Add Pika H3 Flex (MiniMax-H3 plus the released fl2va/ref2va rank-128 adapters) as a FastVideo model family: registry/config/presets, bucket geometry with tiered reference sizing, automatic adapter selection and manifest validation, the fixed six-forward re-noise sampler, T2VA/I2VA/L2VA/FL2VA/Ref2VA pipeline support, an example, docs, and CPU unit tests.
GPU end-to-end validation is still pending. This change does not add PDD, VSA, or new kernel paths.
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.
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.
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.
Adds opt-in regional torch.compile to the YAML-driven modular Wan trainer (training.model.enable_torch_compile / torch_compile_kwargs) with FA4 and Wan modulation forward/backward as opaque custom ops, compiled repeated DiT blocks after FSDP setup, and activation checkpointing applied before FSDP for trainable Wan transformers.
Merged with main's #1716 (WanModel resolves the checkpointing type in __init__; the PR's pre-FSDP wrap uses it and the post-load wrap is gone) and #1757 (pre_fsdp_model_transform runs before this PR's pre_fsdp_transform). Maintainer review (Swarm order #776): AdaLN skip names are canonicalized through checkpoint-wrapper prefixes, test_inference_regional_compile.py joins the unit lane, the Wan modulation custom-op test moved to the transformer lane, regional_compile is classified in the schema inventory, and a 1/2-GPU training smoke covers fully_shard(CheckpointWrapper) in eager and regional-compile modes. Validated on GB200: the FA4 CuTe forward/backward parity and fullgraph-backward tests and all four smoke cases pass.
Also exempts MMAudio recipes, whose student builds its pipeline config at init, from the fine-tuning recipe construction check that #1648 left failing on main.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Adds the Cosmos Predict 2.5 distilled text-to-world sampler (sCM, per-frame timesteps) and the DFD video-to-world pipeline, with their schedulers, registry entries, presets and examples.
Maintainer review fixes (Swarm orders #718 and #748): the DiT receives one shared fps value (its RoPE table is batch-independent, so per-sample vectors crashed batch > 1; differing values now raise), non-finite rollouts raise FloatingPointError instead of being zeroed by nan_to_num, the DFD detector matches "dfd" only in the last path component, the distilled example defaults to 24 fps, and the distilled rollout keeps its state in FP64 through the scheduler step as the official driver does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Retargeted onto fastvideo/models/wan/causal_transformer.py, where main moved the causal Wan implementation (fastvideo/models/dits/causal_wanvideo.py is now a compatibility shim), with the guarded flex-attention trim prepared during the maintainer review (Swarm order #657).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
`load` took the host copies before checking which tensors were missing from
the device, so with `vae_cpu_offload=False` — where the loader builds the VAEs
on CUDA and `unload` is never called — it pulled about 11 GB per rank back over
PCIe once and kept a pinned mirror of it for the life of the process, for a
module that never leaves the device.
It now asks what is absent first and returns early when nothing is. The
offloaded path is unchanged: weights still round-trip and come back bit
identical.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
With the device-to-host copy gone, the remaining host-to-device copy is the
stage's cost, and from pageable memory it is staged through a bounce buffer at
roughly 2.6 GB/s instead of full PCIe speed. The host copies are made once, so
pinning them costs nothing per request.
It does cost the module's size in non-pageable host memory (10.4 GB for the
H3 video VAE, 0.6 GB for the audio VAE, per rank) for as long as the pipeline
is alive. It follows --pin-cpu-memory, which is on by default; --no-pin-cpu-
memory keeps the copies pageable and only loses the bandwidth. Note that this
flag did not already reach the VAE: its other use is FSDP's own CPUOffloadPolicy
and the VAE does not go through FSDP.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ATw7g6cxq9N2LtnMratsru
The decode stages move the 10.4 GB fp32 video VAE (and the 0.6 GB audio VAE)
with module.to(device) / module.to("cpu") on every request. The weights are
frozen -- the whole stage runs under no_grad -- so the copy back to the host
returns bytes the host already has.
Keep one host copy per tensor, made the first time the module is loaded, and
on unload point .data back at it instead of copying. One H2D per request, no
D2H at all; the module still lives on the host between requests. Where the old
path allocated a fresh host tensor per cycle and freed the previous one, this
reuses one buffer, so peak host memory can only go down.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ATw7g6cxq9N2LtnMratsru
aryan5v