nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[feat] add MXFP8 support on H3 (#1796)

Browse Source
This commit is contained in:
Junda Su
2026-09-05 17:08:58 -07:00
committed by GitHub
parent e19913f6e9
commit 4e52f47d1e
15 changed files with 1350 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/inference/basic/basic_fasth3_simplified.py
+147
View File
@@ -0,0 +1,147 @@
# SPDX-License-Identifier: Apache-2.0
"""Generate 5-second FastH3 videos with a supported Preview adapter."""
from __future__ import annotations
import argparse
import os
import statistics
import sys
import time
from pathlib import Path
from huggingface_hub import hf_hub_download
FASTVIDEO_ROOT = Path(__file__).resolve().parents[3]
sys.path.insert(0, str(FASTVIDEO_ROOT))
from fastvideo import VideoGenerator # noqa: E402
from fastvideo.api import ( # noqa: E402
CompileConfig,
ComponentConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
OffloadConfig,
OutputConfig,
ParallelismConfig,
PipelineSelection,
SamplingConfig,
)
# FastVideo exposes these three backend switches only through environment variables.
os.environ.update({
"FASTVIDEO_FA4": "1",
"FASTVIDEO_MINIMAX_H3_FUSIONS": "all",
"FASTVIDEO_VSA_SM100A": "0",
})
os.environ.pop("FASTVIDEO_INFERENCE_TORCH_COMPILE", None)
VARIANT_BACKENDS = {
"dense-datafree": "FLASH_ATTN",
"vsa-datafree": "VIDEO_SPARSE_ATTN_H3",
"vsa-synthetic-step1300": "VIDEO_SPARSE_ATTN_H3",
"vsa-synthetic-step1900": "VIDEO_SPARSE_ATTN_H3",
}
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("variant", choices=VARIANT_BACKENDS)
parser.add_argument("--prompt", required=True)
return parser.parse_args()
def main() -> None:
"""Run one compile warmup and three measured generations with one fixed recipe."""
args = parse_args()
attention_backend = VARIANT_BACKENDS[args.variant]
experimental = {
"attention_backend": attention_backend,
"inference_torch_compile": attention_backend == "FLASH_ATTN",
"vae_parallel_decode": True,
"vae_parallel_decode_strategy": "gather",
}
if attention_backend == "VIDEO_SPARSE_ATTN_H3":
experimental.update({
"VSA_sparsity": 0.9,
"VSA_tile_size": 64,
})
adapter_path = hf_hub_download(
repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
filename=f"{args.variant}/adapter_model.safetensors",
)
output_dir = FASTVIDEO_ROOT / "outputs/fasth3_lora_preview" / args.variant
output_dir.mkdir(parents=True, exist_ok=True)
generator = VideoGenerator.from_config(
GeneratorConfig(
model_path="MiniMaxAI/MiniMax-H3",
pipeline=PipelineSelection(
components=ComponentConfig(lora_path=adapter_path, lora_strength=1.0),
experimental=experimental,
),
engine=EngineConfig(
num_gpus=4,
parallelism=ParallelismConfig(tp_size=1, sp_size=4),
offload=OffloadConfig(dit=False, dit_layerwise=False),
compile=CompileConfig(vae_enabled=True),
),
)
)
generation_count = 4
warmup_count = 1
measured_count = generation_count - warmup_count
measured_seconds: list[float] = []
print(f"Variant: {args.variant} ({attention_backend})")
print(f"Output directory: {output_dir}")
try:
for generation_index in range(generation_count):
measured = generation_index >= warmup_count
if measured:
measured_index = generation_index - warmup_count + 1
label = f"measured {measured_index}/{measured_count}"
output_path = output_dir / f"fasth3_all_run_{measured_index:02d}.mp4"
else:
warmup_index = generation_index + 1
label = f"warmup {warmup_index}/{warmup_count}"
output_path = output_dir / f"_fasth3_warmup_{warmup_index:02d}.mp4"
request = GenerationRequest(
prompt=args.prompt,
negative_prompt="",
sampling=SamplingConfig(
height=768,
width=1344,
num_frames=345, # <-- Change video length: 5sec: 124; 10sec: 243; 15sec: 345.
