search

NeuroBench / neurobench

Benchmark harness and baseline results for the NeuroBench algorithm track.
https://neurobench.readthedocs.io
Apache License 2.0
52 stars 12 forks source link

Footprint metric for LSTM / recurrent networks #89

Closed jasonlyik closed 10 months ago

jasonlyik commented 11 months ago

PyTorch recurrent layers (LSTM/GRU) have two bias vectors for CuDNN compatibility.

In footprint calculation the second bias vector b_hh should be skipped or suppressed

jasonlyik commented 11 months ago

This issue potentially affects the LSTMCell used in the MG task

jasonlyik commented 10 months ago

Will not fix this issue since it is characteristic of the particular model, which is what the footprint metric captures.