fix for protected model_ namespace w pydantic (#1345)
Browse files
README.md
CHANGED
|
@@ -546,7 +546,7 @@ base_model_ignore_patterns:
|
|
| 546 |
# You can set that here, or leave this empty to default to base_model
|
| 547 |
base_model_config: ./llama-7b-hf
|
| 548 |
# You can specify to choose a specific model revision from huggingface hub
|
| 549 |
-
|
| 550 |
# Optional tokenizer configuration path in case you want to use a different tokenizer
|
| 551 |
# than the one defined in the base model
|
| 552 |
tokenizer_config:
|
|
@@ -573,7 +573,7 @@ is_qwen_derived_model:
|
|
| 573 |
is_mistral_derived_model:
|
| 574 |
|
| 575 |
# optional overrides to the base model configuration
|
| 576 |
-
|
| 577 |
# RoPE Scaling https://github.com/huggingface/transformers/pull/24653
|
| 578 |
rope_scaling:
|
| 579 |
type: # linear | dynamic
|
|
|
|
| 546 |
# You can set that here, or leave this empty to default to base_model
|
| 547 |
base_model_config: ./llama-7b-hf
|
| 548 |
# You can specify to choose a specific model revision from huggingface hub
|
| 549 |
+
revision_of_model:
|
| 550 |
# Optional tokenizer configuration path in case you want to use a different tokenizer
|
| 551 |
# than the one defined in the base model
|
| 552 |
tokenizer_config:
|
|
|
|
| 573 |
is_mistral_derived_model:
|
| 574 |
|
| 575 |
# optional overrides to the base model configuration
|
| 576 |
+
overrides_of_model_config:
|
| 577 |
# RoPE Scaling https://github.com/huggingface/transformers/pull/24653
|
| 578 |
rope_scaling:
|
| 579 |
type: # linear | dynamic
|
src/axolotl/utils/config/__init__.py
CHANGED
|
@@ -124,7 +124,7 @@ def normalize_config(cfg):
|
|
| 124 |
(hasattr(model_config, "model_type") and model_config.model_type == "llama")
|
| 125 |
or cfg.is_llama_derived_model
|
| 126 |
or "llama" in cfg.base_model.lower()
|
| 127 |
-
or (cfg.
|
| 128 |
)
|
| 129 |
|
| 130 |
# figure out if the model is falcon
|
|
@@ -140,7 +140,7 @@ def normalize_config(cfg):
|
|
| 140 |
)
|
| 141 |
or cfg.is_falcon_derived_model
|
| 142 |
or "falcon" in cfg.base_model.lower()
|
| 143 |
-
or (cfg.
|
| 144 |
)
|
| 145 |
|
| 146 |
cfg.is_mistral_derived_model = (
|
|
@@ -153,7 +153,7 @@ def normalize_config(cfg):
|
|
| 153 |
)
|
| 154 |
or cfg.is_mistral_derived_model
|
| 155 |
or "mistral" in cfg.base_model.lower().split("/")[-1]
|
| 156 |
-
or (cfg.
|
| 157 |
)
|
| 158 |
|
| 159 |
cfg.is_qwen_derived_model = (
|
|
@@ -379,11 +379,11 @@ def legacy_validate_config(cfg):
|
|
| 379 |
"hub_model_id is set without any models being saved. To save a model, set either save_steps or saves_per_epoch."
|
| 380 |
)
|
| 381 |
|
| 382 |
-
if cfg.gptq and cfg.
|
| 383 |
raise ValueError(
|
| 384 |
-
"
|
| 385 |
+ "Please download the model from HuggingFace Hub manually for correct branch, "
|
| 386 |
-
+ "point to its path, and remove
|
| 387 |
)
|
| 388 |
|
| 389 |
# if cfg.sample_packing and cfg.sdp_attention:
|
|
|
|
| 124 |
(hasattr(model_config, "model_type") and model_config.model_type == "llama")
|
| 125 |
or cfg.is_llama_derived_model
|
| 126 |
or "llama" in cfg.base_model.lower()
|
| 127 |
+
or (cfg.type_of_model and "llama" in cfg.type_of_model.lower())
|
| 128 |
)
|
| 129 |
|
| 130 |
# figure out if the model is falcon
|
|
|
|
| 140 |
)
|
| 141 |
or cfg.is_falcon_derived_model
|
| 142 |
or "falcon" in cfg.base_model.lower()
|
| 143 |
+
or (cfg.type_of_model and "rwforcausallm" in cfg.type_of_model.lower())
|
| 144 |
)
|
| 145 |
|
| 146 |
cfg.is_mistral_derived_model = (
|
|
|
|
| 153 |
)
|
| 154 |
or cfg.is_mistral_derived_model
|
| 155 |
or "mistral" in cfg.base_model.lower().split("/")[-1]
|
| 156 |
+
or (cfg.type_of_model and "mistral" in cfg.type_of_model.lower())
|
| 157 |
)
|
| 158 |
|
| 159 |
cfg.is_qwen_derived_model = (
|
|
|
|
| 379 |
"hub_model_id is set without any models being saved. To save a model, set either save_steps or saves_per_epoch."
