Create qwq.yml
Browse files
qwq.yml
ADDED
@@ -0,0 +1,87 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
base_model: Qwen/QwQ-32B
|
2 |
+
model_type: AutoModelForCausalLM
|
3 |
+
tokenizer_type: AutoTokenizer
|
4 |
+
|
5 |
+
plugins:
|
6 |
+
- axolotl.integrations.liger.LigerPlugin
|
7 |
+
liger_rope: true
|
8 |
+
liger_rms_norm: true
|
9 |
+
liger_swiglu: true
|
10 |
+
liger_fused_linear_cross_entropy: true
|
11 |
+
|
12 |
+
load_in_8bit: false
|
13 |
+
load_in_4bit: false
|
14 |
+
strict: false
|
15 |
+
|
16 |
+
datasets:
|
17 |
+
- path: Mielikki/Erebus-87k
|
18 |
+
type: completion
|
19 |
+
field: body
|
20 |
+
- path: NewEden/Orion-Completion-Asstr-Stories-16K
|
21 |
+
type: completion
|
22 |
+
field: content
|
23 |
+
- path: NewEden/Orion-Completion-LIT
|
24 |
+
type: completion
|
25 |
+
field: text
|
26 |
+
|
27 |
+
shuffle_merged_datasets: true
|
28 |
+
dataset_prepared_path: prepared_data
|
29 |
+
output_dir: ./qvq-cum
|
30 |
+
|
31 |
+
sequence_len: 16384
|
32 |
+
sample_packing: true
|
33 |
+
pad_to_sequence_len: true
|
34 |
+
|
35 |
+
adapter: lora
|
36 |
+
lora_model_dir:
|
37 |
+
lora_r: 128
|
38 |
+
lora_alpha: 16
|
39 |
+
lora_dropout: 0.05
|
40 |
+
lora_target_modules:
|
41 |
+
- gate_proj
|
42 |
+
- down_proj
|
43 |
+
- up_proj
|
44 |
+
- q_proj
|
45 |
+
- v_proj
|
46 |
+
- k_proj
|
47 |
+
- o_proj
|
48 |
+
|
49 |
+
lora_modules_to_save:
|
50 |
+
- embed_tokens
|
51 |
+
- lm_head
|
52 |
+
|
53 |
+
wandb_project: qwq
|
54 |
+
wandb_entity:
|
55 |
+
wandb_watch:
|
56 |
+
wandb_name: Pretrain-pt1
|
57 |
+
wandb_log_model:
|
58 |
+
|
59 |
+
gradient_accumulation_steps: 2
|
60 |
+
micro_batch_size: 2
|
61 |
+
num_epochs: 1
|
62 |
+
optimizer: paged_ademamix_8bit
|
63 |
+
lr_scheduler: cosine
|
64 |
+
learning_rate: 1e-5
|
65 |
+
max_grad_norm: 0.001
|
66 |
+
|
67 |
+
train_on_inputs: false
|
68 |
+
group_by_length: false
|
69 |
+
bf16: auto
|
70 |
+
fp16:
|
71 |
+
tf32: false
|
72 |
+
|
73 |
+
gradient_checkpointing: unsloth
|
74 |
+
early_stopping_patience:
|
75 |
+
resume_from_checkpoint:
|
76 |
+
local_rank:
|
77 |
+
logging_steps: 1
|
78 |
+
xformers_attention:
|
79 |
+
flash_attention: true
|
80 |
+
|
81 |
+
warmup_steps: 40
|
82 |
+
saves_per_epoch: 2
|
83 |
+
debug:
|
84 |
+
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json
|
85 |
+
weight_decay: 0.01
|
86 |
+
fsdp:
|
87 |
+
fsdp_config:
|