fps=24,
num_inference_steps=5,
guidance_scale=1.0,
batch_cfg=False,
seed=1000,
),
output=OutputConfig(
output_path=str(output_path),
save_video=True,
return_frames=False,
),
)
started = time.perf_counter()
result = generator.generate(request)
elapsed = time.perf_counter() - started
if measured:
measured_seconds.append(elapsed)
suffix = "" if measured else " (excluded from median)"
print(f"[{label}] {result.video_path or output_path}: {elapsed:.3f}s{suffix}")
finally:
generator.shutdown()
print(f"Median E2E wall time: {statistics.median(measured_seconds):.3f}s")
if __name__ == "__main__":
main()
examples/inference/basic/basic_fasth3_simplified.sh
+10
View File
@@ -0,0 +1,10 @@
#!/usr/bin/env bash
PROMPT="integrated_multimodal_description: A red fox runs through fresh snow at dawn. overall_soundscape: Fast pawsteps in snow, winter wind, and distant birds."
# Options: dense-datafree, vsa-datafree, vsa-synthetic-step1300, vsa-synthetic-step1900
VARIANT="vsa-datafree"
exec python "$(dirname "$0")/basic_fasth3_simplified.py" \
"$VARIANT" \
--prompt "$PROMPT"
examples/inference/basic/basic_fasth3_simplified_profile.py
+126
View File
@@ -0,0 +1,126 @@
# SPDX-License-Identifier: Apache-2.0
"""Profile one FastH3 Preview generation with NVTX after warmup."""
from __future__ import annotations
import argparse
from pathlib import Path
import torch
from huggingface_hub import hf_hub_download
from fastvideo import VideoGenerator
from fastvideo.api import (
CompileConfig,
ComponentConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
OffloadConfig,
OutputConfig,
ParallelismConfig,
PipelineSelection,
QuantizationConfig,
SamplingConfig,
)
from fastvideo.profiler import nvtx_range
VARIANT_BACKENDS = {
"dense-datafree": "FLASH_ATTN",
"vsa-datafree": "VIDEO_SPARSE_ATTN_H3",
"vsa-synthetic-step1300": "VIDEO_SPARSE_ATTN_H3",
"vsa-synthetic-step1900": "VIDEO_SPARSE_ATTN_H3",
}
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("variant", choices=VARIANT_BACKENDS)
parser.add_argument("--prompt", required=True)
parser.add_argument("--output", required=True)
parser.add_argument("--num-frames", type=int, default=345)
parser.add_argument("--warmup-runs", type=int, default=3)
return parser.parse_args()
def main() -> None:
"""Warm the selected FastH3 recipe, then profile one identical generation."""
args = parse_args()
output_dir = Path(args.output)
output_dir.mkdir(parents=True, exist_ok=True)
attention_backend = VARIANT_BACKENDS[args.variant]
experimental = {
"attention_backend": attention_backend,
"inference_torch_compile": attention_backend == "FLASH_ATTN",
"vae_parallel_decode": True,
"vae_parallel_decode_strategy": "gather",
}
if attention_backend == "VIDEO_SPARSE_ATTN_H3":
experimental.update({
"VSA_sparsity": 0.9,
"VSA_tile_size": 64,
})
adapter_path = hf_hub_download(
repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
filename=f"{args.variant}/adapter_model.safetensors",
)
generator = VideoGenerator.from_config(
GeneratorConfig(
model_path="MiniMaxAI/MiniMax-H3",
pipeline=PipelineSelection(
components=ComponentConfig(lora_path=adapter_path, lora_strength=1.0),