|
| 380 |
)
|
| 381 |
|
| 382 |
+
if cfg.gptq and cfg.revision_of_model:
|
| 383 |
raise ValueError(
|
| 384 |
+
"revision_of_model is not supported for GPTQ models. "
|
| 385 |
+ "Please download the model from HuggingFace Hub manually for correct branch, "
|
| 386 |
+
+ "point to its path, and remove revision_of_model from the config."
|
| 387 |
)
|
| 388 |
|
| 389 |
# if cfg.sample_packing and cfg.sdp_attention:
|
src/axolotl/utils/config/models/input/v0_4_1/__init__.py
CHANGED
|
@@ -47,6 +47,16 @@ class DeprecatedParameters(BaseModel):
|
|
| 47 |
return noisy_embedding_alpha
|
| 48 |
|
| 49 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
class PretrainingDataset(BaseModel):
|
| 51 |
"""pretraining dataset configuration subset"""
|
| 52 |
|
|
@@ -234,12 +244,8 @@ class ModelInputConfig(BaseModel):
|
|
| 234 |
tokenizer_type: Optional[str] = Field(
|
| 235 |
default=None, metadata={"help": "transformers tokenizer class"}
|
| 236 |
)
|
| 237 |
-
model_type: Optional[str] = Field(default=None)
|
| 238 |
-
model_revision: Optional[str] = None
|
| 239 |
trust_remote_code: Optional[bool] = None
|
| 240 |
|
| 241 |
-
model_config_overrides: Optional[Dict[str, Any]] = None
|
| 242 |
-
|
| 243 |
@field_validator("trust_remote_code")
|
| 244 |
@classmethod
|
| 245 |
def hint_trust_remote_code(cls, trust_remote_code):
|
|
@@ -362,11 +368,17 @@ class AxolotlInputConfig(
|
|
| 362 |
HyperparametersConfig,
|
| 363 |
WandbConfig,
|
| 364 |
MLFlowConfig,
|
|
|
|
| 365 |
DeprecatedParameters,
|
| 366 |
BaseModel,
|
| 367 |
):
|
| 368 |
"""wrapper of all config options"""
|
| 369 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 370 |
strict: Optional[bool] = Field(default=False)
|
| 371 |
resume_from_checkpoint: Optional[str] = None
|
| 372 |
auto_resume_from_checkpoints: Optional[bool] = None
|
|
@@ -550,11 +562,11 @@ class AxolotlInputConfig(
|
|
| 550 |
@model_validator(mode="before")
|
| 551 |
@classmethod
|
| 552 |
def check_gptq_w_revision(cls, data):
|
| 553 |
-
if data.get("gptq") and data.get("
|
| 554 |
raise ValueError(
|
| 555 |
-
"
|
| 556 |
+ "Please download the model from HuggingFace Hub manually for correct branch, "
|
| 557 |
-
+ "point to its path, and remove
|
| 558 |
)
|
| 559 |
return data
|
| 560 |
|
|
|
|
| 47 |
return noisy_embedding_alpha
|
| 48 |
|
| 49 |
|
| 50 |
+
class RemappedParameters(BaseModel):
|
| 51 |
+
"""parameters that have been remapped to other names"""
|
| 52 |
+
|
| 53 |
+
overrides_of_model_config: Optional[Dict[str, Any]] = Field(
|
| 54 |
+
default=None, alias="model_config"
|
| 55 |
+
)
|
| 56 |
+
type_of_model: Optional[str] = Field(default=None, alias="model_type")
|
| 57 |
+
revision_of_model: Optional[str] = Field(default=None, alias="model_revision")
|
| 58 |
+
|
| 59 |
+
|
| 60 |
class PretrainingDataset(BaseModel):
|
| 61 |
"""pretraining dataset configuration subset"""
|
| 62 |
|
|
|
|
| 244 |
tokenizer_type: Optional[str] = Field(
|
| 245 |
default=None, metadata={"help": "transformers tokenizer class"}
|
| 246 |
)
|
|
|
|
|
|
|
| 247 |
trust_remote_code: Optional[bool] = None
|
| 248 |
