[docs] Pika H3 Flex cookbook: --weights for the NVFP4 QAD releases and FP4 attention (#2054)
Co-authored-by: Aryan Kumar <aryan5v@users.noreply.github.com>
This commit is contained in:
Aryan Kumar
co-authored by
Aryan Kumar
Aryan Kumar
parent
4b4df1a94e
commit
661d18ba55
@@ -766,7 +766,7 @@
|
||||
"stage": "inference",
|
||||
"task": "Text to video (with audio)",
|
||||
"label": "Pika H3 Flex",
|
||||
"summary": "Pika's post-trained MiniMax-H3: synchronized video and stereo audio from text in six forwards, at 480p or 768p, 5-15 s, with optional NVFP4 decoder and text encoder on Blackwell.",
|
||||
"summary": "Pika's post-trained MiniMax-H3: synchronized video and stereo audio from text in six forwards, at 480p or 768p, 5-15 s, in BF16 or as FastVideo's NVFP4 QAD release on Blackwell, with optional FP4 attention.",
|
||||
"model": "Pika-Labs/Pika-H3-Flex",
|
||||
"source": "examples/inference/basic/basic_pika_h3_flex.py",
|
||||
"command": "FASTVIDEO_ATTENTION_BACKEND=FLASH_ATTN FASTVIDEO_FA4=1 python examples/inference/basic/basic_pika_h3_flex.py --task t2va --offload resident --prompt 'A paper boat drifts down a rainy city street at night; rain patters on the water.'",
|
||||
@@ -783,6 +783,14 @@
|
||||
"expected_artifact": "MP4 with synchronized stereo audio at outputs/pika_h3_flex/pika_h3_flex_t2va.mp4",
|
||||
"modes": ["T2VA"],
|
||||
"knobs": [
|
||||
{
|
||||
"key": "weights",
|
||||
"label": "Weights",
|
||||
"hint": "NVFP4 QAD needs Blackwell",
|
||||
"flag": "--weights",
|
||||
"options": [{"value": "bf16", "label": "BF16 (Pika release)"}, {"value": "nvfp4-qad", "label": "NVFP4 QAD"}, {"value": "nvfp4-qad-fp4-attention", "label": "NVFP4 QAD + FP4 attention"}],
|
||||
"default": "bf16"
|
||||
},
|
||||
{
|
||||
"key": "resolution",
|
||||
"label": "Resolution",
|
||||
@@ -855,6 +863,7 @@
|
||||
}
|
||||
],
|
||||
"limitations": [
|
||||
"NVFP4 QAD release (--weights), one GB200, 768p 16:9, 5 s, warm: 16.8 s with BF16 attention and 15.2 s with FP4 attention; at 768p, 15 s, FP4 attention is 16% faster than BF16 attention (85.3 s vs 101.8 s).",
|
||||
"Measured on one GB200 at 768p 16:9, 5 s, warm: official pika-h3-pro CLI 59.7 s; FastVideo BF16 26.2 s; with the NVFP4 quality-36L decoder 18.0 s; with the fast-8L decoder 17.4 s. The 36L decoder measured 40.5 dB PSNR against the BF16 decoder.",
|
||||
"Peak resident memory was about 129 GiB for the all-BF16 run and 104.6 GiB with the NVFP4 text encoder (25.49 s) at 768p / 5 s on GB200; the official CLI's sampling peak was 71.7 GiB, a different measurement. The NVFP4 text encoder's first request in a fresh process pays a one-time cost of about 183 s.",
|
||||
"--offload resident keeps every component on the GPU (104.6-108.5 GiB peak with the NVFP4 text encoder, about 129 GiB all-BF16, on GB200). --offload lazy loads components on demand for smaller devices; on GB200 it took 48.06 s for the same T2VA request.",
|
||||
@@ -887,6 +896,14 @@
|
||||
"expected_artifact": "MP4 with synchronized stereo audio at outputs/pika_h3_flex/pika_h3_flex_<task>.mp4",
|
||||
"modes": ["I2VA", "L2VA", "FL2VA"],
|
||||
"knobs": [
|
||||
{
|
||||
"key": "weights",
|
||||
"label": "Weights",
|
||||
"hint": "NVFP4 QAD needs Blackwell",
|
||||
"flag": "--weights",
|
||||
"options": [{"value": "bf16", "label": "BF16 (Pika release)"}, {"value": "nvfp4-qad", "label": "NVFP4 QAD"}, {"value": "nvfp4-qad-fp4-attention", "label": "NVFP4 QAD + FP4 attention"}],
|
||||
"default": "bf16"
|
||||
},
|
||||
{
|
||||
"key": "task",
|
||||
"label": "Task",
|
||||
@@ -1002,6 +1019,14 @@
|
||||
"expected_artifact": "MP4 with synchronized stereo audio at outputs/pika_h3_flex/pika_h3_flex_ref2va.mp4",
|
||||
"modes": ["Ref2VA"],
|
||||
"knobs": [
|
||||
{
|
||||
"key": "weights",
|
||||
"label": "Weights",
|
||||
"hint": "NVFP4 QAD needs Blackwell",
|
||||
"flag": "--weights",
|
||||
"options": [{"value": "bf16", "label": "BF16 (Pika release)"}, {"value": "nvfp4-qad", "label": "NVFP4 QAD"}, {"value": "nvfp4-qad-fp4-attention", "label": "NVFP4 QAD + FP4 attention"}],
|
||||
"default": "bf16"
|
||||
},
|
||||
{
|
||||
"key": "resolution",
|
||||
"label": "Resolution",
|
||||
|
||||
@@ -14,7 +14,8 @@ does not expose a step-count flag. ``--lossless-accel`` turns on the existing
|
||||
training-free FastVideo switches (bit-identical kernels, reference memo,
|
||||
parallel VAE tiles) before the generator starts. ``--nvfp4-decoder`` and
|
||||
``--nvfp4-text-encoder`` opt into the existing Blackwell FP4 weight paths;
|
||||
``--nvfp4-dit h3_dit`` loads a packed NVFP4 DiT export (e.g.
