|
|
--- |
|
|
license: apache-2.0 |
|
|
base_model: Qwen/Qwen2.5-1.5B-Instruct |
|
|
tags: |
|
|
- triton |
|
|
- kernel |
|
|
- code-generation |
|
|
- fine-tuned |
|
|
datasets: |
|
|
- triton-kernels-6k |
|
|
language: |
|
|
- en |
|
|
pipeline_tag: text-generation |
|
|
--- |
|
|
|
|
|
# Triton Kernel Code Generation Model |
|
|
|
|
|
This model is a fine-tuned version of Qwen/Qwen2.5-1.5B-Instruct specialized for generating Triton GPU kernels. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Base Model**: Qwen/Qwen2.5-1.5B-Instruct |
|
|
- **Fine-tuned on**: 6000 examples of Triton kernel code |
|
|
- **Eval Loss**: 0.20 |
|
|
- **Eval Perplexity**: 1.22 |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from transformers import AutoTokenizer, AutoModelForCausalLM |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("cdreetz/kwen2.5-1.5b") |
|
|
tokenizer = AutoTokenizer.from_pretrained("cdreetz/kwen2.5-1.5b") |
|
|
|
|
|
prompt = "Write a Triton kernel for element-wise addition:" |
|
|
inputs = tokenizer(prompt, return_tensors="pt") |
|
|
outputs = model.generate(**inputs, max_new_tokens=512) |
|
|
response = tokenizer.decode(outputs[0], skip_special_tokens=True) |
|
|
print(response) |
|
|
``` |
|
|
|
|
|
## Training Details |
|
|
|
|
|
- **Epochs**: 2 |
|
|
- **Batch Size**: 2 |
|
|
- **Learning Rate**: 1e-5 |
|
|
- **Dataset Size**: 6000 examples |
|
|
|
|
|
## Performance |
|
|
|
|
|
The model generates syntactically correct Triton kernels with proper: |
|
|
- `@triton.jit` decorators |
|
|
- Kernel function signatures |
|
|
- Launch function implementations |
|
|
- Memory access patterns |
|
|
- Grid configurations |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Specialized for Triton kernel generation only |
|
|
- May require prompt engineering for optimal results |
|
|
- Generated kernels should be tested before production use |
|
|
|
