mirror of
https://github.com/MLSysBook/TinyTorch.git
synced 2026-03-22 16:39:23 -05:00
96880b3133
Re-exported all modules after restructuring: - Updated _modidx.py with new module locations - Removed outdated autogeneration headers - Updated all core modules (tensor, autograd, layers, etc.) - Updated optimization modules (quantization, compression, etc.) - Updated TITO commands for new structure Changes include: - 24 tinytorch/ module files - 24 tito/ command and core files - Updated references from modules/source/ to modules/ All modules re-exported via nbdev from their new locations.
109 lines
4.2 KiB
Python
Generated
109 lines
4.2 KiB
Python
Generated
# AUTOGENERATED! DO NOT EDIT! File to edit: ../../modules/source/16_quantization/quantization_dev.ipynb.
|
||
|
||
# %% auto 0
|
||
__all__ = ['QuantizationComplete', 'quantize_int8', 'dequantize_int8', 'quantize_model']
|
||
|
||
# %% ../../modules/source/16_quantization/quantization_dev.ipynb 3
|
||
import numpy as np
|
||
import time
|
||
from typing import Tuple, Dict, List, Optional
|
||
import warnings
|
||
|
||
# Import dependencies from other modules
|
||
from ..core.tensor import Tensor
|
||
from ..core.layers import Linear
|
||
from ..core.activations import ReLU
|
||
|
||
print("✅ Quantization module imports complete")
|
||
|
||
# %% ../../modules/source/16_quantization/quantization_dev.ipynb 34
|
||
class QuantizationComplete:
|
||
"""
|
||
Complete quantization system for milestone use.
|
||
|
||
Provides INT8 quantization with calibration for 4× memory reduction.
|
||
"""
|
||
|
||
@staticmethod
|
||
def quantize_tensor(tensor: Tensor) -> Tuple[Tensor, float, int]:
|
||
"""Quantize FP32 tensor to INT8."""
|
||
data = tensor.data
|
||
min_val = float(np.min(data))
|
||
max_val = float(np.max(data))
|
||
|
||
if abs(max_val - min_val) < 1e-8:
|
||
return Tensor(np.zeros_like(data, dtype=np.int8)), 1.0, 0
|
||
|
||
scale = (max_val - min_val) / 255.0
|
||
zero_point = int(np.round(-128 - min_val / scale))
|
||
zero_point = int(np.clip(zero_point, -128, 127))
|
||
|
||
quantized_data = np.round(data / scale + zero_point)
|
||
quantized_data = np.clip(quantized_data, -128, 127).astype(np.int8)
|
||
|
||
return Tensor(quantized_data), scale, zero_point
|
||
|
||
@staticmethod
|
||
def dequantize_tensor(q_tensor: Tensor, scale: float, zero_point: int) -> Tensor:
|
||
"""Dequantize INT8 tensor back to FP32."""
|
||
dequantized_data = (q_tensor.data.astype(np.float32) - zero_point) * scale
|
||
return Tensor(dequantized_data)
|
||
|
||
@staticmethod
|
||
def quantize_model(model, calibration_data: Optional[List[Tensor]] = None) -> Dict[str, any]:
|
||
"""
|
||
Quantize all Linear layers in a model.
|
||
|
||
Returns dictionary with quantization info and memory savings.
|
||
"""
|
||
quantized_layers = {}
|
||
original_size = 0
|
||
quantized_size = 0
|
||
|
||
# Iterate through model parameters
|
||
if hasattr(model, 'parameters'):
|
||
for i, param in enumerate(model.parameters()):
|
||
param_size = param.data.nbytes
|
||
original_size += param_size
|
||
|
||
# Quantize parameter
|
||
q_param, scale, zp = QuantizationComplete.quantize_tensor(param)
|
||
quantized_size += q_param.data.nbytes
|
||
|
||
quantized_layers[f'param_{i}'] = {
|
||
'quantized': q_param,
|
||
'scale': scale,
|
||
'zero_point': zp,
|
||
'original_shape': param.data.shape
|
||
}
|
||
|
||
return {
|
||
'quantized_layers': quantized_layers,
|
||
'original_size_mb': original_size / (1024 * 1024),
|
||
'quantized_size_mb': quantized_size / (1024 * 1024),
|
||
'compression_ratio': original_size / quantized_size if quantized_size > 0 else 1.0
|
||
}
|
||
|
||
@staticmethod
|
||
def compare_models(original_model, quantized_info: Dict) -> Dict[str, float]:
|
||
"""Compare memory usage between original and quantized models."""
|
||
return {
|
||
'original_mb': quantized_info['original_size_mb'],
|
||
'quantized_mb': quantized_info['quantized_size_mb'],
|
||
'compression_ratio': quantized_info['compression_ratio'],
|
||
'memory_saved_mb': quantized_info['original_size_mb'] - quantized_info['quantized_size_mb']
|
||
}
|
||
|
||
# Convenience functions for backward compatibility
|
||
def quantize_int8(tensor: Tensor) -> Tuple[Tensor, float, int]:
|
||
"""Quantize FP32 tensor to INT8."""
|
||
return QuantizationComplete.quantize_tensor(tensor)
|
||
|
||
def dequantize_int8(q_tensor: Tensor, scale: float, zero_point: int) -> Tensor:
|
||
"""Dequantize INT8 tensor back to FP32."""
|
||
return QuantizationComplete.dequantize_tensor(q_tensor, scale, zero_point)
|
||
|
||
def quantize_model(model, calibration_data: Optional[List[Tensor]] = None) -> Dict[str, any]:
|
||
"""Quantize entire model to INT8."""
|
||
return QuantizationComplete.quantize_model(model, calibration_data)
|