generative-recommenders

generative-recommenders copied to clipboard

Fix issues in Hammer compilation

2

Differential Revision: D60771307

mlevesquedion

Triton is running too slow?

1

Compared to the same structure(the qkv attention) I implemented with TensorFlow, triton runs 10 to 20 times slower. With the help of nsight system, I found that cudaMemcpySync takes off...

bzxc

adding bias bwd makes fused attention bwd 30x slower, any suggestions?

1

I have written a bwd triton code , but I found that after adding the bias bwd , the speed is 30 times slower. The following is the time_weight bwd...

zzq96

Use input_precision="tf32x3" when allow_tf32=true

2

Differential Revision: D58453633

htyu

How to combine autoregressive model and classification multi-task head in ranking model?

1

If I understand correctly, autoregressive model has a loss, and also multi-task dense layers followed autoregressive model has a weighted loss. How to combine them? And in ranking model, how...

MoFHeka

How to deal with unknown tokens in practice?

1

Hello, it's a great work! I have some problem about the unknown token(newly created items). Because of the long sequence, only user side category features can be merged into the...

isofun

Evaluation on public dataset

1

Hi, great work! I'm trying to reproduce the results on public datasets. However, I only found the training codes, where the model was evaluated on the eval set (or you...

Blank-z0

Integration with TorchRec

1

Are there any plans to integrate the embedding_modules or custom samplers back into TorchRec?

sbhavani

confused about "SampledSoftmaxLoss" func

3

Hey, Congratulations for your perfect and creative work. when I read the implementation code here, I am very confused about [SampledSoftmaxLoss](https://github.com/facebookresearch/generative-recommenders/blob/54e5240567041b8f74c735b437404270a5b1cf49/generative_recommenders/modeling/sequential/autoregressive_losses.py#L499). I have some questions for this: 1. why do...

zhhu1996

change addmm grid size to 2D to overcome uint16 limitation

2

Differential Revision: D64049725

jiyuanzFB