File size: 3,795 Bytes
aad6de7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | """Frozen source-specific reference-storage contract; no native FP4 runtime."""
import math
from pathlib import PurePosixPath
from typing import Literal
from pydantic import BaseModel,ConfigDict,Field,model_validator
class Strict(BaseModel):
model_config=ConfigDict(extra='forbid',frozen=True)
class FileDigest(Strict):
sha256:str=Field(pattern=r'^[0-9a-f]{64}$')
size_bytes:int=Field(ge=0)
class Tensor(Strict):
shape:tuple[int,...]
dtype:Literal['bfloat16','float32','float16','int64','int32','bool']
file:str
method:Literal['preserve','nvfp4']
global_scale:float|None=Field(default=None,gt=0,allow_inf_nan=False)
group_size:int|None=None
@model_validator(mode='after')
def layout(self):
if any(x<0 for x in self.shape):raise ValueError('Negative shape')
if self.method=='preserve':
if self.global_scale is not None or self.group_size is not None:raise ValueError('Protected tensor has quantization fields')
elif len(self.shape)!=2 or min(self.shape)<1 or self.shape[1]%16 or self.group_size!=16 or self.global_scale is None or self.dtype!='bfloat16':
raise ValueError('Invalid NVFP4 BF16-source matrix')
relative=PurePosixPath(self.file)
if relative.is_absolute() or '..' in relative.parts or '\\' in self.file:raise ValueError('Unsafe tensor file')
return self
class Component(Strict):
class_name:Literal['QwenImage21Transformer2DModel','MiniMaxH3Transformer3DModel']
tensors:dict[str,Tensor]
quantized_matrices:int=Field(gt=0)
preserved_tensors:int=Field(ge=0)
@model_validator(mode='after')
def inventory(self):
selected=sum(x.method=='nvfp4' for x in self.tensors.values())
if selected!=self.quantized_matrices or len(self.tensors)-selected!=self.preserved_tensors:raise ValueError('Inventory counts mismatch')
return self
class DiffusionManifest(Strict):
schema_version:Literal['axq-diffusion.reference/1']='axq-diffusion.reference/1'
format:Literal['nvfp4']='nvfp4'
status:Literal['development']='development'
source_model:Literal['Qwen/Qwen-Image-2.1-Turbo','MiniMaxAI/MiniMax-H3']
source_revision:str=Field(pattern=r'^[0-9a-f]{40}$')
backend:Literal['numpy-reference']='numpy-reference'
execution:Literal['reconstruct-bfloat16-diffusers']='reconstruct-bfloat16-diffusers'
native_low_precision_verified:Literal[False]=False
quality_certified:Literal[False]=False
source_files:dict[str,FileDigest]
components:dict[str,Component]
files:dict[str,FileDigest]
@model_validator(mode='after')
def binding(self):
source={
'Qwen/Qwen-Image-2.1-Turbo':('fc13ba1986fcb905854dc8de2f52eee395f49b5b',{'transformer'},'QwenImage21Transformer2DModel'),
'MiniMaxAI/MiniMax-H3':('42ed227ee7df40d41602854ae760620d6eb651fe',{'transformer','transformer_ref'},'MiniMaxH3Transformer3DModel'),
}
revision,components,class_name=source[self.source_model]
if self.source_revision!=revision or set(self.components)!=components:raise ValueError('Source pin/component identity mismatch')
if not self.files or not self.source_files:raise ValueError('Missing bound files')
for name in [*self.files,*self.source_files]:
relative=PurePosixPath(name)
if relative.is_absolute() or '..' in relative.parts or '\\' in name:raise ValueError('Unsafe manifest member')
for name,spec in self.components.items():
if spec.class_name!=class_name:raise ValueError('Wrong component class')
for tensor in spec.tensors.values():
if tensor.file not in self.files or not tensor.file.startswith(name+'/'):raise ValueError('Unbound tensor member')
return self
|