do not change.
These speed-ups for the MiniMax-H3 pipelines keep the output bit-identical. Each
one has a switch that is off by default:
- FASTVIDEO_MINIMAX_H3_EXACT_KERNELS: Triton RoPE, AdaLN modulation, gated
residual and SwiGLU kernels that round to bf16 at the same points as the
eager ops.
- FASTVIDEO_H3_VSA_HEADS_FIRST_TILE: VSA-H3 tiles are written directly in the
heads-first layout of the kernel.
- FASTVIDEO_H3_VAE_TILE_PARALLEL: VAE spatial tiles are split across the
sequence-parallel ranks.
- FASTVIDEO_H3_REF2VA_MEMO_ENTRIES: repeat reference encodes are reused.
The example basic_fasth3_omniref_pdd.py --lossless-accel turns them all on.
On GB200 with 4 GPUs, MiniMax-H3 text to video with audio (124 frames,
1344x768) gives bit-identical frames and audio with the first three switches
on and off.
MultiprocExecutor created two streaming multiprocessing queues for every
generator, but only StreamingVideoGenerator uses them. Each queue uses POSIX
named semaphores in /dev/shm. When a job epilog deleted these files, spawned
workers failed to start. The executor now creates the queues only when
FastVideoArgs.enable_streaming_ipc_queues is set. StreamingVideoGenerator sets
it on a copy of its arguments in queue mode. Standard generation creates no
streaming semaphores.
The distributed_setup test fixture now sets MASTER_ADDR and MASTER_PORT when
they are not set, with a new port for each test, and restores them after the
test. pyproject.toml now declares orjson and python-multipart, which the
OpenAI-compatible server needs. Two documents about the SSIM reference
bootstrap are corrected.
For Wan VSA inference with 64-token tiles, each attention layer scattered Q, K,
V and the gate into a tile buffer and then made four transposed copies. On
SM100, one Triton kernel (tile_to_bhsd) now does the scatter and the transposes
in one pass. Training, 256-token tiles, other GPUs and builds without the kernel
use the previous path. FASTVIDEO_DISABLE_VSA64_FUSED_LAYOUT=1 turns the fused
path off.
On GB200 the layout and attention step goes from 15.7 ms to 8.6 ms. FastWan 1.3B
frames are bit-exact with the previous path at sequence parallel 1 and 2.
The HunyuanVideo and HunyuanVideo 1.5 transformers failed on the first forward
pass when the batch had more than one sample. The modulation inputs had the
shape [B, C], but the fused ops expect [B, 1, C], so they could not broadcast
over the sequence. The models now add the sequence axis, as the refiner and the
final layer already do. Results for batch size 1 do not change.
On SM100, the Attn-QAT training kernel now uses a faster route for non-causal
BF16 attention with equal query and key lengths of 16384 tokens or more. The
forward pass processes full KV tiles without bounds masks, and a separate masked
step handles the partial tail tile. The backward pass uses 64x128 dQ and dK/dV
tiles. Outputs, softmax statistics and all gradients are bitwise identical to the
previous route. FASTVIDEO_ATTN_QAT_SM100_OPTIMIZED=0 turns off the SM100 route,
and FASTVIDEO_ATTN_QAT_SM100_WIDE_BWD=0 keeps the 64x64 backward tiles.
On GB200 with 6 heads and 31200 tokens, the forward kernel time goes from 9.14 ms
to 6.96 ms and the backward kernel time from 17.41 ms to 15.85 ms.
For n > 1, the image endpoints returned one preview grid instead of n images.
The endpoints now run one model batch, save each decoded sample as its own file,
and return one URL or base64 item for each image. This applies to generations
and edits.
A keyword-only return_samples field (default False) is added to OutputConfig,
SamplingParam and ForwardBatch. Batched image requests set it and do not build
preview frames. The worker keeps the sample tensor for these requests, and
non-output ranks clear the field. Defaults and positional arguments do not
change.
Ray sets CUDA_VISIBLE_DEVICES of each GPU actor to the GPU of that actor. The
import of fastvideo initializes CUDA, so the executor cannot change the device
list later. On a node with more than one GPU, each worker after the first worker
failed with "invalid device ordinal".
On NVIDIA GPUs, each worker actor now keeps the device list of its raylet
(RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=1). Each worker uses the position of
its GPU in that list as its local rank. The safetensors loader now reads each
file on node-group rank 0, which is the rank that sends the broadcast.
ParquetDatasetWriter.flush() named each shard from the chunk index plus the number
of files in the worker directory. When two flushes split the chunks in different
ways, a new shard could get the name of an existing shard and replace it. The
data in the old shard was lost without an error.
New shards now get numbers after the highest existing shard number in all worker
directories. One helper writes each shard to a temporary file and renames it. It
raises FileExistsError instead of an overwrite. flush(write_remainder=True) now
also counts the remainder rows in its return value.
aryan5v