Louis Nguyen

+1

I got same problem of torchao version while installing the torchtune version 0.2.1 ```bash ERROR: Could not find a version that satisfies the requirement torchao==0.3.1 (from torchtune) (from versions: 0.0.1,...

Thanks for yours advice, I'll try to test the token/s

hi @RdoubleA, @felipemello1 I made the test of token/s/gpu: - 1 GPU (RTX 6000ADA) with `tune run lora_finetune_single_device --config phi3/mini_lora_single_device`: ~4100 token/s - 4 GPU (4 x 6000ADA) with `tune...

Hello, regarding the configuration, this is the config file that I used: ```yaml # Fine-tuning arguments epochs: 1 max_steps_per_epoch: null batch_size: 2 gradient_accumulation_steps: 8 compile: False enable_activation_checkpointing: False enable_activation_offloading: False...

For the dataset configuration: ```yaml # Dataset dataset: _component_: torchtune.datasets.text_completion_dataset source: parquet split: train data_dir: ds/sample batch_size: 2 packed: True ``` The error tracking log ```log ╭─────────────────────────────── Traceback (most recent...

this is my full config file. I used for both single_device & distributed ```yaml model: _component_: torchtune.models.phi3.lora_phi3_mini lora_attn_modules: ['q_proj', 'v_proj', 'k_proj', 'o_proj'] apply_lora_to_mlp: False apply_lora_to_output: False lora_rank: 32 lora_alpha: 64...