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

[refactor]: move repository configs and examples off pipeline.experimental

Browse Source 
Every key that repository YAML configs, Python examples, the Dreamverse
backend, and tests passed through pipeline.experimental has a typed
GeneratorConfig field, so they use those fields instead:
engine.attention, engine.compile.regional, pipeline.flow_shift,
pipeline.embedded_cfg_scale, pipeline.dmd_denoising_steps,
pipeline.components.lora_nickname, pipeline.minimax_h3, and
pipeline.longcat. No repository config outside the tests uses
pipeline.experimental any more.

The flat FastVideoArgs keywords that each migrated YAML file and each
basic_fasth3.py option set produces match the previous ones. The golden
snapshots of the two Hunyuan scripts now record embedded_cfg_scale and
flow_shift as floats (6.0 instead of 6), because the typed fields are
floats. The guidance tensor is built as float32 either way. The golden case
untyped_experimental_keys is renamed attention_precision_and_flow_shift with
identical content, and keys_without_typed_fields covers keywords that still
go to pipeline.experimental.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012M91wnVFmPEvJ9h39r5BH7
This commit is contained in:
Davids048
2026-10-02 17:37:22 +00:00
co-authored by Claude Opus 5.5
parent b24396561a
commit ecb15bf9e0
41 changed files with 542 additions and 196 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/inference/basic/basic_dmd_new_api.py
+2 -6
View File
@@ -3,12 +3,12 @@ import time
from fastvideo import VideoGenerator
from fastvideo.api import (
AttentionConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
OffloadConfig,
OutputConfig,
PipelineSelection,
)
OUTPUT_PATH = "video_samples_dmd2_typed"
@@ -29,12 +29,8 @@ def main():
dit=False,
vae=False,
),
attention=AttentionConfig(vsa_sparsity=0.8),
),
# PR 2 still routes a few advanced inference knobs through the
# compatibility bridge until they get first-class typed fields.
pipeline=PipelineSelection(experimental={
"VSA_sparsity": 0.8,
}, ),
)
load_start_time = time.perf_counter()
examples/inference/basic/basic_fasth3.py
+15 -18
View File
@@ -25,11 +25,13 @@ from pathlib import Path
from fastvideo import VideoGenerator
from fastvideo.api import (
AttentionConfig,
CompileConfig,
ComponentConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
MiniMaxH3Config,
OffloadConfig,
OutputConfig,
ParallelismConfig,
@@ -254,23 +256,6 @@ def _execution_backend(args: argparse.Namespace) -> str:
def build_generator_config(args: argparse.Namespace) -> GeneratorConfig:
use_vsa = _uses_vsa(args)
experimental: dict[str, object] = {
"attention_backend": "VIDEO_SPARSE_ATTN_H3" if use_vsa else "FLASH_ATTN",
"inference_torch_compile": args.inference_torch_compile,
"vae_parallel_decode": args.parallel_vae,
"vae_parallel_decode_strategy": "gather",
}
if args.h3_sequential_load is not None:
experimental["h3_sequential_load"] = args.h3_sequential_load
if args.video_decode_backend != "h3-vae":
experimental["video_decode_backend"] = args.video_decode_backend
if args.taeh3_checkpoint is not None:
experimental["taeh3_checkpoint"] = args.taeh3_checkpoint
if use_vsa:
experimental.update({
"VSA_sparsity": args.vsa_sparsity,
"VSA_tile_size": args.vsa_tile_size,
})
return GeneratorConfig(
model_path=args.model_path,
pipeline=PipelineSelection(
@@ -278,7 +263,13 @@ def build_generator_config(args: argparse.Namespace) -> GeneratorConfig:
lora_path=getattr(args, "lora_path", None),
lora_strength=float(getattr(args, "lora_strength", 1.0)),
),
experimental=experimental,
minimax_h3=MiniMaxH3Config(
sequential_load=args.h3_sequential_load,
video_decode_backend=None if args.video_decode_backend == "h3-vae" else args.video_decode_backend,
taeh3_checkpoint=args.taeh3_checkpoint,
vae_parallel_decode=args.parallel_vae,
vae_parallel_decode_strategy="gather",
),
),
engine=EngineConfig(
num_gpus=args.num_gpus,
@@ -297,6 +288,12 @@ def build_generator_config(args: argparse.Namespace) -> GeneratorConfig:
enabled=args.torch_compile,
mode=args.compile_mode,
vae_enabled=args.compile_vae,
regional=args.inference_torch_compile,
),
attention=AttentionConfig(
backend="VIDEO_SPARSE_ATTN_H3" if use_vsa else "FLASH_ATTN",
vsa_sparsity=args.vsa_sparsity if use_vsa else None,
vsa_tile_size=args.vsa_tile_size if use_vsa else None,
),
),
)
examples/inference/basic/basic_fasth3_simplified.py
