Spaces:
Running
Running
adds quantization configuration correctly
Browse files
config/train_gpt_oss_memory_optimized.py
CHANGED
@@ -89,11 +89,8 @@ class GPTOSSMemoryOptimizedConfig:
|
|
89 |
|
90 |
if self.quantization_config is None:
|
91 |
self.quantization_config = {
|
92 |
-
"dequantize": True,
|
93 |
-
"load_in_4bit":
|
94 |
-
"bnb_4bit_compute_dtype": "bfloat16",
|
95 |
-
"bnb_4bit_use_double_quant": True,
|
96 |
-
"bnb_4bit_quant_type": "nf4"
|
97 |
}
|
98 |
|
99 |
if self.model_kwargs is None:
|
|
|
89 |
|
90 |
if self.quantization_config is None:
|
91 |
self.quantization_config = {
|
92 |
+
"dequantize": True, # Use Mxfp4Config as per tutorial
|
93 |
+
"load_in_4bit": False # Only use 4-bit if explicitly needed
|
|
|
|
|
|
|
94 |
}
|
95 |
|
96 |
if self.model_kwargs is None:
|
requirements/requirements_core.txt
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
# Core dependencies for SmolLM3 and GPT-OSS fine-tuning
|
2 |
torch>=2.0.0
|
3 |
-
transformers
|
4 |
datasets>=2.14.0
|
5 |
accelerate>=0.20.0
|
6 |
peft>=0.17.0 # Updated for GPT-OSS LoRA support
|
|
|
1 |
# Core dependencies for SmolLM3 and GPT-OSS fine-tuning
|
2 |
torch>=2.0.0
|
3 |
+
transformers @ git+https://github.com/huggingface/transformers.git # Latest version with GPT-OSS support
|
4 |
datasets>=2.14.0
|
5 |
accelerate>=0.20.0
|
6 |
peft>=0.17.0 # Updated for GPT-OSS LoRA support
|
scripts/training/train_gpt_oss.py
CHANGED
@@ -27,26 +27,38 @@ def load_gpt_oss_model_and_tokenizer(config):
|
|
27 |
|
28 |
# Set up quantization config based on config
|
29 |
if config.quantization_config and config.quantization_config.get("load_in_4bit"):
|
30 |
-
# Use BitsAndBytesConfig for 4-bit quantization
|
31 |
quantization_config = BitsAndBytesConfig(
|
32 |
load_in_4bit=True,
|
33 |
bnb_4bit_compute_dtype=torch.bfloat16,
|
34 |
bnb_4bit_use_double_quant=True,
|
35 |
bnb_4bit_quant_type="nf4"
|
36 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
else:
|
38 |
-
#
|
39 |
quantization_config = None
|
40 |
|
41 |
# Model kwargs as per tutorial
|
42 |
model_kwargs = {
|
43 |
"attn_implementation": "eager",
|
44 |
"torch_dtype": torch.bfloat16,
|
45 |
-
"quantization_config": quantization_config,
|
46 |
"use_cache": False,
|
47 |
"device_map": "auto",
|
48 |
}
|
49 |
|
|
|
|
|
|
|
|
|
50 |
model = AutoModelForCausalLM.from_pretrained(config.model_name, **model_kwargs)
|
51 |
|
52 |
return model, tokenizer
|
|
|
27 |
|
28 |
# Set up quantization config based on config
|
29 |
if config.quantization_config and config.quantization_config.get("load_in_4bit"):
|
30 |
+
# Use BitsAndBytesConfig for 4-bit quantization (memory optimized)
|
31 |
quantization_config = BitsAndBytesConfig(
|
32 |
load_in_4bit=True,
|
33 |
bnb_4bit_compute_dtype=torch.bfloat16,
|
34 |
bnb_4bit_use_double_quant=True,
|
35 |
bnb_4bit_quant_type="nf4"
|
36 |
)
|
37 |
+
elif config.quantization_config and config.quantization_config.get("dequantize"):
|
38 |
+
# Try to use Mxfp4Config if available (as per tutorial)
|
39 |
+
try:
|
40 |
+
from transformers import Mxfp4Config
|
41 |
+
quantization_config = Mxfp4Config(dequantize=True)
|
42 |
+
except ImportError:
|
43 |
+
# Fallback to no quantization if Mxfp4Config not available
|
44 |
+
print("Warning: Mxfp4Config not available, using no quantization")
|
45 |
+
quantization_config = None
|
46 |
else:
|
47 |
+
# No quantization
|
48 |
quantization_config = None
|
49 |
|
50 |
# Model kwargs as per tutorial
|
51 |
model_kwargs = {
|
52 |
"attn_implementation": "eager",
|
53 |
"torch_dtype": torch.bfloat16,
|
|
|
54 |
"use_cache": False,
|
55 |
"device_map": "auto",
|
56 |
}
|
57 |
|
58 |
+
# Only add quantization_config if it's not None
|
59 |
+
if quantization_config is not None:
|
60 |
+
model_kwargs["quantization_config"] = quantization_config
|
61 |
+
|
62 |
model = AutoModelForCausalLM.from_pretrained(config.model_name, **model_kwargs)
|
63 |
|
64 |
return model, tokenizer
|