Thytu
commited on
Commit
•
dd00657
1
Parent(s):
c3d2562
refactor(param): rename load_4bit config param by gptq
Browse files- README.md +1 -1
- configs/quickstart.yml +1 -1
- examples/4bit-lora-7b/config.yml +1 -1
- src/axolotl/utils/models.py +5 -5
- src/axolotl/utils/trainer.py +2 -2
- src/axolotl/utils/validation.py +6 -2
README.md
CHANGED
@@ -176,7 +176,7 @@ tokenizer_type: AutoTokenizer
|
|
176 |
trust_remote_code:
|
177 |
|
178 |
# whether you are training a 4-bit GPTQ quantized model
|
179 |
-
|
180 |
gptq_groupsize: 128 # group size
|
181 |
gptq_model_v1: false # v1 or v2
|
182 |
|
|
|
176 |
trust_remote_code:
|
177 |
|
178 |
# whether you are training a 4-bit GPTQ quantized model
|
179 |
+
gptq: true
|
180 |
gptq_groupsize: 128 # group size
|
181 |
gptq_model_v1: false # v1 or v2
|
182 |
|
configs/quickstart.yml
CHANGED
@@ -40,6 +40,6 @@ early_stopping_patience: 3
|
|
40 |
resume_from_checkpoint:
|
41 |
auto_resume_from_checkpoints: true
|
42 |
local_rank:
|
43 |
-
|
44 |
xformers_attention: true
|
45 |
flash_attention:
|
|
|
40 |
resume_from_checkpoint:
|
41 |
auto_resume_from_checkpoints: true
|
42 |
local_rank:
|
43 |
+
gptq: true
|
44 |
xformers_attention: true
|
45 |
flash_attention:
|
examples/4bit-lora-7b/config.yml
CHANGED
@@ -4,7 +4,7 @@ model_type: LlamaForCausalLM
|
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
trust_remote_code:
|
6 |
load_in_8bit: true
|
7 |
-
|
8 |
datasets:
|
9 |
- path: vicgalle/alpaca-gpt4
|
10 |
type: alpaca
|
|
|
4 |
tokenizer_type: LlamaTokenizer
|
5 |
trust_remote_code:
|
6 |
load_in_8bit: true
|
7 |
+
gptq: true
|
8 |
datasets:
|
9 |
- path: vicgalle/alpaca-gpt4
|
10 |
type: alpaca
|
src/axolotl/utils/models.py
CHANGED
@@ -73,7 +73,7 @@ def load_model(
|
|
73 |
else:
|
74 |
torch_dtype = torch.float32
|
75 |
try:
|
76 |
-
if cfg.
|
77 |
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
|
78 |
replace_peft_model_with_int4_lora_model,
|
79 |
)
|
@@ -95,7 +95,7 @@ def load_model(
|
|
95 |
bnb_4bit_quant_type="nf4",
|
96 |
)
|
97 |
try:
|
98 |
-
if cfg.
|
99 |
from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
|
100 |
from huggingface_hub import snapshot_download
|
101 |
|
@@ -248,7 +248,7 @@ def load_model(
|
|
248 |
|
249 |
if (
|
250 |
((cfg.adapter == "lora" and load_in_8bit) or cfg.adapter == "qlora")
|
251 |
-
and not cfg.
|
252 |
and (load_in_8bit or cfg.load_in_4bit)
|
253 |
):
|
254 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
@@ -259,7 +259,7 @@ def load_model(
|
|
259 |
if cfg.ddp and not load_in_8bit:
|
260 |
model.to(f"cuda:{cfg.local_rank}")
|
261 |
|
262 |
-
if cfg.
|
263 |
# Scales to half
|
264 |
logging.info("Fitting 4bit scales and zeros to half")
|
265 |
for n, m in model.named_modules():
|
@@ -274,7 +274,7 @@ def load_model(
|
|
274 |
if (
|
275 |
torch.cuda.device_count() > 1
|
276 |
and int(os.getenv("WORLD_SIZE", "1")) > 1
|
277 |
-
and cfg.
