FSDP1 -> FSDP2 (#2760)
* FSDP2 args migration implementation This commit implements the migration to FSDP2 arguments including: - FSDP2 support with LoRA training - DPO integration with FSDP2 - Model loading fixes and refactoring - CPU offloading and PEFT handling - Test updates and CI improvements - Bug fixes for dtype errors and various edge cases
This commit is contained in:
salman
GitHub
326
tests/e2e/multigpu/test_fsdp1.py
Normal file
326
tests/e2e/multigpu/test_fsdp1.py
Normal file
@@ -0,0 +1,326 @@
|
||||
"""Test module for FSDP1 multi-GPU functionality."""
|
||||
|
||||
# pylint: disable=duplicate-code
|
||||
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
import yaml
|
||||
from accelerate.test_utils import execute_subprocess_async
|
||||
from tbparse import SummaryReader
|
||||
from transformers.testing_utils import get_torch_dist_unique_port
|
||||
|
||||
from axolotl.utils.dict import DictDefault
|
||||
|
||||
from tests.e2e.utils import most_recent_subdir
|
||||
|
||||
AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
|
||||
|
||||
|
||||
def verify_training_success(temp_dir):
|
||||
"""Verify that training completed successfully by checking artifacts and loss."""
|
||||
output_path = Path(temp_dir)
|
||||
|
||||
model_files = list(output_path.glob("*.bin")) + list(
|
||||
output_path.glob("*.safetensors")
|
||||
)
|
||||
assert len(model_files) > 0, "No model files found - training may have failed"
|
||||
|
||||
checkpoint_files = list(output_path.glob("checkpoint-*"))
|
||||
assert (
|
||||
len(checkpoint_files) > 0
|
||||
), "No checkpoint files found - training may have failed"
|
||||
|
||||
tb_log_path = most_recent_subdir(temp_dir + "/runs")
|
||||
if tb_log_path:
|
||||
event_files = sorted(os.listdir(tb_log_path))
|
||||
if event_files:
|
||||
event_file = os.path.join(tb_log_path, event_files[0])
|
||||
reader = SummaryReader(event_file)
|
||||
df = reader.scalars
|
||||
train_loss_df = df[df.tag == "train/train_loss"]
|
||||
if len(train_loss_df) > 0:
|
||||
final_loss = train_loss_df.value.values[-1]
|
||||
assert not torch.isnan(
|
||||
torch.tensor(final_loss)
|
||||
), f"Training loss is NaN: {final_loss}"
|
||||
|
||||
|
||||
class TestFSDP1:
|
||||
"""Test class for FSDP1 functionality."""
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"fsdp_cpu_ram_efficient_loading",
|
||||
[True, False],
|
||||
)
|
||||
def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
|
||||
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%]",
|
||||
},
|
||||
],
|
||||
"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": "1",
|
||||
"fsdp_config": {
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": fsdp_cpu_ram_efficient_loading,
|
||||
"fsdp_transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_sharding_strategy": "FULL_SHARD",
|
||||
"fsdp_sync_module_states": True,
|
||||
"fsdp_use_orig_params": False,
|
||||
},
|
||||
"use_tensorboard": True,
|
||||
"bf16": 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)
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"adapter_config",
|
||||
[
|
||||
{
|
||||
"adapter": "lora",
|
||||
"load_in_4bit": False,
|
||||
},
|
||||
{
|
||||
"adapter": "qlora",
|
||||
"load_in_4bit": True,
|
||||
},
|
||||
],
|
||||
)
|
||||
def test_lora_sft(self, temp_dir, adapter_config):
|
||||
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": adapter_config["adapter"],
|
||||
"load_in_4bit": adapter_config["load_in_4bit"],
|
||||
"lora_r": 8,
|
||||
"lora_alpha": 16,
|
||||
"lora_dropout": 0.05,
|
||||
"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": "1",
|
||||
"fsdp_config": {
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": True,
|
||||
"fsdp_transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_sharding_strategy": "FULL_SHARD",
|
||||
"fsdp_sync_module_states": True,
|
||||
"fsdp_use_orig_params": False,
|
||||
},
|
||||
"use_tensorboard": True,
|
||||
"bf16": 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)
|
||||
|
||||
