Merge pull request #36 from OpenAccess-AI-Collective/qlora
Browse files- requirements.txt +1 -1
- scripts/finetune.py +3 -0
- src/axolotl/prompters.py +5 -0
- src/axolotl/utils/data.py +5 -0
- src/axolotl/utils/models.py +21 -3
requirements.txt
CHANGED
@@ -1,10 +1,10 @@
|
|
1 |
peft @ git+https://github.com/huggingface/peft.git
|
2 |
transformers @ git+https://github.com/huggingface/transformers.git
|
|
|
3 |
attrdict
|
4 |
fire
|
5 |
PyYAML==6.0
|
6 |
black
|
7 |
-
bitsandbytes==0.37.2
|
8 |
datasets
|
9 |
accelerate>=0.19.0
|
10 |
sentencepiece
|
|
|
1 |
peft @ git+https://github.com/huggingface/peft.git
|
2 |
transformers @ git+https://github.com/huggingface/transformers.git
|
3 |
+
bitsandbytes>=0.39.0
|
4 |
attrdict
|
5 |
fire
|
6 |
PyYAML==6.0
|
7 |
black
|
|
|
8 |
datasets
|
9 |
accelerate>=0.19.0
|
10 |
sentencepiece
|
scripts/finetune.py
CHANGED
@@ -14,6 +14,7 @@ from attrdict import AttrDefault
|
|
14 |
|
15 |
# add src to the pythonpath so we don't need to pip install this
|
16 |
from axolotl.utils.tokenization import check_dataset_labels
|
|
|
17 |
|
18 |
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
|
19 |
src_dir = os.path.join(project_root, "src")
|
@@ -158,6 +159,8 @@ def train(
|
|
158 |
cfg.fp16 = True
|
159 |
cfg.bf16 = False
|
160 |
|
|
|
|
|
161 |
# Load the model and tokenizer
|
162 |
logging.info("loading model, tokenizer, and peft_config...")
|
163 |
model, tokenizer, peft_config = load_model(
|
|
|
14 |
|
15 |
# add src to the pythonpath so we don't need to pip install this
|
16 |
from axolotl.utils.tokenization import check_dataset_labels
|
17 |
+
from axolotl.utils.validation import validate_config
|
18 |
|
19 |
project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
|
20 |
src_dir = os.path.join(project_root, "src")
|
|
|
159 |
cfg.fp16 = True
|
160 |
cfg.bf16 = False
|
161 |
|
162 |
+
validate_config(cfg)
|
163 |
+
|
164 |
# Load the model and tokenizer
|
165 |
logging.info("loading model, tokenizer, and peft_config...")
|
166 |
model, tokenizer, peft_config = load_model(
|
src/axolotl/prompters.py
CHANGED
@@ -11,6 +11,7 @@ class PromptStyle(Enum):
|
|
11 |
instruct = "instruct"
|
12 |
chat = "chat"
|
13 |
|
|
|
14 |
class AlpacaPrompter:
|
15 |
system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
|
16 |
system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
|
@@ -50,6 +51,10 @@ class AlpacaPrompter:
|
|
50 |
return output.split(self.response_split)[1].strip()
|
51 |
|
52 |
|
|
|
|
|
|
|
|
|
53 |
class JeopardyPrompter(AlpacaPrompter):
|
54 |
prompt_input = "Below is a Jeopardy clue paired with input providing the category of the clue. Write a concise response that best answers tbe clue given the category.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n"
|
55 |
|
|
|
11 |
instruct = "instruct"
|
12 |
chat = "chat"
|
13 |
|
14 |
+
|
15 |
class AlpacaPrompter:
|
16 |
system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
|
17 |
system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
|
|
|
51 |
return output.split(self.response_split)[1].strip()
|
52 |
|
53 |
|
54 |
+
class UnpromptedPrompter(AlpacaPrompter):
|
55 |
+
system_prompt = ""
|
56 |
+
system_no_input_prompt = ""
|
57 |
+
|
58 |
class JeopardyPrompter(AlpacaPrompter):