|
278 |
):
|
279 |
# llama is PROBABLY model parallelizable, but the default isn't that it is
|
280 |
# so let's only set it for the 4bit, see
|
|
|
73 |
else:
|
74 |
torch_dtype = torch.float32
|
75 |
try:
|
76 |
+
if cfg.gptq:
|
77 |
from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
|
78 |
replace_peft_model_with_int4_lora_model,
|
79 |
)
|
|
|
95 |
bnb_4bit_quant_type="nf4",
|
96 |
)
|
97 |
try:
|
98 |
+
if cfg.gptq and is_llama_derived_model:
|
99 |
from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
|
100 |
from huggingface_hub import snapshot_download
|
101 |
|
|
|
248 |
|
249 |
if (
|
250 |
((cfg.adapter == "lora" and load_in_8bit) or cfg.adapter == "qlora")
|
251 |
+
and not cfg.gptq
|
252 |
and (load_in_8bit or cfg.load_in_4bit)
|
253 |
):
|
254 |
logging.info("converting PEFT model w/ prepare_model_for_int8_training")
|
|
|
259 |
if cfg.ddp and not load_in_8bit:
|
260 |
model.to(f"cuda:{cfg.local_rank}")
|
261 |
|
262 |
+
if cfg.gptq:
|
263 |
# Scales to half
|
264 |
logging.info("Fitting 4bit scales and zeros to half")
|
265 |
for n, m in model.named_modules():
|
|
|
274 |
if (
|
275 |
torch.cuda.device_count() > 1
|
276 |
and int(os.getenv("WORLD_SIZE", "1")) > 1
|
277 |
+
and cfg.gptq
|
278 |
):
|
279 |
# llama is PROBABLY model parallelizable, but the default isn't that it is
|
280 |
# so let's only set it for the 4bit, see
|
src/axolotl/utils/trainer.py
CHANGED
@@ -63,7 +63,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
63 |
training_arguments_kwargs["warmup_steps"] = warmup_steps
|
64 |
training_arguments_kwargs["logging_steps"] = logging_steps
|
65 |
if cfg.gradient_checkpointing is not None:
|
66 |
-
if cfg.
|
67 |
from alpaca_lora_4bit.gradient_checkpointing import (
|
68 |
apply_gradient_checkpointing,
|
69 |
)
|
@@ -138,7 +138,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
|
|
138 |
importlib.import_module("torchdistx")
|
139 |
if (
|
140 |
cfg.optimizer == "adamw_bnb_8bit"
|
141 |
-
and not cfg.
|
142 |
and not "deepspeed" in training_arguments_kwargs
|
143 |
and not cfg.fsdp
|
144 |
):
|
|
|
63 |
training_arguments_kwargs["warmup_steps"] = warmup_steps
|
64 |
training_arguments_kwargs["logging_steps"] = logging_steps
|
65 |
if cfg.gradient_checkpointing is not None:
|
66 |
+
if cfg.gptq:
|
67 |
from alpaca_lora_4bit.gradient_checkpointing import (
|
68 |
apply_gradient_checkpointing,
|
69 |
)
|
|
|
138 |
importlib.import_module("torchdistx")
|
139 |
if (
|
140 |
cfg.optimizer == "adamw_bnb_8bit"
|
141 |
+
and not cfg.gptq
|
142 |
and not "deepspeed" in training_arguments_kwargs
|
143 |
and not cfg.fsdp
|
144 |
):
|
src/axolotl/utils/validation.py
CHANGED
@@ -2,16 +2,20 @@ import logging
|
|
2 |
|
3 |
|
4 |
def validate_config(cfg):
|
|
|
|
|
|
|
5 |
if cfg.adapter == "qlora":
|
6 |
if cfg.merge_lora:
|
7 |
# can't merge qlora if loaded in 8bit or 4bit
|
8 |
assert cfg.load_in_8bit is False
|
9 |
-
assert cfg.
|
10 |
assert cfg.load_in_4bit is False
|
11 |
else:
|
12 |
assert cfg.load_in_8bit is False
|
13 |
-
assert cfg.
|
14 |
assert cfg.load_in_4bit is True
|
|
|
15 |
if not cfg.load_in_8bit and cfg.adapter == "lora":
|
16 |
logging.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
|
17 |
|
|
|
2 |
|
3 |
|
4 |
def validate_config(cfg):
|
5 |
+
if cfg.load_4bit:
|
6 |
+
raise ValueError("cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq")
|
7 |
+
|
8 |
if cfg.adapter == "qlora":
|
9 |
if cfg.merge_lora:
|
10 |
# can't merge qlora if loaded in 8bit or 4bit
|
11 |
assert cfg.load_in_8bit is False
|
12 |
+
assert cfg.gptq is False
|
13 |
assert cfg.load_in_4bit is False
|
14 |
else:
|
15 |
assert cfg.load_in_8bit is False
|
16 |
+
assert cfg.gptq is False
|
17 |
assert cfg.load_in_4bit is True
|
18 |
+
|
19 |
if not cfg.load_in_8bit and cfg.adapter == "lora":
|
20 |
logging.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
|
21 |
|