141 lines
4.8 KiB
Python
141 lines
4.8 KiB
Python
#!/usr/bin/env python3
|
|
"""Streaming quantization: process one shard at a time to avoid OOM."""
|
|
|
|
import argparse
|
|
import gc
|
|
import torch
|
|
from pathlib import Path
|
|
from transformers import AutoConfig, AutoModelForCausalLM
|
|
from bitsandbytes.nn import Linear4bit
|
|
from torch import nn
|
|
|
|
|
|
def streaming_quantize(model_path, output_path):
|
|
"""Quantize model by processing one shard at a time, using both GPUs."""
|
|
|
|
print(f"Loading config from: {model_path}")
|
|
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
|
|
|
|
# Find all safetensors files
|
|
import glob
|
|
safetensors_files = sorted(glob.glob(f"{model_path}/*.safetensors"))
|
|
print(f"Found {len(safetensors_files)} safetensors files")
|
|
|
|
# Create output directory
|
|
output_path = Path(output_path)
|
|
output_path.mkdir(parents=True, exist_ok=True)
|
|
|
|
# Process each shard
|
|
for shard_idx, shard_file in enumerate(safetensors_files):
|
|
print(f"\n{'='*60}")
|
|
print(f"Processing shard {shard_idx + 1}/{len(safetensors_files)}: {shard_file}")
|
|
print(f"{'='*60}")
|
|
|
|
# Load shard directly to GPU 0
|
|
print(" Loading shard to GPU 0...")
|
|
shard_state_dict = torch.load(shard_file, map_location="cuda:0", weights_only=False)
|
|
|
|
# Quantize Linear layers in this shard using both GPUs
|
|
print(" Quantizing Linear layers (both GPUs)...")
|
|
quantized_keys = 0
|
|
|
|
# Get all weight tensors
|
|
weight_keys = [k for k, v in shard_state_dict.items() if 'weight' in k and v.dim() == 2]
|
|
|
|
# Distribute between GPUs
|
|
gpu0_keys = weight_keys[::2] # Even indices
|
|
gpu1_keys = weight_keys[1::2] # Odd indices
|
|
|
|
# Quantize on GPU 0 (shard already on GPU 0)
|
|
print(" GPU 0: Quantizing...", end=" ")
|
|
for key in gpu0_keys:
|
|
tensor = shard_state_dict[key]
|
|
in_features = tensor.size(1)
|
|
out_features = tensor.size(0)
|
|
|
|
dummy_linear = Linear4bit(
|
|
in_features,
|
|
out_features,
|
|
bias=False,
|
|
compute_dtype=torch.float16,
|
|
quant_type='nf4',
|
|
)
|
|
|
|
with torch.no_grad():
|
|
dummy_linear.weight = nn.Parameter(tensor.clone())
|
|
_ = dummy_linear.weight.quant_state
|
|
|
|
shard_state_dict[key] = dummy_linear.weight
|
|
quantized_keys += 1
|
|
print(f"✓ {len(gpu0_keys)} layers")
|
|
|
|
# Move shard to GPU 1 for second half
|
|
print(" Moving to GPU 1...")
|
|
shard_state_dict = {k: v.to("cuda:1") if isinstance(v, torch.Tensor) else v for k, v in shard_state_dict.items()}
|
|
|
|
# Quantize on GPU 1
|
|
print(" GPU 1: Quantizing...", end=" ")
|
|
for key in gpu1_keys:
|
|
tensor = shard_state_dict[key]
|
|
in_features = tensor.size(1)
|
|
out_features = tensor.size(0)
|
|
|
|
dummy_linear = Linear4bit(
|
|
in_features,
|
|
out_features,
|
|
bias=False,
|
|
compute_dtype=torch.float16,
|
|
quant_type='nf4',
|
|
)
|
|
|
|
with torch.no_grad():
|
|
dummy_linear.weight = nn.Parameter(tensor.clone())
|
|
_ = dummy_linear.weight.quant_state
|
|
|
|
shard_state_dict[key] = dummy_linear.weight
|
|
quantized_keys += 1
|
|
print(f"✓ {len(gpu1_keys)} layers")
|
|
|
|
print(f" ✓ Total: {quantized_keys} layers quantized")
|
|
|
|
# Save quantized shard
|
|
shard_name = f"model_shard_{shard_idx:05d}.safetensors"
|
|
shard_path = output_path / shard_name
|
|
print(f" Saving to: {shard_path}")
|
|
torch.save(shard_state_dict, shard_path)
|
|
|
|
# Free memory
|
|
del shard_state_dict
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
|
|
# Save config
|
|
print(f"\n{'='*60}")
|
|
print("Saving model config...")
|
|
config.save_pretrained(output_path)
|
|
print(f"✓ Model saved to: {output_path}")
|
|
|
|
# Free memory
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
|
|
print("\n✓ Streaming quantization complete!")
|
|
print(f" Used both GPUs in parallel for faster quantization")
|
|
|
|
|
|
def main():
|
|
parser = argparse.ArgumentParser(description="Streaming quantization")
|
|
parser.add_argument("--model-path", type=str,
|
|
default="/data/models/Ornith-1.0-35B",
|
|
help="Path to bf16 model")
|
|
parser.add_argument("--output-path", type=str,
|
|
default="/data/models/Ornith-1.0-35B-streaming-4bit",
|
|
help="Output path for quantized model")
|
|
args = parser.parse_args()
|
|
|
|
streaming_quantize(args.model_path, args.output_path)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|