|
||||
``--weights nvfp4-qad`` (or ``nvfp4-qad-fp4-attention``) selects FastVideo's NVFP4 QAD release;
|
||||
``--nvfp4-dit h3_dit`` loads any packed NVFP4 DiT export (e.g.
|
||||
FastVideo/Pika-H3-Flex-NVFP4-QAD). They change weights, not the six-forward sampler. Kernel/VSA work is separate.
|
||||
|
||||
``--nvfp4-attention`` turns on FlashAttention-4 NVFP4 attention (Q/K in NVFP4, P/V in BF16; Blackwell and FA4 only)
|
||||
@@ -104,9 +105,31 @@ def _reference(value: str) -> MiniMaxH3Reference:
|
||||
return MiniMaxH3Reference(source=source, media_type=media_type)
|
||||
|
||||
|
||||
WEIGHT_PRESETS = {
|
||||
"bf16": {
|
||||
"model_path": "Pika-Labs/Pika-H3-Flex"
|
||||
},
|
||||
"nvfp4-qad": {
|
||||
"model_path": "FastVideo/Pika-H3-Flex-NVFP4-QAD",
|
||||
"nvfp4_dit": "h3_dit"
|
||||
},
|
||||
"nvfp4-qad-fp4-attention": {
|
||||
"model_path": "FastVideo/Pika-H3-Flex-NVFP4-FP4Attn-QAD",
|
||||
"nvfp4_dit": "h3_dit",
|
||||
"nvfp4_attention": "smooth_k+hadamard",
|
||||
"nvfp4_attention_bf16_blocks": "45,49"
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def parse_args() -> argparse.Namespace:
|
||||
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
|
||||
parser.add_argument("--model-path", default="Pika-Labs/Pika-H3-Flex")
|
||||
parser.add_argument("--weights",
|
||||
choices=tuple(WEIGHT_PRESETS),
|
||||
default="bf16",
|
||||
help="bf16: Pika's release; nvfp4-qad: FastVideo's NVFP4 QAD release; "
|
||||
"nvfp4-qad-fp4-attention: NVFP4 QAD with FP4 attention (both need Blackwell)")
|
||||
parser.add_argument("--model-path", help="local model directory or Hub repo (overrides --weights)")
|
||||
parser.add_argument("--task", choices=TASKS, default="t2va")
|
||||
parser.add_argument("--prompt", required=True)
|
||||
parser.add_argument("--image", help="first frame (i2va/fl2va)")
|
||||
@@ -146,10 +169,12 @@ def parse_args() -> argparse.Namespace:
|
||||
choices=tuple(QK_TRANSFORMS),
|
||||
help="FA4 NVFP4 Q/K attention with this exact Q/K preconditioning (default: BF16 attention)")
|
||||
parser.add_argument("--nvfp4-attention-bf16-blocks",
|
||||
default="",
|
||||
help="comma-separated transformer_blocks indices that keep BF16 attention, e.g. 45,49")
|
||||
parser.add_argument("--output", default="outputs/pika_h3_flex")
|
||||
args = parser.parse_args()
|
||||
for name, value in WEIGHT_PRESETS[args.weights].items():
|
||||
if getattr(args, name) is None:
|
||||
setattr(args, name, value)
|
||||
if args.nvfp4_attention_bf16_blocks and not args.nvfp4_attention:
|
||||
parser.error("--nvfp4-attention-bf16-blocks needs --nvfp4-attention")
|
||||
try:
|
||||
@@ -232,7 +257,7 @@ def _apply_nvfp4_attention(args: argparse.Namespace) -> None:
|
||||
return
|
||||
envs.FASTVIDEO_NVFP4_FA4.set(True)
|
||||
envs.FASTVIDEO_NVFP4_FA4_QK_TRANSFORM.set(args.nvfp4_attention)
|
||||
envs.FASTVIDEO_NVFP4_FA4_BF16_BLOCKS.set(args.nvfp4_attention_bf16_blocks)
|
||||
envs.FASTVIDEO_NVFP4_FA4_BF16_BLOCKS.set(args.nvfp4_attention_bf16_blocks or "")
|
||||
envs.setdefault_external("CUTE_DSL_ENABLE_TVM_FFI", "1")
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user