cjg91 / trans-fat

An FPGA Accelerator for Transformer Inference
72 stars 12 forks source link

Stage3 Fused #9

Closed danielstumpp closed 2 years ago

danielstumpp commented 2 years ago

Stage3 fully fused and passes tests w.r.t ground truth.

Verify that the dimensions are correct. Note that the scaling value had to be made very small to prevent gelu from always returning zero.

This includes some stage 4 dev that I used to start this branch. We'll still need additional verification of that as I mentioned on Slack.