Models got fail accuracy on BMG but passed on PVC #1305
Description
🐛 Describe the bug
python benchmarks/dynamo/timm_models.py --accuracy --float16 -d xpu -n10 --training --only botnet26t_256 --backend=inductor
| Category | Model | Accuracy |
|---|---|---|
| timm_models_float16_training | botnet26t_256 | fail_accuracy |
| torchbench_float32_inference | detectron2_maskrcnn_r_50_fpn | fail_accuracy |
| timm_models_amp_fp16_training | fbnetv3_b | fail_accuracy |
| timm_models_bfloat16_training | fbnetv3_b | fail_accuracy |
| timm_models_amp_bf16_training | gernet_l | fail_accuracy |
| torchbench_float16_training | hf_Longformer | fail_accuracy |
| torchbench_bfloat16_training | hf_Longformer | fail_accuracy |
Versions
Environments:
| Name | Version |
|---|---|
| Device | BMG |
| OS | Ubuntu 24.10 |
| Driver | 24.45.31740 |
| Torch | v2.6.0-rc5 |
| Torch-xpu-ops | release/2.6 |
| Triton | 3.2.0 |
| Transformers | 243e186efbf7fb93328dd6b34927a4e8c8f24395 |
| Torchvision | 0.21.0+xpu |
| Torchaudio | 2.6.0+xpu |
| Torchbench | 766a5e3a189384659fd35a68c3b17b88c761aaac |
| Timms | ac3470188b914c5d7a5058a7e28b9eb685a62427 |
| Bundle | 2025.0.4.20241205 |