97 lines
3.0 KiB
Python
97 lines
3.0 KiB
Python
#!/usr/bin/env python3
|
|
"""True streaming 4-bit NF4 quantization - one shard at a time."""
|
|
|
|
import argparse
|
|
import gc
|
|
import torch
|
|
from pathlib import Path
|
|
from safetensors.torch import load_file, save_file
|
|
from bitsandbytes.nn import Linear4bit
|
|
from torch import nn
|
|
from transformers import AutoConfig
|
|
|
|
|
|
def quantize_tensor_to_4bit(tensor: torch.Tensor) -> torch.Tensor:
|
|
"""Quantize a single 2D weight tensor to 4-bit NF4."""
|
|
in_features = tensor.size(1)
|
|
out_features = tensor.size(0)
|
|
|
|
# Create a temporary Linear4bit layer
|
|
linear_4bit = Linear4bit(
|
|
in_features,
|
|
out_features,
|
|
bias=False,
|
|
compute_dtype=torch.bfloat16,
|
|
quant_type="nf4",
|
|
).to(tensor.device)
|
|
|
|
with torch.no_grad():
|
|
linear_4bit.weight = nn.Parameter(tensor.clone())
|
|
# Force quantization to happen
|
|
_ = linear_4bit.weight.quant_state
|
|
|
|
return linear_4bit.weight
|
|
|
|
|
|
def streaming_quantize(model_path: str, output_path: str):
|
|
print(f"Streaming quantization of: {model_path}")
|
|
output_path = Path(output_path)
|
|
output_path.mkdir(parents=True, exist_ok=True)
|
|
|
|
# Load config
|
|
config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
|
|
|
|
# Find all shards
|
|
import glob
|
|
shards = sorted(glob.glob(f"{model_path}/*.safetensors"))
|
|
print(f"Found {len(shards)} shards\n")
|
|
|
|
for idx, shard_path in enumerate(shards):
|
|
print(f"[{idx+1}/{len(shards)}] Processing {Path(shard_path).name}...")
|
|
|
|
# Load shard to GPU 0
|
|
state_dict = load_file(shard_path, device="cuda:0")
|
|
|
|
# Find all 2D weight tensors (Linear layers)
|
|
weight_keys = [k for k, v in state_dict.items()
|
|
if "weight" in k and isinstance(v, torch.Tensor) and v.dim() == 2]
|
|
|
|
print(f" Found {len(weight_keys)} weight tensors to quantize")
|
|
|
|
# Quantize
|
|
for key in weight_keys:
|
|
try:
|
|
state_dict[key] = quantize_tensor_to_4bit(state_dict[key])
|
|
except Exception as e:
|
|
print(f" Warning: Failed to quantize {key}: {e}")
|
|
|
|
# Move to CPU and save
|
|
state_dict = {k: v.cpu() if isinstance(v, torch.Tensor) else v
|
|
for k, v in state_dict.items()}
|
|
|
|
out_shard = output_path / f"model-{idx:05d}-of-{len(shards):05d}.safetensors"
|
|
save_file(state_dict, out_shard)
|
|
|
|
print(f" Saved → {out_shard.name}")
|
|
|
|
# Cleanup
|
|
del state_dict
|
|
gc.collect()
|
|
torch.cuda.empty_cache()
|
|
|
|
# Save config and index
|
|
config.save_pretrained(output_path)
|
|
|
|
# Create model.safetensors.index.json if needed (for multi-shard)
|
|
print(f"\n✅ Streaming quantization complete!")
|
|
print(f" Output: {output_path}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--model-path", type=str, default="/data/models/Ornith-1.0-35B")
|
|
parser.add_argument("--output-path", type=str, default="/data/models/Ornith-1.0-35B-4bit-streaming")
|
|
args = parser.parse_args()
|
|
|
|
streaming_quantize(args.model_path, args.output_path)
|