|
|
|
|
|
|
|
| 249 |
@field_validator("trust_remote_code")
|
| 250 |
@classmethod
|
| 251 |
def hint_trust_remote_code(cls, trust_remote_code):
|
|
|
|
| 368 |
HyperparametersConfig,
|
| 369 |
WandbConfig,
|
| 370 |
MLFlowConfig,
|
| 371 |
+
RemappedParameters,
|
| 372 |
DeprecatedParameters,
|
| 373 |
BaseModel,
|
| 374 |
):
|
| 375 |
"""wrapper of all config options"""
|
| 376 |
|
| 377 |
+
class Config:
|
| 378 |
+
"""Config for alias"""
|
| 379 |
+
|
| 380 |
+
populate_by_name = True
|
| 381 |
+
|
| 382 |
strict: Optional[bool] = Field(default=False)
|
| 383 |
resume_from_checkpoint: Optional[str] = None
|
| 384 |
auto_resume_from_checkpoints: Optional[bool] = None
|
|
|
|
| 562 |
@model_validator(mode="before")
|
| 563 |
@classmethod
|
| 564 |
def check_gptq_w_revision(cls, data):
|
| 565 |
+
if data.get("gptq") and data.get("revision_of_model"):
|
| 566 |
raise ValueError(
|
| 567 |
+
"revision_of_model is not supported for GPTQ models. "
|
| 568 |
+ "Please download the model from HuggingFace Hub manually for correct branch, "
|
| 569 |
+
+ "point to its path, and remove revision_of_model from the config."
|
| 570 |
)
|
| 571 |
return data
|
| 572 |
|
src/axolotl/utils/models.py
CHANGED
|
@@ -86,8 +86,8 @@ def load_model_config(cfg):
|
|
| 86 |
model_config_name = cfg.tokenizer_config
|
| 87 |
trust_remote_code = cfg.trust_remote_code is True
|
| 88 |
config_kwargs = {}
|
| 89 |
-
if cfg.
|
| 90 |
-
config_kwargs["revision"] = cfg.
|
| 91 |
|
| 92 |
try:
|
| 93 |
model_config = AutoConfig.from_pretrained(
|
|
@@ -104,8 +104,8 @@ def load_model_config(cfg):
|
|
| 104 |
)
|
| 105 |
raise err
|
| 106 |
|
| 107 |
-
if cfg.
|
| 108 |
-
for key, val in cfg.
|
| 109 |
setattr(model_config, key, val)
|
| 110 |
|
| 111 |
check_model_config(cfg, model_config)
|
|
@@ -272,7 +272,7 @@ def load_model(
|
|
| 272 |
Load a model for a given configuration and tokenizer.
|
| 273 |
"""
|
| 274 |
base_model = cfg.base_model
|
| 275 |
-
model_type = cfg.
|
| 276 |
model_config = load_model_config(cfg)
|
| 277 |
|
| 278 |
# TODO refactor as a kwarg
|
|
@@ -426,8 +426,8 @@ def load_model(
|
|
| 426 |
if is_deepspeed_zero3_enabled():
|
| 427 |
del model_kwargs["device_map"]
|
| 428 |
|
| 429 |
-
if cfg.
|
| 430 |
-
model_kwargs["revision"] = cfg.
|
| 431 |
if cfg.gptq:
|
| 432 |
if not hasattr(model_config, "quantization_config"):
|
| 433 |
LOG.warning("model config does not contain quantization_config information")
|
|
|
|
| 86 |
model_config_name = cfg.tokenizer_config
|
| 87 |
trust_remote_code = cfg.trust_remote_code is True
|
| 88 |
config_kwargs = {}
|
| 89 |
+
if cfg.revision_of_model:
|
| 90 |
+
config_kwargs["revision"] = cfg.revision_of_model
|
| 91 |
|
| 92 |
try:
|
| 93 |
model_config = AutoConfig.from_pretrained(
|
|
|
|
| 104 |
)
|
| 105 |
raise err
|
| 106 |
|
| 107 |
+
if cfg.overrides_of_model_config:
|
| 108 |
+
for key, val in cfg.overrides_of_model_config.items():
|
| 109 |
setattr(model_config, key, val)
|
| 110 |
|
| 111 |
check_model_config(cfg, model_config)
|
|
|
|
| 272 |
Load a model for a given configuration and tokenizer.
