Vezora commited on
Commit
a4df890
Β·
verified Β·
1 Parent(s): 6ebf78b

Upload quant.py

Browse files
Files changed (1) hide show
  1. quant/quant.py +84 -0
quant/quant.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from transformers import AutoTokenizer
3
+ from llmcompressor.transformers import SparseAutoModelForCausalLM
4
+ from llmcompressor.transformers import oneshot
5
+ from llmcompressor.modifiers.quantization import QuantizationModifier
6
+
7
+ def get_user_input():
8
+ """Get model configuration from user input"""
9
+ print("\n=== Model Quantization Configuration ===")
10
+
11
+ while True:
12
+ model_id = input("\nEnter the HuggingFace model ID (e.g., meta-llama/Llama-2-7b-chat-hf): ").strip()
13
+ if model_id:
14
+ break
15
+ print("Model ID cannot be empty. Please try again.")
16
+
17
+ return model_id
18
+
19
+ def quantize_model_fp8(model_id):
20
+ """
21
+ Quantize a model to FP8 Dynamic format using llm-compressor on CPU.
22
+
23
+ Args:
24
+ model_id (str): HuggingFace model ID
25
+ """
26
+ try:
27
+ print(f"\nLoading model and tokenizer: {model_id}")
28
+ model = SparseAutoModelForCausalLM.from_pretrained(
29
+ model_id,
30
+ device_map="cpu",
31
+ torch_dtype="auto"
32
+ )
33
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
34
+
35
+ print("\nConfiguring FP8 quantization recipe...")
36
+ recipe = QuantizationModifier(
37
+ targets="Linear",
38
+ scheme="FP8_DYNAMIC",
39
+ ignore=["lm_head"]
40
+ )
41
+
42
+ print("\nApplying quantization (this may take a while)...")
43
+ oneshot(model=model, recipe=recipe)
44
+
45
+ model_name = model_id.split("/")[-1]
46
+ save_dir = f"{model_name}-FP8-Dynamic"
47
+
48
+ print(f"\nSaving quantized model to: {save_dir}")
49
+ model.save_pretrained(save_dir, save_compressed=True)
50
+ tokenizer.save_pretrained(save_dir)
51
+
52
+ print("\nβœ… Quantization completed successfully!")
53
+ print(f"πŸ“ Quantized model saved to: {os.path.abspath(save_dir)}")
54
+ return save_dir
55
+
56
+ except Exception as e:
57
+ print(f"\n❌ Error during quantization: {str(e)}")
58
+ return None
59
+
60
+ if __name__ == "__main__":
61
+ print("""
62
+ ╔══════════════════════════════════════╗
63
+ β•‘ Model Quantization to FP8 β•‘
64
+ β•‘ (Dynamic Per-Token) β•‘
65
+ β•šβ•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•
66
+ """)
67
+
68
+ model_id = get_user_input()
69
+
70
+ print("\n=== Configuration Summary ===")
71
+ print(f"Model ID: {model_id}")
72
+ print("Quantization Type: FP8 Dynamic (per-token)")
73
+ print("Device: CPU")
74
+
75
+ while True:
76
+ confirm = input("\nProceed with quantization? (y/n): ").lower().strip()
77
+ if confirm in ['y', 'n']:
78
+ break
79
+ print("Please enter 'y' for yes or 'n' for no.")
80
+
81
+ if confirm == 'y':
82
+ quantized_model_path = quantize_model_fp8(model_id)
83
+ else:
84
+ print("\nQuantization cancelled.")