phoenix-oss/llama-stack
1
0
Fork 0
forked from phoenix-oss/llama-stack-mirror
Code Pull requests Releases Packages 2 Activity Actions

[Post training] make validation steps configurable (#715)

Browse source 
## what does this PR do? 
The current code hardcode the validation steps to run (forgot to change
it after testing). in this PR, we make it configurable by training
config

## test 
On client side, issue a post training request with 20 validation steps,
server side logging shows that it runs 20 validation steps successfully
<img width="1128" alt="Screenshot 2025-01-02 at 8 21 06 PM"
src="https://github.com/user-attachments/assets/7a757516-c6ba-41d4-85c5-361a80ecf46e"
/>
This commit is contained in:
Botao Chen 2025-01-03 08:43:24 -08:00 • committed by GitHub
parent f450a0fd32
commit 4320b0ebb2
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 3 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

1
llama_stack/apis/post_training/post_training.py
View file

@ -58,6 +58,7 @@ class TrainingConfig(BaseModel):
n_epochs: int n_epochs: int
max_steps_per_epoch: int max_steps_per_epoch: int
gradient_accumulation_steps: int gradient_accumulation_steps: int
max_validation_steps: int
data_config: DataConfig data_config: DataConfig
optimizer_config: OptimizerConfig optimizer_config: OptimizerConfig
efficiency_config: Optional[EfficiencyConfig] = None efficiency_config: Optional[EfficiencyConfig] = None

3
llama_stack/providers/inline/post_training/torchtune/recipes/lora_finetuning_single_device.py
View file

@ -137,6 +137,7 @@ class LoraFinetuningSingleDevice:
self.global_step = 0 self.global_step = 0
self._gradient_accumulation_steps = training_config.gradient_accumulation_steps self._gradient_accumulation_steps = training_config.gradient_accumulation_steps
self.max_validation_steps = training_config.max_validation_steps
self._clip_grad_norm = 1.0 self._clip_grad_norm = 1.0
self._enable_activation_checkpointing = ( self._enable_activation_checkpointing = (
@ -583,7 +584,7 @@ class LoraFinetuningSingleDevice:
log.info("Starting validation...") log.info("Starting validation...")
pbar = tqdm(total=len(self._validation_dataloader)) pbar = tqdm(total=len(self._validation_dataloader))
for idx, batch in enumerate(self._validation_dataloader): for idx, batch in enumerate(self._validation_dataloader):
if idx == 10: if idx == self.max_validation_steps:
break break
torchtune_utils.batch_to_device(batch, self._device) torchtune_utils.batch_to_device(batch, self._device)