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

Feat: Add support for gemma3_text and add e2e for gemma2 (#2406)

Browse Source
This commit is contained in:
NanoCode012
2025-03-23 07:33:21 +07:00
committed by GitHub
parent 86bac48d14
commit 9f00465a5c
12 changed files with 348 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py
View File

@@ -144,7 +144,7 @@ def test_swiglu_mlp_integration(small_llama_model):
def test_geglu_model_integration():
"""Test GeGLU activation with Gemma model."""
model = AutoModelForCausalLM.from_pretrained(
"mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda"
"mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="auto"
)
peft_config = get_peft_config(
{
@@ -347,7 +347,7 @@ def test_model_architecture(model_config):
"""Test LoRA kernel patches across different model architectures."""
# Load model with appropriate dtype
model = AutoModelForCausalLM.from_pretrained(
model_config["name"], torch_dtype=model_config["dtype"], device_map="cuda"
model_config["name"], torch_dtype=model_config["dtype"], device_map="auto"
)
# Apply LoRA configuration

2
tests/e2e/test_deepseekv3.py
View File

@@ -1,5 +1,5 @@
"""
E2E tests for lora llama
E2E tests for deepseekv3
"""
import logging

133
tests/e2e/test_gemma2.py Normal file
View File

@@ -0,0 +1,133 @@
"""
E2E tests for gemma2
"""
import logging
import os
from pathlib import Path
import pytest
from axolotl.cli.args import TrainerCliArgs
from axolotl.common.datasets import load_datasets
from axolotl.train import train
from axolotl.utils.config import normalize_config, validate_config
from axolotl.utils.dict import DictDefault
LOG = logging.getLogger("axolotl.tests.e2e")
os.environ["WANDB_DISABLED"] = "true"
class TestGemma2:
"""
Test case for Gemma2 models
"""
@pytest.mark.parametrize(
"sample_packing",
[True, False],
)
def test_lora_gemma2(self, temp_dir, sample_packing):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "axolotl-ai-co/gemma-2-33M",
"trust_remote_code": True,
"sample_packing": sample_packing,
"flash_attention": True,
"sequence_len": 2048,
"adapter": "lora",
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"lora_target_linear": True,
"val_set_size": 0,
"datasets": [
{
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"field_messages": "conversations",
"message_property_mappings": {
"role": "from",
"content": "value",
},
"drop_system_message": True,
"split": "train[:1%]",
},
],
"special_tokens": {
"bos_token": "<bos>",
"eos_token": "<eos>",
},
"chat_template": "gemma", # gemma2's template is same as gemma
"num_epochs": 1,
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_bnb_8bit",
"lr_scheduler": "cosine",
"max_steps": 5,
"save_safetensors": True,
"bf16": True,
}
)
cfg = validate_config(cfg)
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, dataset_meta=dataset_meta)
assert (Path(temp_dir) / "adapter_model.safetensors").exists()
@pytest.mark.parametrize(
"sample_packing",
[True, False],
)
def test_fft_gemma2(self, temp_dir, sample_packing):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "axolotl-ai-co/gemma-2-33M",
"trust_remote_code": True,
"sample_packing": sample_packing,
"flash_attention": True,
"sequence_len": 2048,
"val_set_size": 0,
"datasets": [
{
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"field_messages": "conversations",
"message_property_mappings": {
"role": "from",
"content": "value",
},
"split": "train[:1%]",
"drop_system_message": True,
},
],
"chat_template": "gemma", # gemma2's template is same as gemma
"special_tokens": {
"bos_token": "<bos>",
"eos_token": "<eos>",
},
"num_epochs": 1,
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_bnb_8bit",
"lr_scheduler": "cosine",
"max_steps": 5,
"save_safetensors": True,
"bf16": True,
}
)
cfg = validate_config(cfg)
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, dataset_meta=dataset_meta)
assert (Path(temp_dir) / "model.safetensors").exists()

131
tests/e2e/test_gemma3_text.py Normal file
View File

@@ -0,0 +1,131 @@
"""
E2E tests for gemma3_text
"""
import logging
import os
from pathlib import Path
import pytest
from axolotl.cli.args import TrainerCliArgs
from axolotl.common.datasets import load_datasets
from axolotl.train import train
from axolotl.utils.config import normalize_config, validate_config
from axolotl.utils.dict import DictDefault
LOG = logging.getLogger("axolotl.tests.e2e")
os.environ["WANDB_DISABLED"] = "true"
class TestGemma3Text:
"""
Test case for Gemma3Text models
"""
@pytest.mark.parametrize(
"sample_packing",
[True, False],
)
def test_lora_gemma3_text(self, temp_dir, sample_packing):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "axolotl-ai-co/gemma-3-34M",
"trust_remote_code": True,
"sample_packing": sample_packing,
"flash_attention": True,
"sequence_len": 2048,
"adapter": "lora",
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"lora_target_linear": True,
"val_set_size": 0,
"datasets": [
{
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"field_messages": "conversations",
"message_property_mappings": {
"role": "from",
"content": "value",
},
"split": "train[:1%]",
},
],
"special_tokens": {
"bos_token": "<bos>",
"eos_token": "<eos>",
},
"chat_template": "gemma3_text",
"num_epochs": 1,
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_bnb_8bit",
"lr_scheduler": "cosine",
"max_steps": 5,
"save_safetensors": True,
"bf16": True,
}
)
cfg = validate_config(cfg)
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, dataset_meta=dataset_meta)
assert (Path(temp_dir) / "adapter_model.safetensors").exists()
@pytest.mark.parametrize(
"sample_packing",
[True, False],
)
def test_fft_gemma3_text(self, temp_dir, sample_packing):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "axolotl-ai-co/gemma-3-34M",
"trust_remote_code": True,
"sample_packing": sample_packing,
"flash_attention": True,
"sequence_len": 2048,
"val_set_size": 0,
"datasets": [
{
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"field_messages": "conversations",
"message_property_mappings": {
"role": "from",
"content": "value",
},
"split": "train[:1%]",
},
],
"chat_template": "gemma3_text",
"special_tokens": {
"bos_token": "<bos>",
"eos_token": "<eos>",
},
"num_epochs": 1,
"micro_batch_size": 1,
"gradient_accumulation_steps": 4,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_bnb_8bit",
"lr_scheduler": "cosine",
"max_steps": 5,
"save_safetensors": True,
"bf16": True,
}
)
cfg = validate_config(cfg)
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, dataset_meta=dataset_meta)
assert (Path(temp_dir) / "model.safetensors").exists()

2
tests/e2e/test_schedulers.py
View File

@@ -54,7 +54,7 @@ class TestCustomSchedulers(unittest.TestCase):
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_hf",
"optimizer": "adamw_torch_fused",
"max_steps": 20,
"lr_scheduler": "rex",
"warmup_steps": 5,