docs: add fine-tuned model section
Browse files
README.md
CHANGED
|
@@ -27,6 +27,12 @@ ROCmPort AI orchestrates a migration loop:
|
|
| 27 |
|
| 28 |
Complex kernels can fail conversion due to architecture assumptions, undefined behavior, inline PTX, or handcrafted memory logic. The value is reduced migration time and faster debug loops.
|
| 29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 30 |
## Dataset
|
| 31 |
|
| 32 |
Training data: [tazwarrrr/cuda-to-rocm-wavefront-bugs](https://huggingface.co/datasets/tazwarrrr/cuda-to-rocm-wavefront-bugs)
|
|
|
|
| 27 |
|
| 28 |
Complex kernels can fail conversion due to architecture assumptions, undefined behavior, inline PTX, or handcrafted memory logic. The value is reduced migration time and faster debug loops.
|
| 29 |
|
| 30 |
+
## Fine-tuned Model
|
| 31 |
+
|
| 32 |
+
[tazwarrrr/rocmport-qwen-wavefront-finetuned](https://huggingface.co/tazwarrrr/rocmport-qwen-wavefront-finetuned)
|
| 33 |
+
Qwen2.5-Coder-7B-Instruct fine-tuned on 153 CUDA→ROCm wavefront-64 bug examples.
|
| 34 |
+
Trained on AMD Instinct MI300X (gfx942) with ROCm 6.2 in 79 seconds.
|
| 35 |
+
|
| 36 |
## Dataset
|
| 37 |
|
| 38 |
Training data: [tazwarrrr/cuda-to-rocm-wavefront-bugs](https://huggingface.co/datasets/tazwarrrr/cuda-to-rocm-wavefront-bugs)
|