support llama-adapter zero init attention
Browse files- scripts/finetune.py +4 -4
- src/axolotl/utils/models.py +50 -21
scripts/finetune.py
CHANGED
|
@@ -146,8 +146,8 @@ def train(
|
|
| 146 |
cfg.bf16 = False
|
| 147 |
|
| 148 |
# Load the model and tokenizer
|
| 149 |
-
logging.info("loading model, tokenizer, and
|
| 150 |
-
model, tokenizer,
|
| 151 |
cfg.base_model,
|
| 152 |
cfg.base_model_config,
|
| 153 |
cfg.model_type,
|
|
@@ -186,9 +186,9 @@ def train(
|
|
| 186 |
model = torch.compile(model)
|
| 187 |
|
| 188 |
# go ahead and presave, so we have the adapter config available to inspect
|
| 189 |
-
if
|
| 190 |
logging.info(f"Pre-saving adapter config to {cfg.output_dir}")
|
| 191 |
-
|
| 192 |
|
| 193 |
# In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
|
| 194 |
if cfg.local_rank == 0:
|
|
|
|
| 146 |
cfg.bf16 = False
|
| 147 |
|
| 148 |
# Load the model and tokenizer
|
| 149 |
+
logging.info("loading model, tokenizer, and peft_config...")
|
| 150 |
+
model, tokenizer, peft_config = load_model(
|
| 151 |
cfg.base_model,
|
| 152 |
cfg.base_model_config,
|
| 153 |
cfg.model_type,
|
|
|
|
| 186 |
model = torch.compile(model)
|
| 187 |
|
| 188 |
# go ahead and presave, so we have the adapter config available to inspect
|
| 189 |
+
if peft_config:
|
| 190 |
logging.info(f"Pre-saving adapter config to {cfg.output_dir}")
|
| 191 |
+
peft_config.save_pretrained(cfg.output_dir)
|
| 192 |
|
| 193 |
# In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
|
| 194 |
if cfg.local_rank == 0:
|
src/axolotl/utils/models.py
CHANGED
|
@@ -195,11 +195,41 @@ def load_adapter(model, cfg, adapter):
|
|
| 195 |
return model, None
|
| 196 |
if adapter == "lora":
|
| 197 |
return load_lora(model, cfg)
|
| 198 |
-
|
|
|
|
| 199 |
|
| 200 |
raise NotImplementedError(f"{adapter} peft adapter not available")
|
| 201 |
|
| 202 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 203 |
def load_lora(model, cfg):
|
| 204 |
# type: (PreTrainedModel, AttrDefault) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
|
| 205 |
|
|
@@ -211,27 +241,26 @@ def load_lora(model, cfg):
|
|
| 211 |
|
| 212 |
lora_config = None
|
| 213 |
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
)
|
| 224 |
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
|
| 235 |
-
|
| 236 |
|
| 237 |
return model, lora_config
|
|
|
|
| 195 |
return model, None
|
| 196 |
if adapter == "lora":
|
| 197 |
return load_lora(model, cfg)
|
| 198 |
+
if adapter == "llama-adapter":
|
| 199 |
+
return load_llama_adapter(model, cfg)
|
| 200 |
|
| 201 |
raise NotImplementedError(f"{adapter} peft adapter not available")
|
| 202 |
|
| 203 |
|
| 204 |
+
def load_llama_adapter(model, cfg):
|
| 205 |
+
# type: (PreTrainedModel, AttrDefault) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
|
| 206 |
+
from peft import (
|
| 207 |
+
AdaptionPromptConfig,
|
| 208 |
+
get_peft_model,
|
| 209 |
+
PeftModel,
|
| 210 |
+
)
|
| 211 |
+
|
| 212 |
+
peft_config = AdaptionPromptConfig(
|
| 213 |
+
adapter_layers=cfg.peft_adapter.layers, # layers (L)
|
| 214 |
+
adapter_len=cfg.peft_adapter.len, # prompt length (K)
|
| 215 |
+
task_type="CAUSAL_LM",
|
| 216 |
+
)
|
| 217 |
+
|
| 218 |
+
if cfg.peft_model_dir:
|
| 219 |
+
model = PeftModel.from_pretrained(
|
| 220 |
+
model,
|
| 221 |
+
cfg.lora_model_dir,
|
| 222 |
+
device_map=cfg.device_map,
|
| 223 |
+
torch_dtype=torch.float16,
|
| 224 |
+
)
|
| 225 |
+
else:
|
| 226 |
+
model = get_peft_model(model, peft_config)
|
| 227 |
+
|
| 228 |
+
model.print_trainable_parameters()
|
| 229 |
+
|
| 230 |
+
return model, peft_config
|
| 231 |
+
|
| 232 |
+
|
| 233 |
def load_lora(model, cfg):
|
| 234 |
# type: (PreTrainedModel, AttrDefault) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
|
| 235 |
|
|
|
|
| 241 |
|
| 242 |
lora_config = None
|
| 243 |
|
| 244 |
+
lora_config = LoraConfig(
|
| 245 |
+
r=cfg.lora_r,
|
| 246 |
+
lora_alpha=cfg.lora_alpha,
|
| 247 |
+
target_modules=cfg.lora_target_modules,
|
| 248 |
+
lora_dropout=cfg.lora_dropout,
|
| 249 |
+
fan_in_fan_out=cfg.lora_fan_in_fan_out,
|
| 250 |
+
bias="none",
|
| 251 |
+
task_type="CAUSAL_LM",
|
| 252 |
+
)
|
|
|
|
| 253 |
|
| 254 |
+
if cfg.lora_model_dir:
|
| 255 |
+
model = PeftModel.from_pretrained(
|
| 256 |
+
model,
|
| 257 |
+
cfg.lora_model_dir,
|
| 258 |
+
device_map=cfg.device_map,
|
| 259 |
+
torch_dtype=torch.float16,
|
| 260 |
+
)
|
| 261 |
+
else:
|
| 262 |
+
model = get_peft_model(model, lora_config)
|
| 263 |
|
| 264 |
+
model.print_trainable_parameters()
|
| 265 |
|
| 266 |
return model, lora_config
|