Driss Guessous

@pytorchbot merge -l

@prav2019 So what exactly solved your OOM issue, was it setting the max-requests?

Notes: - Did not implement context manager versions of use_fused and use_math - Not tested (YAY!). Not totally true since we have test_transformers that exercised fused_sdp function. However we do...

This branch has been corrupted. Lets try this again: #85880

fixes: #85064

@cpuhrsch the one user of _scaled_dot_product_attention is multiheaded_attention_forward and the change broke with: ``` File "/opt/conda/lib/python3.7/site-packages/torch/nn/functional.py", line 5170, in multi_head_attention_forward attn_output, attn_output_weights = _scaled_dot_product_attention(q, k, v, attn_mask, dropout_p) RuntimeError: _scaled_dot_product_attention:...

@cpuhrsch yeah onnx failure is related. Need to figure where the docs are on that one. the other test failure: ``` AssertionError: The supported dtypes for nn.functional._scaled_dot_product_attention on device type...

@pytorchbot merge -l

@pytorchbot merge -l

@pytorchbot merge -f "Onnx failure is unrelated to my change"