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

fix flaky tests; should be using train loss from final step rather than final avg train loss

Browse Source
This commit is contained in:
Wing Lian
2026-03-22 10:38:46 -04:00
parent 5b2e3f00ce
commit 6130e40c37
21 changed files with 37 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/e2e/multigpu/patched/test_sp.py
View File

@@ -94,7 +94,7 @@ class TestSequenceParallelism:
check_tensorboard(
temp_dir + "/runs",
"train/train_loss",
"train/loss",
threshold,
"Train Loss (%s) is too high",
)

2
tests/e2e/multigpu/solo/test_flex.py
View File

@@ -86,5 +86,5 @@ class TestPackedFlex:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.1, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.1, "Train Loss (%s) is too high"
)

2
tests/e2e/multigpu/test_dist_muon_fsdp2.py
View File

@@ -37,7 +37,7 @@ def verify_training_success(temp_dir):
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"]
train_loss_df = df[df.tag == "train/loss"]
if len(train_loss_df) > 0:
final_loss = train_loss_df.value.values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (

2
tests/e2e/multigpu/test_fp8_fsdp2.py
View File

@@ -37,7 +37,7 @@ def verify_fp8_training_success(temp_dir):
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"]
train_loss_df = df[df.tag == "train/loss"]
if len(train_loss_df) > 0:
final_loss = train_loss_df.value.values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (

2
tests/e2e/multigpu/test_fsdp1.py
View File

@@ -38,7 +38,7 @@ def verify_training_success(temp_dir):
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"]
train_loss_df = df[df.tag == "train/loss"]
if len(train_loss_df) > 0:
final_loss = train_loss_df.value.values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (

2
tests/e2e/multigpu/test_fsdp2.py
View File

@@ -38,7 +38,7 @@ def verify_training_success(temp_dir):
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"]
train_loss_df = df[df.tag == "train/loss"]
if len(train_loss_df) > 0:
final_loss = train_loss_df.value.values[-1]
assert not torch.isnan(torch.tensor(final_loss)), (

2
tests/e2e/multigpu/test_gemma3.py
View File

@@ -94,5 +94,5 @@ class TestMultiGPUGemma3:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 1.8, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 1.8, "Train Loss (%s) is too high"
)

24
tests/e2e/multigpu/test_llama.py
View File

@@ -90,7 +90,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.8, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.8, "Train Loss (%s) is too high"
)
@pytest.mark.parametrize(
@@ -156,7 +156,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
def test_dpo_lora_ddp(self, temp_dir):
@@ -233,7 +233,7 @@ class TestMultiGPULlama:
loss_threshold = 2.3
check_tensorboard(
temp_dir + "/runs",
"train/train_loss",
"train/loss",
loss_threshold,
"Train Loss (%s) is too high",
)
@@ -312,7 +312,7 @@ class TestMultiGPULlama:
loss_threshold = 2.3
check_tensorboard(
temp_dir + "/runs",
"train/train_loss",
"train/loss",
loss_threshold,
"Train Loss (%s) is too high",
)
@@ -385,7 +385,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@pytest.mark.parametrize(
@@ -461,7 +461,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@require_torch_2_6_0
@@ -543,7 +543,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.1, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.1, "Train Loss (%s) is too high"
)
def test_fsdp_qlora_prequant_packed(self, temp_dir):
@@ -623,7 +623,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@pytest.mark.parametrize(
@@ -708,7 +708,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.45, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.45, "Train Loss (%s) is too high"
)
@pytest.mark.parametrize(
@@ -784,7 +784,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@pytest.mark.parametrize(
@@ -859,7 +859,7 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.5, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.5, "Train Loss (%s) is too high"
)
@pytest.mark.skip(
@@ -925,5 +925,5 @@ class TestMultiGPULlama:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 4.0, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 4.0, "Train Loss (%s) is too high"
)

6
tests/e2e/multigpu/test_ray.py
View File

@@ -79,7 +79,7 @@ class TestMultiGPURay:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@require_torch_2_7_0
@@ -138,7 +138,7 @@ class TestMultiGPURay:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)
@require_torch_2_7_0
@@ -205,5 +205,5 @@ class TestMultiGPURay:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 2.3, "Train Loss (%s) is too high"
)

2
tests/e2e/multigpu/test_tp.py
View File

@@ -64,5 +64,5 @@ class TestTensorParallel:
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 1.0, "Train Loss (%s) is too high"
temp_dir + "/runs", "train/loss", 1.0, "Train Loss (%s) is too high"
)