Upload 2-PKTDC-llama-30B-gptq-lora-24gb.yml
Browse files
trainer config/2-PKTDC-llama-30B-gptq-lora-24gb.yml
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# accelerate launch ./scripts/finetune.py 2-PKTDC-llama-30B-gptq-lora-24gb.yml
|
2 |
+
#
|
3 |
+
# base model settings (local or huggingface repo)
|
4 |
+
base_model: PocketDoc/llama-30b-gptq-4bit-128g
|
5 |
+
base_model_config: PocketDoc/llama-30b-gptq-4bit-128g
|
6 |
+
model_type: LlamaForCausalLM
|
7 |
+
tokenizer_type: LlamaTokenizer
|
8 |
+
trust_remote_code:
|
9 |
+
|
10 |
+
# wandb configuration
|
11 |
+
wandb_project: llama-30b-gptq-4bit-128g-lora
|
12 |
+
wandb_watch:
|
13 |
+
wandb_run_id:
|
14 |
+
wandb_log_model:
|
15 |
+
|
16 |
+
# where to save the finished model to
|
17 |
+
output_dir: ./llama-30b-gptq-4bit-128g-lora
|
18 |
+
|
19 |
+
# dataset settings (local or huggingface repo)
|
20 |
+
datasets:
|
21 |
+
- path: dansmeth.json
|
22 |
+
type: pygmalion
|
23 |
+
|
24 |
+
dataset_prepared_path: data/last_run_prepared
|
25 |
+
|
26 |
+
# percentage of the dataset to set aside as evaluation.
|
27 |
+
val_set_size: 0.02
|
28 |
+
|
29 |
+
# max token length / prompt
|
30 |
+
sequence_len: 2048
|
31 |
+
|
32 |
+
# max sequence length to concatenate training samples together up to
|
33 |
+
# inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
|
34 |
+
max_packed_sequence_len: 2048
|
35 |
+
|
36 |
+
# quantized model loading settings
|
37 |
+
gptq: true
|
38 |
+
gptq_groupsize: 128 # group size
|
39 |
+
gptq_model_v1: false # v1 or v2
|
40 |
+
strict: false
|
41 |
+
|
42 |
+
# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
|
43 |
+
load_in_8bit: true
|
44 |
+
|
45 |
+
load_in_4bit:
|
46 |
+
|
47 |
+
# Use CUDA bf16
|
48 |
+
bf16: false
|
49 |
+
# Use CUDA fp16
|
50 |
+
fp16: true
|
51 |
+
# Use CUDA tf32
|
52 |
+
tf32: true
|
53 |
+
|
54 |
+
# training hyperparameters
|
55 |
+
gradient_accumulation_steps: 32
|
56 |
+
micro_batch_size: 1
|
57 |
+
eval_batch_size: 1
|
58 |
+
num_epochs: 3
|
59 |
+
warmup_steps: 350
|
60 |
+
learning_rate: 0.00003
|
61 |
+
|
62 |
+
logging_steps: 1
|
63 |
+
eval_steps: 25
|
64 |
+
save_steps: 175
|
65 |
+
|
66 |
+
# stop training after this many evaluation losses have increased in a row
|
67 |
+
# https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
|
68 |
+
early_stopping_patience:
|
69 |
+
# specify a scheduler to use with the optimizer. only one_cycle is supported currently
|
70 |
+
lr_scheduler: linear
|
71 |
+
# specify optimizer
|
72 |
+
optimizer: paged_adamw_8bit
|
73 |
+
# specify weight decay
|
74 |
+
weight_decay: 0.05
|
75 |
+
|
76 |
+
|
77 |
+
# if you already have a lora model trained that you want to load, put that here
|
78 |
+
lora_model_dir:
|
79 |
+
|
80 |
+
# LoRA hyperparameters
|
81 |
+
adapter: lora # blank for full finetune
|
82 |
+
lora_r: 32
|
83 |
+
lora_alpha: 64
|
84 |
+
lora_dropout: 0.05
|
85 |
+
lora_target_linear:
|
86 |
+
lora_target_modules:
|
87 |
+
- q_proj
|
88 |
+
- v_proj
|
89 |
+
# - k_proj
|
90 |
+
# - o_proj
|
91 |
+
# - gate_proj
|
92 |
+
# - down_proj
|
93 |
+
# - up_proj
|
94 |
+
lora_modules_to_save:
|
95 |
+
# - embed_tokens
|
96 |
+
# - lm_head
|
97 |
+
lora_out_dir:
|
98 |
+
lora_fan_in_fan_out: false
|
99 |
+
|
100 |
+
|
101 |
+
# whether to mask out or include the human's prompt from the training labels
|
102 |
+
train_on_inputs: false
|
103 |
+
# don't use this, leads to wonky training (according to someone on the internet)
|
104 |
+
group_by_length: true
|
105 |
+
|
106 |
+
|
107 |
+
# does not work with current implementation of 4-bit LoRA
|
108 |
+
gradient_checkpointing: true
|
109 |
+
|
110 |
+
|
111 |
+
# whether to use xformers attention patch https://github.com/facebookresearch/xformers:
|
112 |
+
xformers_attention: true
|
113 |
+
# whether to use flash attention patch https://github.com/HazyResearch/flash-attention:
|
114 |
+
flash_attention: # require a100 for llama
|
115 |
+
# whether to use scaled-dot-product attention
|
116 |
+
# https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
|
117 |
+
sdp_attention:
|
118 |
+
|
119 |
+
|
120 |
+
# resume from a specific checkpoint dir
|
121 |
+
resume_from_checkpoint:
|
122 |
+
# if resume_from_checkpoint isn't set and you simply want it to start where it left off
|
123 |
+
# be careful with this being turned on between different models
|
124 |
+
auto_resume_from_checkpoints:
|
125 |
+
|
126 |
+
|
127 |
+
# don't mess with this, it's here for accelerate and torchrun
|
128 |
+
local_rank:
|
129 |
+
|
130 |
+
# add or change special tokens
|
131 |
+
special_tokens:
|
132 |
+
# sys_role_token: "<|system|>"
|
133 |
+
# user_role_token: "<|user|>"
|
134 |
+
# model_role_token: "<|model|>"
|
135 |
+
bos_token: "<s>"
|
136 |
+
eos_token: "</s>"
|
137 |
+
unk_token: "<unk>"
|
138 |
+
|
139 |
+
# add extra tokens
|
140 |
+
tokens:
|
141 |
+
|
142 |
+
|
143 |
+
# FSDP
|
144 |
+
fsdp:
|
145 |
+
|
146 |
+
fsdp_config:
|
147 |
+
|
148 |
+
# Deepspeed
|
149 |
+
deepspeed:
|
150 |
+
|
151 |
+
# TODO
|
152 |
+
torchdistx_path:
|
153 |
+
|
154 |
+
# Debug mode
|
155 |
+
debug:
|