|
59 |
prompt_input = "Below is a Jeopardy clue paired with input providing the category of the clue. Write a concise response that best answers tbe clue given the category.\n\n### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n"
|
60 |
|
src/axolotl/utils/data.py
CHANGED
@@ -98,6 +98,11 @@ def load_tokenized_prepared_datasets(tokenizer, cfg, default_dataset_prepared_pa
|
|
98 |
ds = load_dataset("json", data_files=fp, streaming=False, split=None)
|
99 |
if not ds:
|
100 |
raise Exception("unhandled dataset load")
|
|
|
|
|
|
|
|
|
|
|
101 |
d_type = d.type
|
102 |
d_type_split = d_type.split(":")
|
103 |
d_base_type = d_type_split[0]
|
|
|
98 |
ds = load_dataset("json", data_files=fp, streaming=False, split=None)
|
99 |
if not ds:
|
100 |
raise Exception("unhandled dataset load")
|
101 |
+
# support for using a subset of the data
|
102 |
+
if d.shards:
|
103 |
+
ds = ds.shuffle(seed=42)["train"].shard(
|
104 |
+
num_shards=cfg.shards, index=0
|
105 |
+
)
|
106 |
d_type = d.type
|
107 |
d_type_split = d_type.split(":")
|
108 |
d_base_type = d_type_split[0]
|
src/axolotl/utils/models.py
CHANGED
@@ -6,11 +6,12 @@ from typing import Optional, Tuple, TYPE_CHECKING
|
|
6 |
|
7 |
import torch
|
8 |
import transformers
|
|
|
9 |
from transformers import (
|
10 |
AutoModelForCausalLM,
|
11 |
AutoTokenizer,
|
12 |
PreTrainedModel,
|
13 |
-
AutoConfig,
|
14 |
)
|
15 |
|
16 |
try:
|
@@ -81,6 +82,16 @@ def load_model(
|
|
81 |
logging.exception(e)
|
82 |
raise e
|
83 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
try:
|
85 |
if cfg.load_4bit and is_llama_derived_model:
|
86 |
from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
|
@@ -123,8 +134,10 @@ def load_model(
|
|
123 |
model = LlamaForCausalLM.from_pretrained(
|
124 |
base_model,
|
125 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
|
|
126 |
torch_dtype=torch_dtype,
|
127 |
device_map=cfg.device_map,
|
|
|
128 |
)
|
129 |
# elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
|
130 |
# This is a WIP, still an issue with the backward pass
|
@@ -156,9 +169,11 @@ def load_model(
|
|
156 |
model = getattr(transformers, model_type).from_pretrained(
|
157 |
base_model,
|
158 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
|
|
159 |
torch_dtype=torch_dtype,
|
160 |
device_map=cfg.device_map,
|
161 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
|
|
162 |
)
|
163 |
else:
|
164 |
config = AutoConfig.from_pretrained(
|
@@ -169,9 +184,11 @@ def load_model(
|
|
169 |
base_model,
|
170 |
config=config,
|
171 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
|
|
172 |
torch_dtype=torch_dtype,
|
173 |
device_map=cfg.device_map,
|
174 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
|
|
175 |
)
|
176 |
except Exception as e:
|
177 |
logging.error(
|
@@ -184,6 +201,7 @@ def load_model(
|
|
184 |
torch_dtype=torch_dtype,
|
185 |
device_map=cfg.device_map,
|
186 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
|
|
187 |
)
|
188 |
|
189 |
if not tokenizer:
|
@@ -225,7 +243,7 @@ def load_model(
|
|
225 |
embeddings_len = math.ceil(len(tokenizer) / 32) * 32
|
226 |
model.resize_token_embeddings(embeddings_len)
|
227 |
|
228 |
-
if cfg.adapter and load_in_8bit and not cfg.load_4bit:
|