def test_dpo_fft(self, temp_dir):
|
||||
cfg = DictDefault(
|
||||
{
|
||||
"base_model": "Qwen/Qwen2.5-0.5B",
|
||||
"sequence_len": 2048,
|
||||
"val_set_size": 0.01,
|
||||
"rl": "dpo",
|
||||
"chat_template": "chatml",
|
||||
"datasets": [
|
||||
{
|
||||
"path": "Intel/orca_dpo_pairs",
|
||||
"split": "train",
|
||||
"type": "chatml.intel",
|
||||
},
|
||||
],
|
||||
"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": "1",
|
||||
"fsdp_config": {
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": True,
|
||||
"fsdp_transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_sharding_strategy": "FULL_SHARD",
|
||||
"fsdp_sync_module_states": True,
|
||||
"fsdp_use_orig_params": False,
|
||||
},
|
||||
"use_tensorboard": 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)
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
"adapter_config",
|
||||
[
|
||||
{
|
||||
"adapter": "lora",
|
||||
"load_in_4bit": False,
|
||||
},
|
||||
{
|
||||
"adapter": "qlora",
|
||||
"load_in_4bit": True,
|
||||
},
|
||||
],
|
||||
)
|
||||
def test_dpo_lora(self, temp_dir, adapter_config):
|
||||
cfg = DictDefault(
|
||||
{
|
||||
"base_model": "Qwen/Qwen2.5-0.5B",
|
||||
"load_in_4bit": adapter_config["load_in_4bit"],
|
||||
"rl": "dpo",
|
||||
"chat_template": "chatml",
|
||||
"sequence_len": 2048,
|
||||
"adapter": adapter_config["adapter"],
|
||||
"lora_r": 8,
|
||||
"lora_alpha": 16,
|
||||
"lora_dropout": 0.05,
|
||||
"lora_target_linear": True,
|
||||
"val_set_size": 0.01,
|
||||
"datasets": [
|
||||
{
|
||||
"path": "Intel/orca_dpo_pairs",
|
||||
"split": "train",
|
||||
"type": "chatml.intel",
|
||||
},
|
||||
],
|
||||
"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": "1",
|
||||
"fsdp_config": {
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": True,
|
||||
"fsdp_transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_sharding_strategy": "FULL_SHARD",
|
||||
"fsdp_sync_module_states": True,
|
||||
"fsdp_use_orig_params": False,
|
||||
},
|
||||
"use_tensorboard": True,
|
||||
"bf16": "auto",
|
||||
"tf32": 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)
|
||||
355
tests/e2e/multigpu/test_fsdp2.py
Normal file
355
tests/e2e/multigpu/test_fsdp2.py
Normal file
@@ -0,0 +1,355 @@
|
||||
"""Test module for FSDP2 multi-GPU functionality."""
|
||||
|
||||
# pylint: disable=duplicate-code
|
||||
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
import yaml
|
||||
from accelerate.test_utils import execute_subprocess_async
|
||||
from tbparse import SummaryReader
|
||||
from transformers.testing_utils import get_torch_dist_unique_port
|
||||
|
||||
from axolotl.utils.dict import DictDefault
|
||||
|
||||
from tests.e2e.utils import most_recent_subdir, require_torch_2_7_0
|
||||
|
||||
AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
|
||||
|
||||
|
||||
def verify_training_success(temp_dir):
|
||||
"""Verify that training completed successfully by checking artifacts and loss."""
|
||||
output_path = Path(temp_dir)
|
||||
|
||||
model_files = list(output_path.glob("*.bin")) + list(
|
||||
output_path.glob("*.safetensors")
|
||||
)
|
||||
assert len(model_files) > 0, "No model files found - training may have failed"
|
||||
|
||||
checkpoint_files = list(output_path.glob("checkpoint-*"))
|
||||
assert (
|
||||
len(checkpoint_files) > 0
|
||||
), "No checkpoint files found - training may have failed"
|
||||
|
||||
tb_log_path = most_recent_subdir(temp_dir + "/runs")
|
||||
if tb_log_path:
|
||||
event_files = sorted(os.listdir(tb_log_path))
|
||||
if event_files:
|
||||
event_file = os.path.join(tb_log_path, event_files[0])
|
||||
reader = SummaryReader(event_file)
|
||||
df = reader.scalars
|
||||
train_loss_df = df[df.tag == "train/train_loss"]
|
||||
if len(train_loss_df) > 0:
|
||||
final_loss = train_loss_df.value.values[-1]
|
||||
assert not torch.isnan(
|
||||
torch.tensor(final_loss)
|
||||
), f"Training loss is NaN: {final_loss}"
|
||||
|
||||
|
||||
class TestFSDP2:
|
||||
"""Test class for FSDP2 functionality."""
