Merge pull request #485 from maximegmd/patch-4
Browse filesfix: finetune model inference needs the dtype fix to work with flash-attn
- src/axolotl/utils/models.py +12 -9
src/axolotl/utils/models.py
CHANGED
@@ -355,6 +355,7 @@ def load_model(
|
|
355 |
if hasattr(module, "weight"):
|
356 |
module.to(torch.float32)
|
357 |
|
|
|
358 |
if not cfg.gptq and (
|
359 |
(cfg.adapter == "lora" and load_in_8bit)
|
360 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
@@ -363,16 +364,18 @@ def load_model(
|
|
363 |
model = prepare_model_for_kbit_training(
|
364 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
365 |
)
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
|
|
|
|
|
|
|
|
|
|
372 |
module.to(cfg.torch_dtype)
|
373 |
-
if "lm_head" in name or "embed_tokens" in name:
|
374 |
-
if hasattr(module, "weight"):
|
375 |
-
module.to(cfg.torch_dtype)
|
376 |
|
377 |
model, lora_config = load_adapter(model, cfg, cfg.adapter)
|
378 |
|
|
|
355 |
if hasattr(module, "weight"):
|
356 |
module.to(torch.float32)
|
357 |
|
358 |
+
needs_fa2_dtype = cfg.adapter is not None
|
359 |
if not cfg.gptq and (
|
360 |
(cfg.adapter == "lora" and load_in_8bit)
|
361 |
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
|
|
|
364 |
model = prepare_model_for_kbit_training(
|
365 |
model, use_gradient_checkpointing=cfg.gradient_checkpointing
|
366 |
)
|
367 |
+
needs_fa2_dtype = True
|
368 |
+
|
369 |
+
# LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
|
370 |
+
# convert them back to fp16/bf16 for flash-attn compatibility.
|
371 |
+
if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
|
372 |
+
LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
|
373 |
+
for name, module in model.named_modules():
|
374 |
+
if "norm" in name:
|
375 |
+
module.to(cfg.torch_dtype)
|
376 |
+
if "lm_head" in name or "embed_tokens" in name:
|
377 |
+
if hasattr(module, "weight"):
|
378 |
module.to(cfg.torch_dtype)
|
|
|
|
|
|
|
379 |
|
380 |
model, lora_config = load_adapter(model, cfg, cfg.adapter)
|
381 |
|