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

use smaller pretrained models for ci

Browse Source
This commit is contained in:
Wing Lian
2026-04-23 13:51:01 +00:00
parent 1bf65c500e
commit 431888c1de
24 changed files with 614 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
tests/e2e/test_mistral.py
View File

@@ -11,7 +11,11 @@ from axolotl.train import train
from axolotl.utils.config import normalize_config, validate_config
from axolotl.utils.dict import DictDefault
from .utils import check_model_output_exists, with_temp_dir
from .utils import (
check_model_output_exists,
check_tensorboard_loss_decreased,
with_temp_dir,
)
class TestMistral(unittest.TestCase):
@@ -23,7 +27,7 @@ class TestMistral(unittest.TestCase):
def test_lora(self, temp_dir):
cfg = DictDefault(
{
"base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
"base_model": "axolotl-ai-co/tiny-mistral-25m",
"flash_attention": True,
"sequence_len": 1024,
"load_in_8bit": True,
@@ -45,16 +49,18 @@ class TestMistral(unittest.TestCase):
},
],
"num_epochs": 2,
"micro_batch_size": 2,
"micro_batch_size": 4,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"max_steps": 20,
"save_steps": 10,
"eval_steps": 10,
"max_steps": 50,
"logging_steps": 1,
"save_steps": 50,
"eval_steps": 50,
"save_first_step": False,
"use_tensorboard": True,
}
)
@@ -64,12 +70,19 @@ class TestMistral(unittest.TestCase):
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=5,
final_window=5,
max_initial=4.5,
max_final=4.3,
)
@with_temp_dir
def test_ft(self, temp_dir):
cfg = DictDefault(
{
"base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
"base_model": "axolotl-ai-co/tiny-mistral-25m",
"flash_attention": True,
"sequence_len": 1024,
"val_set_size": 0.02,
@@ -85,16 +98,18 @@ class TestMistral(unittest.TestCase):
},
],
"num_epochs": 2,
"micro_batch_size": 2,
"micro_batch_size": 4,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"lr_scheduler": "cosine",
"max_steps": 20,
"save_steps": 10,
"eval_steps": 10,
"max_steps": 50,
"logging_steps": 1,
"save_steps": 50,
"eval_steps": 50,
"save_first_step": False,
"use_tensorboard": True,
}
)
if is_torch_bf16_gpu_available():
@@ -108,3 +123,10 @@ class TestMistral(unittest.TestCase):
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=5,
final_window=5,
max_initial=4.5,
max_final=4.3,
)