support for replit lm
Browse files
examples/replit-3b/config-lora.yml
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: replit/replit-code-v1-3b
|
2 |
+
base_model_config: replit/replit-code-v1-3b
|
3 |
+
trust_remote_code: true
|
4 |
+
load_in_8bit: false
|
5 |
+
datasets:
|
6 |
+
- path: vicgalle/alpaca-gpt4
|
7 |
+
type: alpaca
|
8 |
+
dataset_prepared_path: last_run_prepared
|
9 |
+
val_set_size: 0.05
|
10 |
+
adapter: lora
|
11 |
+
lora_model_dir:
|
12 |
+
sequence_len: 2048
|
13 |
+
max_packed_sequence_len:
|
14 |
+
lora_r: 8
|
15 |
+
lora_alpha: 16
|
16 |
+
lora_dropout: 0.05
|
17 |
+
lora_target_modules:
|
18 |
+
- Wqkv
|
19 |
+
- mlp_up
|
20 |
+
- mlp_down
|
21 |
+
lora_fan_in_fan_out:
|
22 |
+
wandb_project: lora-replit
|
23 |
+
wandb_watch:
|
24 |
+
wandb_run_id:
|
25 |
+
wandb_log_model:
|
26 |
+
output_dir: ./lora-replit
|
27 |
+
batch_size: 8
|
28 |
+
micro_batch_size: 1
|
29 |
+
num_epochs: 3
|
30 |
+
optimizer:
|
31 |
+
torchdistx_path:
|
32 |
+
lr_scheduler:
|
33 |
+
learning_rate: 0.00001
|
34 |
+
train_on_inputs: false
|
35 |
+
group_by_length: false
|
36 |
+
bf16: true
|
37 |
+
tf32: true
|
38 |
+
gradient_checkpointing:
|
39 |
+
early_stopping_patience:
|
40 |
+
resume_from_checkpoint:
|
41 |
+
local_rank:
|
42 |
+
logging_steps: 1
|
43 |
+
xformers_attention:
|
44 |
+
flash_attention:
|
45 |
+
gptq_groupsize:
|
46 |
+
gptq_model_v1:
|
47 |
+
warmup_steps: 20
|
48 |
+
eval_steps: 50
|
49 |
+
save_steps:
|
50 |
+
debug:
|
51 |
+
deepspeed:
|
52 |
+
weight_decay: 0
|
53 |
+
fsdp:
|
54 |
+
fsdp_config:
|
55 |
+
#special_tokens:
|
src/axolotl/utils/models.py
CHANGED
@@ -163,11 +163,20 @@ def load_model(
|
|
163 |
if not tokenizer:
|
164 |
try:
|
165 |
if is_llama_derived_model and "LlamaTokenizer" in globals():
|
166 |
-
tokenizer = LlamaTokenizer.from_pretrained(
|
|
|
|
|
|
|
167 |
else:
|
168 |
-
tokenizer = getattr(transformers, tokenizer_type).from_pretrained(
|
|
|
|
|
|
|
169 |
except:
|
170 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
|
|
|
|
|
|
171 |
|
172 |
logging.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
|
173 |
logging.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
|
|
|
163 |
if not tokenizer:
|
164 |
try:
|
165 |
if is_llama_derived_model and "LlamaTokenizer" in globals():
|
166 |
+
tokenizer = LlamaTokenizer.from_pretrained(
|
167 |
+
model,
|
168 |
+
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
169 |
+
)
|
170 |
else:
|
171 |
+
tokenizer = getattr(transformers, tokenizer_type).from_pretrained(
|
172 |
+
model,
|
173 |
+
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
174 |
+
)
|
175 |
except:
|
176 |
+
tokenizer = AutoTokenizer.from_pretrained(
|
177 |
+
base_model_config,
|
178 |
+
trust_remote_code=True if cfg.trust_remote_code is True else False,
|
179 |
+
)
|
180 |
|
181 |
logging.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
|
182 |
logging.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
|