|
||||
|
||||
@require_torch_2_7_0
|
||||
@pytest.mark.parametrize(
|
||||
"fsdp_cpu_ram_efficient_loading",
|
||||
[True, False],
|
||||
)
|
||||
def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
|
||||
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%]",
|
||||
},
|
||||
],
|
||||
"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": fsdp_cpu_ram_efficient_loading,
|
||||
"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,
|
||||
}
|
||||
)
|
||||
|
||||
# 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
|
||||
@pytest.mark.parametrize("peft_use_dora", [True, False])
|
||||
def test_lora_sft(self, temp_dir, peft_use_dora):
|
||||
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%]",
|
||||
},
|
||||
],
|
||||
"peft_use_dora": peft_use_dora,
|
||||
"adapter": "lora",
|
||||
"lora_r": 8,
|
||||
"lora_alpha": 16,
|
||||
"lora_dropout": 0.05,
|
||||
"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,
|
||||
}
|
||||
)
|
||||
|
||||
# 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(
|
||||
{
|
||||
"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_dropout": 0.05,
|
||||
"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,
|
||||
}
|
||||
)
|
||||
|
||||
# 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(
|
||||
{
|
||||
"base_model": "Qwen/Qwen2.5-0.5B",
|
||||
"sequence_len": 2048,
|
||||
"val_set_size": 0.01,
|
||||
"rl": "dpo",
|
||||
"chat_template": "chatml",
|
||||
"datasets": [
|
||||
{
|
||||
"path": "Intel/orca_dpo_pairs",
|
||||
"split": "train",
|
||||
"type": "chatml.intel",
|
||||
},
|
||||
],
|
||||
"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,
|
||||
}
|
||||
)
|
||||
|
||||
# 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_lora(self, temp_dir):
|
||||
cfg = DictDefault(
|
||||
{
|
||||
"base_model": "Qwen/Qwen2.5-0.5B",
|
||||
"sequence_len": 2048,
|
||||
"rl": "dpo",
|
||||
"chat_template": "chatml",
|
||||
"datasets": [
|
||||
{
|
||||
"path": "Intel/orca_dpo_pairs",
|
||||
"split": "train",
|
||||
"type": "chatml.intel",
|
||||
},
|
||||
],
|
||||
"adapter": "lora",
|
||||
"lora_r": 8,
|
||||
"lora_alpha": 16,
|
||||
"lora_dropout": 0.05,
|
||||
"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,
|
||||
}
|
||||
)
|
||||
|
||||
# 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)
|
||||
@@ -1,93 +0,0 @@
|
||||
"""
|
||||
E2E tests for multigpu qwen2
|
||||
"""
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
import pytest
|
||||
import yaml
|
||||
from accelerate.test_utils import execute_subprocess_async
|
||||
from transformers.testing_utils import get_torch_dist_unique_port
|
||||
|
||||
from axolotl.utils.dict import DictDefault
|
||||
|
||||
|
||||
class TestMultiGPUQwen2:
|
||||
"""
|
||||
Test case for Llama models using LoRA
|
||||
"""
|
||||
|
||||
@pytest.mark.parametrize("base_model", ["Qwen/Qwen2-0.5B", "Qwen/Qwen2.5-0.5B"])
|
||||
def test_qlora_fsdp_dpo(self, base_model, temp_dir):
|
||||
# pylint: disable=duplicate-code
|
||||
cfg = DictDefault(
|
||||
{
|
||||
"base_model": base_model,
|
||||
"load_in_4bit": True,
|
||||
"rl": "dpo",
|
||||
"chat_template": "chatml",
|
||||
"sequence_len": 2048,
|
||||
"adapter": "qlora",
|
||||
"lora_r": 8,
|
||||
"lora_alpha": 16,
|
||||
"lora_dropout": 0.05,
|
||||
