Charles Yang comments

Results 9 comments of


                                            Charles Yang

3D reconstructed quality is poor with ROCm as video, 2000 iteration

@pfxuan this is another dataset, truck, training 15000 iterations, the final loss >0.4, seems not converged. Radeon 7900 XTX + ROCm 6.3.3 + Ubuntu22.04 + torch 2.1.2. [truck15k.zip](https://github.com/user-attachments/files/19718681/truck15k.zip) [15ktruck.txt](https://github.com/user-attachments/files/19718683/15ktruck.txt) ![Image](https://github.com/user-attachments/assets/eb7dfefd-8f12-4fcd-b6a0-a521b7c81394)

有没有方法把7b模型替换成更大的？

same question, qwen2 is relative out of time, is qwen3 verified? @Vincentwei1021 @wangxiongts @linhaojia13 @lxysl @BradyFU

rocm6.3.3 on Ubuntu 24.04, loss won't go down after 2000 iters, while Ubuntu22.04 is fine

@jcaesar @eokeeffe @pierotofy @pfxuan could you give some advice?

rocm6.3.3 on Ubuntu 24.04, loss won't go down after 2000 iters, while Ubuntu22.04 is fine

@pfxuan @eokeeffe @pierotofy I made system experiments. The issue is caused by Ubuntu24.04. Same hardware, same version of ROCm and PyTorch, it works on Ubuntu 22.04 but failed on Ubuntu...

rocm6.3.3 on Ubuntu 24.04, loss won't go down after 2000 iters, while Ubuntu22.04 is fine

@pfxuan @eokeeffe @pierotofy Even I running on a Ubuntu24.04 host and Ubuntu 22.04 container, it won't work.

rocm6.3.3 on Ubuntu 24.04, loss won't go down after 2000 iters, while Ubuntu22.04 is fine

So apologize for any disturbing. I just want to raise attention for this issue which show stopper on AMD hardware. take it easy and thanks for reminding.

这个本地可以使用AMD显卡跑吗

mark，同求amd方案

显卡

同求 AMD方案

AMD Radeon 780M Graphics 显卡是否可以安装成功？

同求AMD Radeon方案