1b53c49e1a
* diffusion training plugin * cleanup * nits * fixes + improvements * add back in reinit_weights (clobbered?); masking / pretrain fixes * nits * cleanup; tests draft * sample generation, tests fixes * fixes * nits * add inference support; add auto-mask token support * nits * nits * progress * simplify logging * lint * prefix args with diffusion_ * coderabbito * tests fix * nit * nits * cleanup + nits * nits * fix SFT sample gen * fixes * fix * comments * comments * lint * reward model lora fix * cleanup; fix pretraining_dataset case * gradio inference * update cfgs * update cfgs * train, generation parity, cleanup * fix * simplify * test * test fix
57 lines
1.0 KiB
YAML
57 lines
1.0 KiB
YAML
base_model: meta-llama/Llama-3.2-1B
|
|
# Automatically upload checkpoint and final model to HF
|
|
# hub_model_id: username/custom_model_name
|
|
|
|
pretraining_dataset:
|
|
- path: wikitext
|
|
name: wikitext-103-raw-v1
|
|
type: completion
|
|
field: text
|
|
|
|
plugins:
|
|
- axolotl.integrations.diffusion.DiffusionPlugin
|
|
|
|
diffusion:
|
|
noise_schedule: cosine
|
|
min_mask_ratio: 0.15
|
|
max_mask_ratio: 0.85
|
|
num_diffusion_steps: 128
|
|
eps: 5e-4
|
|
importance_weighting: true
|
|
mask_token_id: 128002
|
|
generate_samples: true
|
|
generation_interval: 250
|
|
|
|
output_dir: ./outputs/model-out
|
|
|
|
sequence_len: 512
|
|
sample_packing: true
|
|
|
|
gradient_accumulation_steps: 8
|
|
micro_batch_size: 4
|
|
max_steps: 10000
|
|
warmup_ratio: 0.1
|
|
|
|
optimizer: adamw_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 3e-4
|
|
sdp_attention: true
|
|
|
|
bf16: auto
|
|
tf32: true
|
|
|
|
logging_steps: 1
|
|
save_strategy: steps
|
|
save_steps: 1000
|
|
|
|
special_tokens:
|
|
pad_token: "<|end_of_text|>"
|
|
|
|
wandb_project:
|
|
wandb_entity:
|
|
wandb_watch:
|
|
wandb_name:
|
|
wandb_log_model:
|
|
|
|
# save_first_step: true # uncomment this to validate checkpoint saving works with your config
|