"lora_target_linear": True,
|
||||
"val_set_size": 0.01,
|
||||
"datasets": [
|
||||
{
|
||||
"path": "Intel/orca_dpo_pairs",
|
||||
"split": "train",
|
||||
"type": "chatml.intel",
|
||||
},
|
||||
],
|
||||
"num_epochs": 1,
|
||||
"max_steps": 2,
|
||||
"warmup_steps": 20,
|
||||
"micro_batch_size": 2,
|
||||
"gradient_accumulation_steps": 2,
|
||||
"output_dir": temp_dir,
|
||||
"dataset_prepared_path": temp_dir + "/last_run_prepared",
|
||||
"learning_rate": 0.00001,
|
||||
"optimizer": "adamw_torch_fused",
|
||||
"lr_scheduler": "cosine",
|
||||
"flash_attention": True,
|
||||
"bf16": "auto",
|
||||
"tf32": True,
|
||||
# "gradient_checkpointing": True,
|
||||
"gradient_checkpointing_kwargs": {
|
||||
"use_reentrant": False,
|
||||
},
|
||||
"fsdp": [
|
||||
"full_shard",
|
||||
"auto_wrap",
|
||||
],
|
||||
"fsdp_config": {
|
||||
"fsdp_limit_all_gathers": True,
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_sync_module_states": True,
|
||||
"fsdp_use_orig_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": False,
|
||||
"fsdp_transformer_layer_cls_to_wrap": "Qwen2DecoderLayer",
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_sharding_strategy": "FULL_SHARD",
|
||||
},
|
||||
}
|
||||
)
|
||||
|
||||
# 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()}",
|
||||
]
|
||||
)
|
||||
@@ -77,6 +77,18 @@ def require_torch_2_6_0(test_case):
|
||||
return unittest.skipUnless(is_min_2_6_0(), "test requires torch>=2.6.0")(test_case)
|
||||
|
||||
|
||||
def require_torch_2_7_0(test_case):
|
||||
"""
|
||||
Decorator marking a test that requires torch >= 2.7.0
|
||||
"""
|
||||
|
||||
def is_min_2_7_0():
|
||||
torch_version = version.parse(torch.__version__)
|
||||
return torch_version >= version.parse("2.7.0")
|
||||
|
||||
return unittest.skipUnless(is_min_2_7_0(), "test requires torch>=2.7.0")(test_case)
|
||||
|
||||
|
||||
def require_torch_lt_2_6_0(test_case):
|
||||
"""
|
||||
Decorator marking a test that requires torch < 2.6.0
|
||||
|
||||
@@ -5,7 +5,11 @@ Test classes for checking functionality of the cfg normalization
|
||||
import unittest
|
||||
from unittest.mock import patch
|
||||
|
||||
from axolotl.utils.config import normalize_cfg_datasets, normalize_config
|
||||
from axolotl.utils.config import (
|
||||
migrate_fsdp_config,
|
||||
normalize_cfg_datasets,
|
||||
normalize_config,
|
||||
)
|
||||
from axolotl.utils.dict import DictDefault
|
||||
|
||||
|
||||
@@ -90,3 +94,104 @@ class NormalizeConfigTestCase(unittest.TestCase):
|
||||
|
||||
self.assertTrue(cfg.bf16)
|
||||
self.assertFalse(cfg.fp16)
|
||||
|
||||
def test_migrate_fsdp_config(self):
|
||||
"""Test basic FSDP config migration with and without fsdp_version"""
|
||||
cfg_with_version = DictDefault(
|
||||
{
|
||||
"fsdp_config": {
|
||||
"fsdp_version": 2,
|
||||
"fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
|
||||
"fsdp_offload_params": False,
|
||||
"fsdp_cpu_ram_efficient_loading": True,
|
||||
"regular_param": "value",
|
||||
}
|
||||
}
|
||||
)
|
||||
|
||||
migrate_fsdp_config(cfg_with_version)
|
||||
|
||||
self.assertEqual(cfg_with_version.fsdp_version, 2)
|
||||
self.assertEqual(
|
||||
cfg_with_version.fsdp_config.auto_wrap_policy, "TRANSFORMER_BASED_WRAP"
|
||||
)
|
||||
