Upload training_config.yml with huggingface_hub
Browse files- training_config.yml +82 -0
training_config.yml
ADDED
|
@@ -0,0 +1,82 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
model:
|
| 2 |
+
_component_: models.lora_mmllama3_8b
|
| 3 |
+
lora_attn_modules:
|
| 4 |
+
- q_proj
|
| 5 |
+
- v_proj
|
| 6 |
+
apply_lora_to_mlp: false
|
| 7 |
+
apply_lora_to_output: false
|
| 8 |
+
lora_rank: 32
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
perception_tokens: 2
|
| 11 |
+
use_clip: false
|
| 12 |
+
tokenizer:
|
| 13 |
+
_component_: models.a2a_tokenizer
|
| 14 |
+
path: models/tokenizer.model
|
| 15 |
+
checkpointer:
|
| 16 |
+
_component_: torchtune.utils.FullModelMetaCheckpointer
|
| 17 |
+
checkpoint_dir:
|
| 18 |
+
checkpoint_files:
|
| 19 |
+
-
|
| 20 |
+
adapter_checkpoint: null
|
| 21 |
+
recipe_checkpoint: null
|
| 22 |
+
output_dir: output_checkpoints/experiment_1
|
| 23 |
+
model_type: LLAMA3
|
| 24 |
+
resume_from_checkpoint: false
|
| 25 |
+
interim_checkpoint_steps: 15000
|
| 26 |
+
interim_gen_steps: null
|
| 27 |
+
max_new_tokens: 77
|
| 28 |
+
temperature: 0.6
|
| 29 |
+
top_k: 231
|
| 30 |
+
dataset:
|
| 31 |
+
_component_: ds.EvenBatcher
|
| 32 |
+
buffer_size: 72
|
| 33 |
+
dataset:
|
| 34 |
+
_component_: ds.RoundRobinDataset
|
| 35 |
+
datasets:
|
| 36 |
+
- _component_: ds.OmegaVideoCaptionDataset
|
| 37 |
+
length: 500000
|
| 38 |
+
- _component_: ds.LlavaInstructDataset
|
| 39 |
+
dataset_path: ds/coco_llava_instruct/output.parquet
|
| 40 |
+
train_on_input: false
|
| 41 |
+
- _component_: ds.LlavaInstructDataset
|
| 42 |
+
dataset_path: ds/vision_flan/output.parquet
|
| 43 |
+
train_on_input: false
|
| 44 |
+
- _component_: ds.CaptionInstructDataset
|
| 45 |
+
dataset_path: ds/sam_llava/output.parquet
|
| 46 |
+
train_on_input: false
|
| 47 |
+
seed: null
|
| 48 |
+
shuffle: true
|
| 49 |
+
batch_size: 6
|
| 50 |
+
optimizer:
|
| 51 |
+
_component_: torch.optim.AdamW
|
| 52 |
+
weight_decay: 0.99
|
| 53 |
+
lr: 1e-11
|
| 54 |
+
lr_scheduler:
|
| 55 |
+
_component_: torchtune.modules.get_cosine_schedule_with_warmup
|
| 56 |
+
num_warmup_steps: 0
|
| 57 |
+
loss:
|
| 58 |
+
_component_: torch.nn.CrossEntropyLoss
|
| 59 |
+
epochs: 1
|
| 60 |
+
max_steps_per_epoch: null
|
| 61 |
+
gradient_accumulation_steps: 260
|
| 62 |
+
compile: false
|
| 63 |
+
output_dir: /tmp/lora_finetune_output
|
| 64 |
+
metric_logger:
|
| 65 |
+
_component_: torchtune.utils.metric_logging.DiskLogger
|
| 66 |
+
log_dir: ${output_dir}
|
| 67 |
+
log_every_n_steps: null
|
| 68 |
+
device: cuda
|
| 69 |
+
dtype: bf16
|
| 70 |
+
enable_activation_checkpointing: false
|
| 71 |
+
profiler:
|
| 72 |
+
_component_: torchtune.utils.profiler
|
| 73 |
+
enabled: false
|
| 74 |
+
inference:
|
| 75 |
+
prompt_template: 'Video:
|
| 76 |
+
{video}
|
| 77 |
+
Caption the previous video.'
|
| 78 |
+
max_new_tokens: 231
|
| 79 |
+
temperature: 0.8
|
| 80 |
+
top_k: 231
|
| 81 |
+
quantizer: null
|
| 82 |
+
gradient-accumulation-steps: 32
|