Zhen

Results 9 comments of Zhen

Implementation Bug & Error ?

Sorry, I still can not achieve the results mentioned in the paper.

Add Ascend NPU as a backend for single device recipes

It is very helpful for me, nice work!

[recipe] fix: Qwen3-vl npu patch

> I’m on board with the PR itself, but our NPU-patch has added more and more Transformers compatibility changes over time—leading to noticeable maintenance difficulties. It’s time to upgrade the...

load_dataset("livecodebench/code_generation_lite", version_tag="release_v2") TypeError: 'NoneType' object is not callable

@Serzhanov @zwxandy I have met the same problem, have this problem be solved?

load_dataset("livecodebench/code_generation_lite", version_tag="release_v2") TypeError: 'NoneType' object is not callable

> [@Serzhanov](https://github.com/Serzhanov) [@zwxandy](https://github.com/zwxandy) I have met the same problem, have this problem be solved? I try to downgrade datasets version to 2.20.0，and it works for me @Serzhanov @dshwei , hope...

[rollout] fix: RolloutConfig should support repetition_penalty config…

@Lokiscripter Some CI failed, please fix pre-commit and rebase your code to latest version

[fsdp] feat: update NPU fused kernels for Qwen3 moe block

e2e_ppo_trainer_megatron_vllm_2 fail but not related with this PR, ignore.

[fsdp] feat: add NPU fusion kernels for Qwen2 and Qwen2.5 dense model.

@ZLiao097 please pull and rebase the newest code from `main`, there are some modifications about `e2e_ascend`.

[WIP][megatron] feat: support discrete profiling for mindspeed

> After configuring profiling for tests/speical_npu/run_qwen2_5_05b_grpo.sh, the profiling directory is as follows, > > * * root > | - actor_compute_log_prob > | - actor_update > | - ref_compute_log_prob >...