self.assertEqual(cfg_with_version.fsdp_config.offload_params, False)
|
||||
self.assertEqual(cfg_with_version.fsdp_config.cpu_ram_efficient_loading, True)
|
||||
self.assertEqual(cfg_with_version.fsdp_config.regular_param, "value")
|
||||
|
||||
self.assertNotIn("fsdp_auto_wrap_policy", cfg_with_version.fsdp_config)
|
||||
self.assertNotIn("fsdp_offload_params", cfg_with_version.fsdp_config)
|
||||
self.assertNotIn("fsdp_cpu_ram_efficient_loading", cfg_with_version.fsdp_config)
|
||||
self.assertNotIn("fsdp_version", cfg_with_version.fsdp_config)
|
||||
self.assertNotIn("version", cfg_with_version.fsdp_config)
|
||||
|
||||
cfg_without_version = DictDefault(
|
||||
{
|
||||
"fsdp_config": {
|
||||
"fsdp_auto_wrap_policy": "SIZE_BASED_WRAP",
|
||||
"fsdp_offload_params": True,
|
||||
"regular_param": "value",
|
||||
}
|
||||
}
|
||||
)
|
||||
|
||||
migrate_fsdp_config(cfg_without_version)
|
||||
|
||||
self.assertNotIn("fsdp_version", cfg_without_version)
|
||||
self.assertEqual(
|
||||
cfg_without_version.fsdp_config.auto_wrap_policy, "SIZE_BASED_WRAP"
|
||||
)
|
||||
self.assertEqual(cfg_without_version.fsdp_config.offload_params, True)
|
||||
self.assertEqual(cfg_without_version.fsdp_config.regular_param, "value")
|
||||
|
||||
self.assertNotIn("fsdp_auto_wrap_policy", cfg_without_version.fsdp_config)
|
||||
self.assertNotIn("fsdp_offload_params", cfg_without_version.fsdp_config)
|
||||
|
||||
def test_migrate_fsdp_config_no_fsdp_config(self):
|
||||
"""Test that function doesn't crash when no fsdp_config is present"""
|
||||
cfg = DictDefault({"some_other_config": "value"})
|
||||
|
||||
migrate_fsdp_config(cfg)
|
||||
|
||||
self.assertNotIn("fsdp_config", cfg)
|
||||
self.assertNotIn("fsdp_version", cfg)
|
||||
self.assertEqual(cfg.some_other_config, "value")
|
||||
|
||||
def test_migrate_fsdp_config_empty_fsdp_config(self):
|
||||
"""Test migration with empty fsdp_config"""
|
||||
cfg = DictDefault({"fsdp_config": {}})
|
||||
|
||||
migrate_fsdp_config(cfg)
|
||||
|
||||
self.assertNotIn("fsdp_version", cfg)
|
||||
self.assertEqual(cfg.fsdp_config, {})
|
||||
|
||||
def test_migrate_fsdp_config_mixed_keys(self):
|
||||
"""Test migration with a mix of fsdp_ and non-fsdp_ keys"""
|
||||
cfg = DictDefault(
|
||||
{
|
||||
"fsdp_config": {
|
||||
"fsdp_version": 1,
|
||||
"fsdp_state_dict_type": "FULL_STATE_DICT",
|
||||
"mixed_precision_policy": "fp16",
|
||||
"activation_checkpointing": True,
|
||||
"fsdp_reshard_after_forward": False,
|
||||
}
|
||||
}
|
||||
)
|
||||
|
||||
migrate_fsdp_config(cfg)
|
||||
|
||||
self.assertEqual(cfg.fsdp_version, 1)
|
||||
self.assertEqual(cfg.fsdp_config.state_dict_type, "FULL_STATE_DICT")
|
||||
self.assertEqual(cfg.fsdp_config.reshard_after_forward, False)
|
||||
self.assertEqual(cfg.fsdp_config.mixed_precision_policy, "fp16")
|
||||
self.assertEqual(cfg.fsdp_config.activation_checkpointing, True)
|
||||
|
||||
# Check original fsdp_ keys are removed
|
||||
self.assertNotIn("fsdp_version", cfg.fsdp_config)
|
||||
self.assertNotIn("fsdp_state_dict_type", cfg.fsdp_config)
|
||||
self.assertNotIn("fsdp_reshard_after_forward", cfg.fsdp_config)
|
||||
|
||||
# Ensure no duplicate version key
|
||||
self.assertNotIn("version", cfg.fsdp_config)
|
||||
|
||||
Reference in New Issue
Block a user