Ningxin Zheng comments

Results 12 comments of


                                            Ningxin Zheng

assert len(set(num_channels_list)) == 1

The `size` operation on the graph misleads the model topology analysis. ![image](https://user-images.githubusercontent.com/49771382/133862794-2b265de2-14a8-4a50-b100-966547193599.png)

assert len(set(num_channels_list)) == 1

I'll submit a PR to fix this ASAP. I'll let you know when this pr merged. Thanks!

assert len(set(num_channels_list)) == 1

@HappyPeanuts Sincerely sorry for the late PR, already working on this pr and need some time. Thanks~

[New Feature Request] support 3D model compression

You are right, we do not support Conv3d yet.

build Fail: Unsupported gpu architecture 'compute_20'

Got the same error~ Have you fixed this?

[QUESTION] Not supported on A6000?

Yes, Flux only compiled the architectures 80, 89, and 90 for now. However, I suspect that CUTLASS v2 should directly support architecture number 86. Could you try adding the corresponding...

[BUG] Failing to install byte-flux from pypi

Hello @tlrmchlsmth @Rainlin007 , sorry for the late reply. Could you please let me know your torch.__version__ and torch.version.cuda versions?

[BUG] Failing to install byte-flux from pypi

@ZSL98

Are there any difficulties in implementing gemm-allreduce?

@kongroo

[QUESTION] The gemm time on GPU of different rank under tp8 is very different , and cause low performance

You can check whether the frequencies of different GPUs on the server are the same; some GPUs might have downclocked.