|
| 273 |
"""
|
| 274 |
base_model = cfg.base_model
|
| 275 |
+
model_type = cfg.type_of_model
|
| 276 |
model_config = load_model_config(cfg)
|
| 277 |
|
| 278 |
# TODO refactor as a kwarg
|
|
|
|
| 426 |
if is_deepspeed_zero3_enabled():
|
| 427 |
del model_kwargs["device_map"]
|
| 428 |
|
| 429 |
+
if cfg.revision_of_model:
|
| 430 |
+
model_kwargs["revision"] = cfg.revision_of_model
|
| 431 |
if cfg.gptq:
|
| 432 |
if not hasattr(model_config, "quantization_config"):
|
| 433 |
LOG.warning("model config does not contain quantization_config information")
|
tests/test_validation.py
CHANGED
|
@@ -3,6 +3,7 @@
|
|
| 3 |
|
| 4 |
import logging
|
| 5 |
import os
|
|
|
|
| 6 |
from typing import Optional
|
| 7 |
|
| 8 |
import pytest
|
|
@@ -14,6 +15,8 @@ from axolotl.utils.dict import DictDefault
|
|
| 14 |
from axolotl.utils.models import check_model_config
|
| 15 |
from axolotl.utils.wandb_ import setup_wandb_env_vars
|
| 16 |
|
|
|
|
|
|
|
| 17 |
|
| 18 |
@pytest.fixture(name="minimal_cfg")
|
| 19 |
def fixture_cfg():
|
|
@@ -190,6 +193,45 @@ class TestValidation(BaseValidation):
|
|
| 190 |
|
| 191 |
assert new_cfg.learning_rate == 0.00005
|
| 192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
def test_qlora(self, minimal_cfg):
|
| 194 |
base_cfg = (
|
| 195 |
DictDefault(
|
|
|
|
| 3 |
|
| 4 |
import logging
|
| 5 |
import os
|
| 6 |
+
import warnings
|
| 7 |
from typing import Optional
|
| 8 |
|
| 9 |
import pytest
|
|
|
|
| 15 |
from axolotl.utils.models import check_model_config
|
| 16 |
from axolotl.utils.wandb_ import setup_wandb_env_vars
|
| 17 |
|
| 18 |
+
warnings.filterwarnings("error")
|
| 19 |
+
|
| 20 |
|
| 21 |
@pytest.fixture(name="minimal_cfg")
|
| 22 |
def fixture_cfg():
|
|
|
|
| 193 |
|
| 194 |
assert new_cfg.learning_rate == 0.00005
|
| 195 |
|
| 196 |
+
def test_model_config_remap(self, minimal_cfg):
|
| 197 |
+
cfg = (
|
| 198 |
+
DictDefault(
|
| 199 |
+
{
|
| 200 |
+
"model_config": {"model_type": "mistral"},
|
| 201 |
+
}
|
| 202 |
+
)
|
| 203 |
+
| minimal_cfg
|
| 204 |
+
)
|
| 205 |
+
|
| 206 |
+
new_cfg = validate_config(cfg)
|
| 207 |
+
assert new_cfg.overrides_of_model_config["model_type"] == "mistral"
|
| 208 |
+
|
| 209 |
+
def test_model_type_remap(self, minimal_cfg):
|
| 210 |
+
cfg = (
|
| 211 |
+
DictDefault(
|
| 212 |
+
{
|
| 213 |
+
"model_type": "AutoModelForCausalLM",
|
| 214 |
+
}
|
| 215 |
+
)
|
| 216 |
+
| minimal_cfg
|
| 217 |
+
)
|
| 218 |
+
|
| 219 |
+
new_cfg = validate_config(cfg)
|
| 220 |
+
assert new_cfg.type_of_model == "AutoModelForCausalLM"
|
| 221 |
+
|
| 222 |
+
def test_model_revision_remap(self, minimal_cfg):
|
| 223 |
+
cfg = (
|
| 224 |
+
DictDefault(
|
| 225 |
+
{
|
| 226 |
+
"model_revision": "main",
|
| 227 |
+
}
|
| 228 |
+
)
|
| 229 |
+
| minimal_cfg
|
| 230 |
+
)
|
| 231 |
+
|
| 232 |
+
new_cfg = validate_config(cfg)
|
| 233 |
+
assert new_cfg.revision_of_model == "main"
|
| 234 |
+
|
| 235 |
def test_qlora(self, minimal_cfg):
|
| 236 |
base_cfg = (
|
| 237 |
DictDefault(
|