+10 -13
View File
@@ -17,11 +17,13 @@ sys.path.insert(0, str(FASTVIDEO_ROOT))
from fastvideo import VideoGenerator # noqa: E402
from fastvideo.api import ( # noqa: E402
AttentionConfig,
CompileConfig,
ComponentConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
MiniMaxH3Config,
OffloadConfig,
OutputConfig,
ParallelismConfig,
@@ -57,17 +59,7 @@ def main() -> None:
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,
})
use_vsa = attention_backend == "VIDEO_SPARSE_ATTN_H3"
adapter_path = hf_hub_download(
repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
@@ -81,13 +73,18 @@ def main() -> None:
model_path="MiniMaxAI/MiniMax-H3",
pipeline=PipelineSelection(
components=ComponentConfig(lora_path=adapter_path, lora_strength=1.0),
experimental=experimental,
minimax_h3=MiniMaxH3Config(vae_parallel_decode=True, vae_parallel_decode_strategy="gather"),
),
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),
compile=CompileConfig(vae_enabled=True, regional=attention_backend == "FLASH_ATTN"),
attention=AttentionConfig(
backend=attention_backend,
vsa_sparsity=0.9 if use_vsa else None,
vsa_tile_size=64 if use_vsa else None,
),
),
)
)
examples/inference/basic/basic_fasth3_simplified_profile.py
+10 -13
View File
@@ -11,11 +11,13 @@ from huggingface_hub import hf_hub_download
from fastvideo import VideoGenerator
from fastvideo.api import (
AttentionConfig,
CompileConfig,
ComponentConfig,
EngineConfig,
GenerationRequest,
GeneratorConfig,
MiniMaxH3Config,
OffloadConfig,
OutputConfig,
ParallelismConfig,
@@ -50,17 +52,7 @@ def main() -> None:
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,
})
use_vsa = attention_backend == "VIDEO_SPARSE_ATTN_H3"
adapter_path = hf_hub_download(
repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
@@ -71,13 +63,18 @@ def main() -> None:
model_path="MiniMaxAI/MiniMax-H3",
pipeline=PipelineSelection(
components=ComponentConfig(lora_path=adapter_path, lora_strength=1.0),
experimental=experimental,
minimax_h3=MiniMaxH3Config(vae_parallel_decode=True, vae_parallel_decode_strategy="gather"),
),
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),
compile=CompileConfig(vae_enabled=True, regional=attention_backend == "FLASH_ATTN"),
attention=AttentionConfig(
backend=attention_backend,
vsa_sparsity=0.9 if use_vsa else None,
vsa_tile_size=64 if use_vsa else None,
),
quantization=QuantizationConfig(transformer_quant="MXFP8"),
),
)
examples/inference/basic/basic_fasth3_spark.yaml
+6 -5
View File
@@ -30,12 +30,13 @@ generator:
compile:
enabled: false
vae_enabled: true
regional: true
attention:
backend: VIDEO_SPARSE_ATTN_H3
vsa_sparsity: 0.9
vsa_tile_size: 64
pipeline:
experimental:
attention_backend: VIDEO_SPARSE_ATTN_H3
VSA_sparsity: 0.9
VSA_tile_size: 64
inference_torch_compile: true
minimax_h3:
vae_parallel_decode: true
vae_parallel_decode_strategy: gather
request:
examples/inference/basic/basic_fasth3_spark_pair.yaml
+6 -5
View File
@@ -30,12 +30,13 @@ generator:
compile:
enabled: false
vae_enabled: true
regional: true
attention:
backend: VIDEO_SPARSE_ATTN_H3
vsa_sparsity: 0.9
vsa_tile_size: 64
pipeline:
experimental:
attention_backend: VIDEO_SPARSE_ATTN_H3
VSA_sparsity: 0.9
VSA_tile_size: 64
inference_torch_compile: true
minimax_h3:
vae_parallel_decode: true
vae_parallel_decode_strategy: gather
request:
examples/inference/basic/basic_minimax_h3_t2v.py
+1 -8
View File
@@ -16,7 +16,6 @@ from fastvideo.api import (
OffloadConfig,
OutputConfig,
ParallelismConfig,
PipelineSelection,
SamplingConfig,
)
@@ -82,17 +81,10 @@ def main() -> None:
output_dir = Path(args.output)
output_dir.mkdir(parents=True, exist_ok=True)
# Boot-time run configuration folded into FastVideoArgs (the same
# experimental-dict route basic_fasth3.py uses for the VSA knobs).
experimental: dict[str, object] = {}
if args.inference_torch_compile:
experimental["inference_torch_compile"] = True
execution_backend = args.execution_backend or ("ray" if os.environ.get("RAY_ADDRESS") else "mp")
generator = VideoGenerator.from_config(
GeneratorConfig(
model_path=args.model_path,
pipeline=PipelineSelection(experimental=experimental),
engine=EngineConfig(
num_gpus=args.num_gpus,
execution_backend=execution_backend,
@@ -110,6 +102,7 @@ def main() -> None:
enabled=args.torch_compile,
mode=args.compile_mode,
vae_enabled=args.compile_vae,
regional=True if args.inference_torch_compile else None,
),
),
))