Views
No views yet
0.10.0.dev01# === Model Configuration ===
2base_model: Qwen/Qwen3-30B-A3B-Base # e.g. "mistralai/Mistral-Small-24B-Instruct-2501"
3load_in_8bit: false
4load_in_4bit: false
5
6# === Training Setup ===
7num_epochs: 2
8micro_batch_size: 2
9gradient_accumulation_steps: 1
10sequence_len: 8192
11sample_packing: true
12pad_to_sequence_len: true
13
14# === Hyperparameter Configuration ===
15optimizer: adamw_torch_fused
16# Apollo-mini configuration:
17#optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"
18# Regular Apollo configuration:
19# optim_args:
20#optim_target_modules: all_linear
21learning_rate: 1e-5
22lr_scheduler: rex
23weight_decay: 0.01
24warmup_ratio: 0.05
25cosine_min_lr_ratio: 0.1
26
27# === LoRA Configuration ===
28adapter: lora
29lora_r: 128
30lora_alpha: 16
31lora_dropout: 0.35
32lora_target_modules:
33lora_target_linear: true
34peft_use_rslora: true
35
36lora_mlp_kernel: true
37lora_qkv_kernel: true
38lora_o_kernel: true
39
40# === Data Configuration ===
41datasets:
42 - path: allura-forge/fuckedup-inkmix
43 type: chat_template
44 split: train
45 field_messages: conversations
46 message_field_role: from
47 message_field_content: value
48
49dataset_prepared_path: last_run_prepared
50chat_template: chatml
51
52# === Plugins ===
53plugins:
54 - axolotl.integrations.liger.LigerPlugin
55 - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
56
57# === Hardware Optimization ===
58#51;33;32Mgradient_checkpointing: offload
59#gradient_checkpointing_kwargs:
60# use_reentrant: false
61liger_rope: true
62liger_rms_norm: true
63liger_glu_activation: true
64cut_cross_entropy: true
65
66# === Wandb Tracking ===
67wandb_project: q3-30b-fuckedup-inkmix
68
69# === Checkpointing ===
70saves_per_epoch: 2
71save_total_limit: 3
72
73# === Advanced Settings ===
74output_dir: ./ckpts
75bf16: auto
76flash_attention: true
77train_on_inputs: false
78group_by_length: false
79logging_steps: 1
80trust_remote_code: true
81fsdp:
82 - full_shard
83 - auto_wrap
84fsdp_config:
85 fsdp_activation_checkpointing: true
86 fsdp_limit_all_gathers: true
87 fsdp_sync_module_states: true
88 fsdp_use_orig_params: false
89 fsdp_cpu_ram_efficient_loading: true
90 fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
91 fsdp_transformer_layer_cls_to_wrap: Qwen3MoeDecoderLayer
92 fsdp_state_dict_type: SHARDED_STATE_DICT
93 fsdp_reshard_after_forward: true
94 fsdp_version: 2
95