Boyuan Feng

"Then I started the training using the default configuration. But the model cannot detect any object. I have tested after training 1, 2, ..., 8 epoches. The training loss looks...

@1292765944 I use two gpus to train and get [email protected]

> @BoyuanFeng can this have a prototype by Aug 12 or is that too early? I am still working on manta-rs such as circuits and trusted setup migration. I need...

I confronted the same issue. I guess the reason is that, at Line [340](https://github.com/pcyin/pytorch_basic_nmt/blob/75bafa2d05118a4e8a0e8691a986ab9dd09bb93d/nmt.py#L340), you changed the "prev_hyp_ids = top_cand_hyp_pos / len(self.vocab.tgt)" to something like "prev_hyp_ids = top_cand_hyp_pos / (1.0*len(self.vocab.tgt))"....

Thanks for quick reply! I tried `webdriver.json` with `{ "goog:chromeOptions": { "args": [ "disable-timeouts-for-profiling" ] } }` However, this one does not work... The same timeout error appears. Do you...

Parallelization could definitely reduce the latency of proof generation. Parallelization can help even if we only generate proof for 1 transaction. More specifically, many time-consuming operations in proof generation such...

nit: noop elimination for slice errors when end = -1. repro: ``` import torch def dims_equivalent(dim, i_dim) -> bool: # Case 1 and 2 if dim == i_dim or dim...

issue: #17078 Btw, pytorch supports noop elimination for [view](https://github.com/pytorch/pytorch/pull/151095), [slice, and slice_scatter](https://github.com/pytorch/pytorch/pull/151175) now, which should be equivalent with `noop_elimination.py`.

TLDR: The issue seems to related to [train.scheduler](https://github.com/JonasGeiping/cramming/blob/main/cramming/config/train/bert-o4.yaml#L20) in user code instead of cudagraph. 1. I can repeat the precision behavior using the provided command and user code. To simplify...

@pytorchbot merge