---
library_name: transformers
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-14B
tags:
- generated_from_trainer
datasets:
- winglian/gpumode-py2triton-reasoning
model-index:
- name: outputs/out-fft
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.8.0`
```yaml
base_model: Qwen/Qwen2.5-Coder-14B
strict: false       

plugins:             
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.liger.LigerPlugin
                                                               
cut_cross_entropy: true
  # liger_rope: true
liger_rms_norm: true
liger_layer_norm: true

# gemma3 doesn't seem to play nice with ddp
ddp_find_unused_parameters: true
                                                               
chat_template: qwen_25
datasets:           
  - path: winglian/gpumode-py2triton-reasoning
    type: chat_template
                                                               
dataset_prepared_path: last_run_prepared
val_set_size: 0.0
output_dir: ./outputs/out-fft
save_safetensors: true
save_only_model: true

sequence_len: 32768
sample_packing: true 
pad_to_sequence_len: true
                                                               
sequence_parallel_degree: 1

# unfrozen_parameters:
#   - language_model.model

wandb_project: qwen25-kernel-llm
wandb_entity: axolotl-ai
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 3
num_epochs: 3
optimizer: adamw_8bit
lr_scheduler: rex
learning_rate: 3.0e-6
lr_groups:
  - name: embeddings
    lr: 3.0e-5
    modules:
      - lm_head
      - embed_tokens

train_on_inputs: false
group_by_length: false
bf16: true
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
logging_steps: 1
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 1
saves_per_epoch: 1
  # deepspeed: deepspeed_configs/zero1.json
weight_decay: 0.0
deepspeed: deepspeed_configs/zero1.json
tokens:
 - <think>
 - </think>
special_tokens:
  eos_token: <|im_end|>
fix_untrained_tokens:
  - 151665
  - 151666


```

</details><br>

# outputs/out-fft

This model is a fine-tuned version of [Qwen/Qwen2.5-Coder-14B](https://huggingface.co/Qwen/Qwen2.5-Coder-14B) on the winglian/gpumode-py2triton-reasoning dataset.

## Model description

More information needed

## Intended uses & limitations

More information needed

## Training and evaluation data

More information needed

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 3e-06
- train_batch_size: 3
- eval_batch_size: 3
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- total_train_batch_size: 24
- total_eval_batch_size: 24
- optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 27
- num_epochs: 3.0

### Training results



### Framework versions

- Transformers 4.51.3
- Pytorch 2.6.0+cu124
- Datasets 3.5.0
- Tokenizers 0.21.1