tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Merge pull request #485 from maximegmd/patch-4

Browse Source 
fix: finetune model inference needs the dtype fix to work with flash-attn
This commit is contained in:
Aman Gupta Karmani
2023-08-27 16:27:47 -04:00
committed by GitHub
parent 35130711d6 3a011ea1ef
commit f144e98a32
1 changed files with 11 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
src/axolotl/utils/models.py
View File

@@ -355,6 +355,7 @@ def load_model(
if hasattr(module, "weight"):
module.to(torch.float32)
needs_fa2_dtype = cfg.adapter is not None
if not cfg.gptq and (
(cfg.adapter == "lora" and load_in_8bit)
or (cfg.adapter == "qlora" and cfg.load_in_4bit)
@@ -363,16 +364,18 @@ def load_model(
model = prepare_model_for_kbit_training(
model, use_gradient_checkpointing=cfg.gradient_checkpointing
)
needs_fa2_dtype = True
# LlamaRMSNorm layers are in fp32 after kbit_training, so we need to
# convert them back to fp16/bf16 for flash-attn compatibility.
if cfg.flash_attention and cfg.is_llama_derived_model:
for name, module in model.named_modules():
if "norm" in name:
# LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
# convert them back to fp16/bf16 for flash-attn compatibility.
if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
for name, module in model.named_modules():
if "norm" in name:
module.to(cfg.torch_dtype)
if "lm_head" in name or "embed_tokens" in name:
if hasattr(module, "weight"):
module.to(cfg.torch_dtype)
if "lm_head" in name or "embed_tokens" in name:
if hasattr(module, "weight"):
module.to(cfg.torch_dtype)
model, lora_config = load_adapter(model, cfg, cfg.adapter)