experimental=experimental,
),
engine=EngineConfig(
num_gpus=4,
parallelism=ParallelismConfig(tp_size=1, sp_size=4),
offload=OffloadConfig(dit=False, dit_layerwise=False),
compile=CompileConfig(vae_enabled=True),
quantization=QuantizationConfig(transformer_quant="MXFP8"),
),
)
)
request = GenerationRequest(
prompt=args.prompt,
negative_prompt="",
sampling=SamplingConfig(
height=768,
width=1344,
num_frames=args.num_frames,
fps=24,
num_inference_steps=5,
guidance_scale=1.0,
batch_cfg=False,
seed=1000,
),
output=OutputConfig(
output_path=str(output_dir / "fasth3_profile.mp4"),
save_video=True,
return_frames=False,
),
)
try:
for warmup_index in range(args.warmup_runs):
warmup_result = generator.generate(request)
print(f"Warmup {warmup_index + 1}/{args.warmup_runs}: {warmup_result.video_path}")
torch.cuda.profiler.start()
try:
with nvtx_range("fasth3.profiled_generation"):
measured_result = generator.generate(request)
finally:
torch.cuda.profiler.stop()
print(f"Profiled output: {measured_result.video_path}")
if measured_result.generation_time is not None:
print(f"Generation time: {measured_result.generation_time:.2f}s")
finally:
generator.shutdown()
if __name__ == "__main__":
main()
examples/inference/basic/basic_fasth3_simplified_profile.sh
Executable
+53
View File
@@ -0,0 +1,53 @@
#!/usr/bin/env bash
# Capture one FastH3 Preview generation with CUDA and FastVideo NVTX ranges.
set -euo pipefail
SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)"
WORKTREE_ROOT="$(cd -- "${SCRIPT_DIR}/../../.." && pwd)"
WORKSPACE_ROOT="$(cd -- "${WORKTREE_ROOT}/.." && pwd)"
PYTHON_BIN="${WORKSPACE_ROOT}/.venv-fv/bin/python"
NSYS_BIN=/usr/local/cuda/bin/nsys
PROFILE_SCRIPT="${SCRIPT_DIR}/basic_fasth3_simplified_profile.py"
export CUDA_VISIBLE_DEVICES=0,1,2,3
export FASTVIDEO_FA4=1
export FASTVIDEO_INFERENCE_TORCH_COMPILE=0
export FASTVIDEO_MINIMAX_H3_FUSIONS=all
export FASTVIDEO_NVTX_PROFILE=1
export FASTVIDEO_VSA_SM100A=0
export PYTHONUNBUFFERED=1
export PYTHONPATH="${WORKTREE_ROOT}${PYTHONPATH:+:${PYTHONPATH}}"
PROMPT="integrated_multimodal_description: A red fox runs through fresh snow at dawn. overall_soundscape: Fast pawsteps in snow, winter wind, and distant birds."
# Options: dense-datafree, vsa-datafree, vsa-synthetic-step1300, vsa-synthetic-step1900
VARIANT="vsa-datafree"
NUM_FRAMES=345
WARMUP_RUNS=3
PROFILE_ID="${VARIANT}_4gpu_sp4_${NUM_FRAMES}frames"
RUN_ID="$(date -u +%Y%m%dT%H%M%SZ)-$$"
RESULT_DIR="${WORKTREE_ROOT}/runs/fasth3_lora_preview_profile/${PROFILE_ID}/${RUN_ID}"
MEDIA_DIR="${RESULT_DIR}/media"
RUN_LOG="${RESULT_DIR}/run.log"
mkdir -p "${MEDIA_DIR}"
"${NSYS_BIN}" profile \
--trace=cuda,nvtx \
--sample=none \
--cpuctxsw=none \
--capture-range=cudaProfilerApi \
--capture-range-end=stop \
--output="${RESULT_DIR}/${PROFILE_ID}" \
"${PYTHON_BIN}" "${PROFILE_SCRIPT}" \
"${VARIANT}" \
--num-frames "${NUM_FRAMES}" \
--warmup-runs "${WARMUP_RUNS}" \
--prompt "${PROMPT}" \
--output "${MEDIA_DIR}" \
2>&1 | tee "${RUN_LOG}"
printf 'RESULT_DIR=%s\n' "${RESULT_DIR}"
printf 'NSYS_REPORT=%s\n' "${RESULT_DIR}/${PROFILE_ID}.nsys-rep"
printf 'MEDIA_DIR=%s\n' "${MEDIA_DIR}"