tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

more steps for loss check

Browse Source
This commit is contained in:
Wing Lian
2026-04-23 18:43:18 +00:00
parent 431888c1de
commit 5db4272f69
3 changed files with 57 additions and 120 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
tests/e2e/multigpu/test_dist_muon_fsdp2.py
View File

@@ -1,23 +1,22 @@
"""Test module for DistMuon optimizer with FSDP2 multi-GPU functionality."""
import os
from pathlib import Path
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
from tests.e2e.utils import check_tensorboard_loss_decreased, 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."""
"""Verify that training completed successfully — artifacts, no-NaN, loss
stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
"""
output_path = Path(temp_dir)
model_files = list(output_path.glob("*.bin")) + list(
@@ -30,33 +29,13 @@ def verify_training_success(temp_dir):
"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/loss"]
if len(train_loss_df) == 0:
train_loss_df = df[df.tag == "train/train_loss"]
if len(train_loss_df) > 0:
values = train_loss_df.value.values
final_loss = values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (
f"Training loss is NaN: {final_loss}"
)
if len(values) >= 2:
initial_loss = float(values[0])
assert float(final_loss) <= initial_loss * 1.10, (
f"Training loss regressed: initial={initial_loss:.4f}, "
f"final={final_loss:.4f} — likely silent bug (e.g. "
"bad label masking) pushed loss scale up."
)
assert float(final_loss) <= 10.0, (
f"Final loss {final_loss:.4f} above sanity bound 10.0 "
"— absolute scale wrong."
)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=3,
final_window=3,
max_initial=5.0,
max_final=4.7,
)
class TestDistMuon:
@@ -77,7 +56,7 @@ class TestDistMuon:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
@@ -139,7 +118,7 @@ class TestDistMuon:
"lora_dropout": 0.05,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,

59
tests/e2e/multigpu/test_fsdp1.py
View File

@@ -1,24 +1,23 @@
"""Test module for FSDP1 multi-GPU functionality."""
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
from tests.e2e.utils import check_tensorboard_loss_decreased
AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
def verify_training_success(temp_dir):
"""Verify that training completed successfully by checking artifacts and loss."""
"""Verify that training completed successfully — artifacts, no-NaN, loss
stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
"""
output_path = Path(temp_dir)
model_files = list(output_path.glob("*.bin")) + list(
@@ -31,33 +30,13 @@ def verify_training_success(temp_dir):
"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/loss"]
if len(train_loss_df) == 0:
train_loss_df = df[df.tag == "train/train_loss"]
if len(train_loss_df) > 0:
values = train_loss_df.value.values
final_loss = values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (
f"Training loss is NaN: {final_loss}"
)
if len(values) >= 2:
initial_loss = float(values[0])
assert float(final_loss) <= initial_loss * 1.10, (
f"Training loss regressed: initial={initial_loss:.4f}, "
f"final={final_loss:.4f} — likely silent bug (e.g. "
"bad label masking) pushed loss scale up."
)
assert float(final_loss) <= 10.0, (
f"Final loss {final_loss:.4f} above sanity bound 10.0 "
"— absolute scale wrong."
)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=3,
final_window=3,
max_initial=5.0,
max_final=4.7,
)
class TestFSDP1:
@@ -81,11 +60,11 @@ class TestFSDP1:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -157,11 +136,11 @@ class TestFSDP1:
"lora_dropout": 0.05,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -217,11 +196,11 @@ class TestFSDP1:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -295,11 +274,11 @@ class TestFSDP1:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,

71
tests/e2e/multigpu/test_fsdp2.py
View File

@@ -1,24 +1,23 @@
"""Test module for FSDP2 multi-GPU functionality."""
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
from tests.e2e.utils import check_tensorboard_loss_decreased, 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."""
"""Verify that training completed successfully — artifacts, no-NaN, loss
stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
"""
output_path = Path(temp_dir)
model_files = list(output_path.glob("*.bin")) + list(
@@ -31,33 +30,13 @@ def verify_training_success(temp_dir):
"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/loss"]
if len(train_loss_df) == 0:
train_loss_df = df[df.tag == "train/train_loss"]
if len(train_loss_df) > 0:
values = train_loss_df.value.values
final_loss = values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (
f"Training loss is NaN: {final_loss}"
)
if len(values) >= 2:
initial_loss = float(values[0])
assert float(final_loss) <= initial_loss * 1.10, (
f"Training loss regressed: initial={initial_loss:.4f}, "
f"final={final_loss:.4f} — likely silent bug (e.g. "
"bad label masking) pushed loss scale up."
)
assert float(final_loss) <= 10.0, (
f"Final loss {final_loss:.4f} above sanity bound 10.0 "
"— absolute scale wrong."
)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=3,
final_window=3,
max_initial=5.0,
max_final=4.7,
)
class TestFSDP2:
@@ -82,11 +61,11 @@ class TestFSDP2:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -145,11 +124,11 @@ class TestFSDP2:
"lora_dropout": 0.05,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -209,11 +188,11 @@ class TestFSDP2:
"lora_alpha": 16,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -274,11 +253,11 @@ class TestFSDP2:
"lora_dropout": 0.05,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -335,11 +314,11 @@ class TestFSDP2:
"lora_alpha": 16,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -397,11 +376,11 @@ class TestFSDP2:
},
],
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,
@@ -459,11 +438,11 @@ class TestFSDP2:
"lora_dropout": 0.05,
"lora_target_linear": True,
"num_epochs": 1,
"max_steps": 2,
"max_steps": 20,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"flash_attention": True,