diodiogod
3d7e8dfaa1
Fix Step Audio EditX generation and runtime compatibility Technical details: - Restore the correct Step prompt, token layout, untied audio head, and memory-safe attention behavior - Add shared and dedicated Transformers 4 runtime routing with worker progress reporting - Preserve runtime selection across TTS, SRT, caching, and inline audio editing - Fix character voice handling, empty audio validation, sample-rate interface, and deprecated dependency calls
131 lines
4.2 KiB
Python
131 lines
4.2 KiB
Python
from __future__ import annotations
|
|
|
|
"""
|
|
Named isolated-runtime profiles.
|
|
|
|
Profiles describe which external Python runtime should be used for a fragile
|
|
engine family. Paths stay user-configurable; these names are the stable keys.
|
|
"""
|
|
|
|
from dataclasses import dataclass, field
|
|
from typing import Dict, List, Optional
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class RuntimeProfile:
|
|
name: str
|
|
engine_names: List[str]
|
|
python_path_hint: Optional[str] = None
|
|
description: str = ""
|
|
runtime_mode: str = "isolated"
|
|
env_vars: Dict[str, str] = field(default_factory=dict)
|
|
inherit_base_site_packages: bool = False
|
|
pip_packages: List[str] = field(default_factory=list)
|
|
pip_packages_no_deps: List[str] = field(default_factory=list)
|
|
|
|
|
|
_VIBEVOICE_T4_PACKAGES = [
|
|
"numpy>=1.26.4,<2.3.0",
|
|
"soundfile>=0.12.0",
|
|
"omegaconf>=2.3.0",
|
|
"transformers>=4.51.3,<=4.57.3",
|
|
"kernels>=0.6.1,<=0.9",
|
|
"accelerate",
|
|
"requests",
|
|
"av",
|
|
"bitsandbytes>=0.47.0",
|
|
"safetensors>=0.6.2",
|
|
"sentencepiece>=0.2.1",
|
|
"tqdm",
|
|
"scipy",
|
|
"librosa",
|
|
"llvmlite>=0.40.0",
|
|
"numba>=0.57.0",
|
|
"diffusers",
|
|
"ml-collections",
|
|
"absl-py",
|
|
"conformer>=0.3.2",
|
|
"x-transformers",
|
|
]
|
|
|
|
_QWEN3_T4_PACKAGES = [
|
|
"numpy>=1.26.4,<2.3.0",
|
|
"soundfile>=0.12.0",
|
|
"librosa",
|
|
"transformers>=4.51.3,<=4.57.3",
|
|
"kernels>=0.6.1,<=0.9",
|
|
"accelerate",
|
|
"huggingface-hub<1.0",
|
|
"safetensors>=0.6.2",
|
|
]
|
|
|
|
|
|
RUNTIME_PROFILES: Dict[str, RuntimeProfile] = {
|
|
"vibevoice_transformers4_shared": RuntimeProfile(
|
|
name="vibevoice_transformers4_shared",
|
|
engine_names=["vibevoice", "step_audio_editx"],
|
|
python_path_hint="runtimes/shared_legacy_t4/Scripts/python.exe",
|
|
description="Shared legacy Transformers 4 runtime for VibeVoice/Kugel, Step Audio EditX, and similar engines.",
|
|
inherit_base_site_packages=True,
|
|
pip_packages=list(_VIBEVOICE_T4_PACKAGES),
|
|
pip_packages_no_deps=[
|
|
"git+https://github.com/FushionHub/VibeVoice.git",
|
|
],
|
|
),
|
|
"vibevoice_transformers4_dedicated": RuntimeProfile(
|
|
name="vibevoice_transformers4_dedicated",
|
|
engine_names=["vibevoice"],
|
|
python_path_hint="runtimes/vibevoice_t4_dedicated/Scripts/python.exe",
|
|
description="Dedicated legacy Transformers 4 runtime reserved for VibeVoice/Kugel only.",
|
|
inherit_base_site_packages=True,
|
|
pip_packages=list(_VIBEVOICE_T4_PACKAGES),
|
|
pip_packages_no_deps=[
|
|
"git+https://github.com/FushionHub/VibeVoice.git",
|
|
],
|
|
),
|
|
"qwen3_tts_transformers4_dedicated": RuntimeProfile(
|
|
name="qwen3_tts_transformers4_dedicated",
|
|
engine_names=["qwen3_tts"],
|
|
python_path_hint="runtimes/qwen3_tts_t4_dedicated/Scripts/python.exe",
|
|
description="Dedicated legacy Transformers 4 runtime reserved for Qwen3-TTS.",
|
|
inherit_base_site_packages=True,
|
|
pip_packages=list(_QWEN3_T4_PACKAGES),
|
|
),
|
|
"step_audio_editx_transformers4": RuntimeProfile(
|
|
name="step_audio_editx_transformers4",
|
|
engine_names=["step_audio_editx"],
|
|
python_path_hint="runtimes/step_audio_editx_t4/Scripts/python.exe",
|
|
description="Dedicated Step Audio EditX runtime matching its released PyTorch stack.",
|
|
inherit_base_site_packages=True,
|
|
pip_packages=[
|
|
"transformers==4.57.6",
|
|
"huggingface-hub<1.0",
|
|
"tokenizers<0.23",
|
|
"accelerate>=1.10.1",
|
|
"safetensors>=0.6.2",
|
|
],
|
|
),
|
|
"moss_tts_transformers5": RuntimeProfile(
|
|
name="moss_tts_transformers5",
|
|
engine_names=["moss_tts"],
|
|
python_path_hint="runtimes/moss_tts_t5/Scripts/python.exe",
|
|
description="Dedicated MOSS-TTS runtime isolated from VibeVoice/Qwen.",
|
|
),
|
|
"higgs_audio_embedded": RuntimeProfile(
|
|
name="higgs_audio_embedded",
|
|
engine_names=["higgs_audio"],
|
|
description="Placeholder profile for future Higgs isolation if needed.",
|
|
runtime_mode="embedded",
|
|
),
|
|
}
|
|
|
|
|
|
def get_runtime_profile(profile_name: Optional[str]) -> Optional[RuntimeProfile]:
|
|
if not profile_name:
|
|
return None
|
|
return RUNTIME_PROFILES.get(profile_name)
|
|
|
|
|
|
def list_runtime_profiles() -> List[RuntimeProfile]:
|
|
return list(RUNTIME_PROFILES.values())
|