Repository navigation
[OOB] use sdpa fusion in eager models #3778
Description
change and upstream model script to enable sdpa in the following models in inference. Torch compile has fused them by pattern matcher.
| # | Model |
|---|---|
| 1 | volo_d1_224_bs2048 |
| 2 | tnt_s_patch16_224_bs2048 |
| 3 | LayoutLMForSequenceClassification_bs256 |
| 4 | hf_T5_large_bs32 |
| 5 | hf_T5_bs16 |
| 6 | T5ForConditionalGeneration_bs32 |
| 7 | T5Small_bs32 |
| 8 | hf_T5_generate_bs16 |
| 9 | visformer_small_bs2048 |
| 10 | hf_T5_base_bs1 |
| 11 | MT5ForConditionalGeneration_bs64 |
Summary: 11 model(s) could use F.scaled_dot_product_attention in eager code.
Activity
daisyden commentedon Nov 11, 2026
label-issue: #3778
Root cause: enablement task, not a failure — 11 eager models could use
F.scaled_dot_product_attention(already fused by torch.compile pattern matcher); requires upstreaming model-script changes. Owner: model scripts / torch-xpu-ops OOB.typeFeaturetesttest: oobtest: oob; [OOB] out-of-box eager model suite.modulemodule: sdpapriorityLowinferenceApplied to this issue. Generated by a local
label-issuerun.