1eaf4d7418
* mxfp4 axo * import lint * test for qat mxfp4 * config for mxfp4 * add qat: * pass base config * MXFakeQuantizeConfig * lint * tune config so it fits in 32GB VRAM --------- Co-authored-by: Wing Lian <wing@axolotl.ai>
164 lines
5.7 KiB
Python
164 lines
5.7 KiB
Python
"""
|
|
E2E tests for QAT
|
|
"""
|
|
|
|
from pathlib import Path
|
|
|
|
from axolotl.common.datasets import load_datasets, load_preference_datasets
|
|
from axolotl.train import train
|
|
from axolotl.utils.config import normalize_config, validate_config
|
|
from axolotl.utils.dict import DictDefault
|
|
from axolotl.utils.schemas.enums import TorchAOQuantDType
|
|
from axolotl.utils.schemas.quantization import QATConfig, validate_ao_dtype
|
|
|
|
from .utils import check_model_output_exists, check_tensorboard
|
|
|
|
|
|
class TestQATLlama:
|
|
"""
|
|
Test case for QAT Llama models
|
|
"""
|
|
|
|
def test_qat(self, temp_dir):
|
|
cfg = DictDefault(
|
|
{
|
|
"base_model": "HuggingFaceTB/SmolLM2-135M",
|
|
"tokenizer_type": "AutoTokenizer",
|
|
"sequence_len": 1024,
|
|
"special_tokens": {
|
|
"pad_token": "<|endoftext|>",
|
|
},
|
|
"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%]",
|
|
},
|
|
],
|
|
"chat_template": "chatml",
|
|
"qat": {
|
|
"quantize_embedding": True,
|
|
"activation_dtype": "int8",
|
|
"weight_dtype": "int4",
|
|
"group_size": 8,
|
|
},
|
|
"num_epochs": 1,
|
|
"micro_batch_size": 1,
|
|
"gradient_accumulation_steps": 2,
|
|
"output_dir": temp_dir,
|
|
"learning_rate": 0.00001,
|
|
"optimizer": "adamw_bnb_8bit",
|
|
"lr_scheduler": "cosine",
|
|
"max_steps": 5,
|
|
"bf16": True,
|
|
"save_first_step": False,
|
|
}
|
|
)
|
|
cfg = validate_config(cfg)
|
|
normalize_config(cfg)
|
|
dataset_meta = load_datasets(cfg=cfg)
|
|
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
check_model_output_exists(Path(temp_dir) / "checkpoint-5", cfg)
|
|
|
|
def test_qat_dpo(self, temp_dir):
|
|
cfg = DictDefault(
|
|
{
|
|
"base_model": "HuggingFaceTB/SmolLM2-135M",
|
|
"sequence_len": 2048,
|
|
"sample_packing": False,
|
|
"eval_sample_packing": False,
|
|
"pad_to_sequence_len": True,
|
|
"val_set_size": 0.01,
|
|
"special_tokens": {
|
|
"pad_token": "<|endoftext|>",
|
|
},
|
|
"rl": "dpo",
|
|
"chat_template": "chatml",
|
|
"datasets": [
|
|
{
|
|
"path": "fozziethebeat/alpaca_messages_2k_dpo_test",
|
|
"type": "chat_template.default",
|
|
"field_messages": "conversation",
|
|
"field_chosen": "chosen",
|
|
"field_rejected": "rejected",
|
|
"message_field_role": "role",
|
|
"message_field_content": "content",
|
|
"roles": {
|
|
"system": ["system"],
|
|
"user": ["user"],
|
|
"assistant": ["assistant"],
|
|
},
|
|
},
|
|
],
|
|
"num_epochs": 1,
|
|
"max_steps": 5,
|
|
"micro_batch_size": 2,
|
|
"gradient_accumulation_steps": 2,
|
|
"output_dir": temp_dir,
|
|
"warmup_steps": 0,
|
|
"learning_rate": 0.00001,
|
|
"optimizer": "adamw_torch_fused",
|
|
"lr_scheduler": "cosine",
|
|
"flash_attention": True,
|
|
"use_tensorboard": True,
|
|
"bf16": True,
|
|
"qat": {
|
|
"quantize_embedding": True,
|
|
"activation_dtype": "int8",
|
|
"weight_dtype": "int4",
|
|
"group_size": 8,
|
|
},
|
|
"save_first_step": False,
|
|
}
|
|
)
|
|
cfg = validate_config(cfg)
|
|
normalize_config(cfg)
|
|
dataset_meta = load_preference_datasets(cfg=cfg)
|
|
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
check_model_output_exists(Path(temp_dir) / "checkpoint-5", cfg)
|
|
|
|
loss_threshold = 2.3
|
|
check_tensorboard(
|
|
temp_dir + "/runs",
|
|
"train/train_loss",
|
|
loss_threshold,
|
|
"Train Loss (%s) is too high",
|
|
)
|
|
|
|
|
|
class TestMXFP4Schema:
|
|
"""Test MXFP4 schema validation"""
|
|
|
|
def test_validate_mxfp4_dtype(self):
|
|
result = validate_ao_dtype("mxfp4")
|
|
assert result == TorchAOQuantDType.mxfp4
|
|
|
|
def test_qat_config_with_mxfp4(self):
|
|
"""Test QATConfig accepts mxfp4 weight_dtype"""
|
|
config = QATConfig(
|
|
weight_dtype="mxfp4",
|
|
group_size=32,
|
|
quantize_embedding=False,
|
|
)
|
|
assert config.weight_dtype == TorchAOQuantDType.mxfp4
|
|
assert config.group_size == 32
|
|
|
|
def test_qat_config_mxfp4_invalid_group_size(self):
|
|
"""Test that invalid group_size raises appropriate error during quantization"""
|
|
# Note: Schema validation doesn't check group_size compatibility,
|
|
# that happens in get_quantization_config
|
|
config = QATConfig(
|
|
weight_dtype="mxfp4",
|
|
group_size=16, # Invalid for mxfp4, but schema allows it
|
|
)
|
|
assert config.group_size == 16 # Schema accepts it
|
|
# Actual validation happens at runtime in get_quantization_config
|