generative-recommenders

generative-recommenders copied to clipboard

when i see: B = all_timestamps.size(0) N = self._max_seq_len t = F.pad(self._pos_w[:2 * N - 1], [0, N]).repeat(N) t = t[..., :-N].reshape(1, N, 3 * N - 2) r =...

ShuaiXuan

updates README.md

2

Differential Revision: D56863393

jiaqizhai

Add position index input to support different timestamp chunks in HSTU attwntion kernel for context parallelism

1

Reviewed By: hanli0612 Differential Revision: D57754655

yoyoyocmu

Support no-mask in the HSTU attention kernel in both forward and backward

3

Differential Revision: D57292670

yoyoyocmu

Separate normalization factor from max_seq_len

2

Reviewed By: hanli0612 Differential Revision: D57320475

yoyoyocmu

https://github.com/facebookresearch/generative-recommenders/blob/main/data/dataset.py#L142 你好，我看代码这里的每个item代表一个token，但是我看论文里说是item0，action0，item1，action1这种形式来作为输入的格式的，不知道我是哪里理解有问题？

llmmodel123

NOTE: This is an experiment, and a draft. Do not review. The following change requires a private patchset that is not yet available outside of https://github.com/plotfi/triton/pull/4 This patch adds usage...

plotfi

[WIP] TMA Version of HSTU (Autotuned)

4

Based on #57, this version uses the autotuned to toggle use of TMA.

plotfi

TMA version of hstu

1

The TMA version is added as a separate file, we have seen it improves perf in some cases and regresses perf in other cases. We are also working on improving...

manman-ren

What computing power configuration is required for trainging the GRs? With the experiment setup in the article, i.e. 256 H100, how long should the model be trained?

fyx18