229 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
230 |
model = prepare_model_for_int8_training(model)
|
231 |
|
@@ -270,7 +288,7 @@ def load_adapter(model, cfg, adapter):
|
|
270 |
|
271 |
if adapter is None:
|
272 |
return model, None
|
273 |
-
if adapter == "lora":
|
274 |
return load_lora(model, cfg)
|
275 |
if adapter == "llama-adapter":
|
276 |
return load_llama_adapter(model, cfg)
|
|
|
6 |
|
7 |
import torch
|
8 |
import transformers
|
9 |
+
from torch import nn
|
10 |
from transformers import (
|
11 |
AutoModelForCausalLM,
|
12 |
AutoTokenizer,
|
13 |
PreTrainedModel,
|
14 |
+
AutoConfig, BitsAndBytesConfig,
|
15 |
)
|
16 |
|
17 |
try:
|
|
|
82 |
logging.exception(e)
|
83 |
raise e
|
84 |
|
85 |
+
model_kwargs = {}
|
86 |
+
if cfg.adapter == "qlora":
|
87 |
+
model_kwargs["quantization_config"] = BitsAndBytesConfig(
|
88 |
+
load_in_4bit=True,
|
89 |
+
llm_int8_threshold=6.0,
|
90 |
+
llm_int8_has_fp16_weight=False,
|
91 |
+
bnb_4bit_compute_dtype=torch.float16,
|
92 |
+
bnb_4bit_use_double_quant=True,
|
93 |
+
bnb_4bit_quant_type="nf4",
|
94 |
+
)
|
95 |
try:
|
96 |
if cfg.load_4bit and is_llama_derived_model:
|
97 |
from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
|
|
|
134 |
model = LlamaForCausalLM.from_pretrained(
|
135 |
base_model,
|
136 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
137 |
+
load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
|
138 |
torch_dtype=torch_dtype,
|
139 |
device_map=cfg.device_map,
|
140 |
+
**model_kwargs,
|
141 |
)
|
142 |
# elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
|
143 |
# This is a WIP, still an issue with the backward pass
|
|
|
169 |
model = getattr(transformers, model_type).from_pretrained(
|
170 |
base_model,
|
171 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
172 |
+
load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
|
173 |
torch_dtype=torch_dtype,
|
174 |
device_map=cfg.device_map,
|
175 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
176 |
+
**model_kwargs,
|
177 |
)
|
178 |
else:
|
179 |
config = AutoConfig.from_pretrained(
|
|
|
184 |
base_model,
|
185 |
config=config,
|
186 |
load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
|
187 |
+
load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
|
188 |
torch_dtype=torch_dtype,
|
189 |
device_map=cfg.device_map,
|
190 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
191 |
+
**model_kwargs,
|
192 |
)
|
193 |
except Exception as e:
|
194 |
logging.error(
|
|
|
201 |
torch_dtype=torch_dtype,
|
202 |
device_map=cfg.device_map,
|
203 |
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
204 |
+
**model_kwargs,
|
205 |
)
|
206 |
|
207 |
if not tokenizer:
|
|
|
243 |
embeddings_len = math.ceil(len(tokenizer) / 32) * 32
|
244 |
model.resize_token_embeddings(embeddings_len)
|
245 |
|
246 |
+
if ((cfg.adapter == "lora" and load_in_8bit) or cfg.adapter == "qlora") and not cfg.load_4bit:
|
247 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
248 |
model = prepare_model_for_int8_training(model)
|
249 |
|
|
|
288 |
|
289 |
if adapter is None:
|
290 |
return model, None
|
291 |
+
if adapter == "lora" or adapter == "qlora":
|
292 |
return load_lora(model, cfg)
|
293 |
if adapter == "llama-adapter":
|
294 |
return load_llama_adapter(model, cfg)
|