PowerMachine's picture
V5: W8A8 aprimorado (α aprendível via STE) + bug fix apply_w8a8_to_model + 4 hiperparâmetros auto-ajustáveis (T, τ, λ_ent, init_gate) + OOM-Killer fixes + num_layers_hyp=8 fixo
cbf308e verified
Raw
History Blame Contribute Delete
1.18 kB
"""Quantização W8A8 (Lema 3).
V5: módulos novos:
- SmoothQuantW8A8 (α aprendível via STE + sigmoid)
- TrustGuidedW8A8 (redução iterativa de erro, self-contained)
- MultiLayerW8A8Reducer (aplica a múltiplas camadas)
- apply_w8a8_to_model (bug fix crítico — substitui in-place)
"""
from .quantized_linear import (
QuantizedLinear,
quantize_tensor,
apply_w8a8,
apply_w8a8_to_model,
)
from .w8a8_smoothquant import (
SmoothQuantW8A8,
calibrate_smoothquant,
quantize_per_channel_symmetric,
quantize_per_token_symmetric,
STEQuantize,
ste_quantize,
)
from .w8a8_error_reduction import (
W8A8ConvergenceError,
analytic_trust,
W8A8ReductionResult,
TrustGuidedW8A8,
MultiLayerW8A8Reducer,
)
__all__ = [
# V1-V4
"QuantizedLinear",
"quantize_tensor",
"apply_w8a8",
# V5
"apply_w8a8_to_model",
"SmoothQuantW8A8",
"calibrate_smoothquant",
"quantize_per_channel_symmetric",
"quantize_per_token_symmetric",
"STEQuantize",
"ste_quantize",
"W8A8ConvergenceError",
"analytic_trust",
"W8A8ReductionResult",
"TrustGuidedW8A8",
"MultiLayerW8A8Reducer",
]