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

FSDP2 + LoRA kernels (#2992)

Browse Source 
* impl fix

* smoke tests

* patches for fsdp2 + qlora compat

* nit

* working fix

* working fix

* fix merge

* minifying patches; update bnb dep

* renaming; adding tests

* remove duplicate test, add dora guard

* generalize __torch_function__

* revert generalization

* update comments
This commit is contained in:
Dan Saunders
2025-08-03 20:05:17 -04:00
committed by GitHub
parent deac7b18a1
commit e758343cac
7 changed files with 520 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

127
tests/e2e/multigpu/test_fsdp2.py
View File

@@ -174,6 +174,69 @@ class TestFSDP2:
verify_training_success(temp_dir)
@require_torch_2_7_0
def test_lora_sft_kernels(self, temp_dir):
cfg = DictDefault(
{
"base_model": "Qwen/Qwen2.5-0.5B",
"sequence_len": 2048,
"val_set_size": 0.01,
"datasets": [
{
"path": "tatsu-lab/alpaca",
"type": "alpaca",
"split": "train[:10%]",
},
],
"adapter": "lora",
"lora_r": 8,
"lora_alpha": 16,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
"fsdp_version": 2,
"fsdp_config": {
"offload_params": False,
"cpu_ram_efficient_loading": False,
"transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
"state_dict_type": "FULL_STATE_DICT",
"auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
"reshard_after_forward": True,
},
"use_tensorboard": True,
"bf16": True,
"lora_mlp_kernel": True,
"lora_qkv_kernel": True,
"lora_o_kernel": True,
}
)
# write cfg to yaml file
Path(temp_dir).mkdir(parents=True, exist_ok=True)
with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
execute_subprocess_async(
[
"axolotl",
"train",
str(Path(temp_dir) / "config.yaml"),
"--num-processes",
"2",
"--main-process-port",
f"{get_torch_dist_unique_port()}",
]
)
verify_training_success(temp_dir)
@require_torch_2_7_0
def test_qlora_sft(self, temp_dir):
cfg = DictDefault(
@@ -236,6 +299,70 @@ class TestFSDP2:
verify_training_success(temp_dir)
@require_torch_2_7_0
def test_qlora_sft_kernels(self, temp_dir):
cfg = DictDefault(
{
"base_model": "Qwen/Qwen2.5-0.5B",
"sequence_len": 2048,
"val_set_size": 0.01,
"datasets": [
{
"path": "tatsu-lab/alpaca",
"type": "alpaca",
"split": "train[:10%]",
},
],
"load_in_4bit": True,
"adapter": "qlora",
"lora_r": 8,
"lora_alpha": 16,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
"fsdp_version": 2,
"fsdp_config": {
"offload_params": False,
"cpu_ram_efficient_loading": False,
"transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
"state_dict_type": "FULL_STATE_DICT",
"auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
"reshard_after_forward": True,
},
"use_tensorboard": True,
"bf16": True,
"lora_mlp_kernel": True,
"lora_qkv_kernel": True,
"lora_o_kernel": True,
}
)
# write cfg to yaml file
Path(temp_dir).mkdir(parents=True, exist_ok=True)
with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
execute_subprocess_async(
[
"axolotl",
"train",
str(Path(temp_dir) / "config.yaml"),
"--num-processes",
"2",
"--main-process-port",
f"{get_torch_dist_unique_port()}",
]
)
verify_training_success(temp_dir)
@require_torch_2_7_0
def test_dpo_fft(self, temp_dir):
cfg = DictDefault(