Upload scripts/train_orpo_n8n_thinking.py with huggingface_hub
Browse files
scripts/train_orpo_n8n_thinking.py
CHANGED
|
@@ -9,7 +9,6 @@
|
|
| 9 |
# "bitsandbytes",
|
| 10 |
# "sentencepiece",
|
| 11 |
# "protobuf",
|
| 12 |
-
# "flash-attn",
|
| 13 |
# ]
|
| 14 |
# ///
|
| 15 |
|
|
@@ -71,7 +70,7 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
| 71 |
MODEL_NAME,
|
| 72 |
quantization_config=bnb_config,
|
| 73 |
device_map="auto",
|
| 74 |
-
attn_implementation="
|
| 75 |
)
|
| 76 |
|
| 77 |
# LoRA configuration for efficient training on 7B model
|
|
|
|
| 9 |
# "bitsandbytes",
|
| 10 |
# "sentencepiece",
|
| 11 |
# "protobuf",
|
|
|
|
| 12 |
# ]
|
| 13 |
# ///
|
| 14 |
|
|
|
|
| 70 |
MODEL_NAME,
|
| 71 |
quantization_config=bnb_config,
|
| 72 |
device_map="auto",
|
| 73 |
+
attn_implementation="sdpa", # Use scaled dot-product attention
|
| 74 |
)
|
| 75 |
|
| 76 |
# LoRA configuration for efficient training on 7B model
|