Dinghao Zhou

拆分该pr：https://github.com/wenet-e2e/wenet/pull/2460

fix rdzv_id

2

rdzv_id may be not equal in multi node setup using RANDOM

when to support gpu?

2

great work! But when to support/release gpu?

多模态和部分大模型中采用qk norm 来稳定训练 （apple dmel中也用到了）(有益于bestrq 训练 和稳定梯度) TODO: - [ ] conformer result

sensevoice small arch 和paraformer 的encoder 几乎一致， cmvn也是一样 这里支持下sensevoice的fintune - [ ] convert script

- [ ] 需要验证下 ```python # usage1: json_files = ["1.jsonl", "2.jsonl", "3.jsonl"] dataset = WenetRawDatasetSource(json_files, partiaion=True, shard_by_files=True) # usage2: json_files = "all.jsonl" dataset = WenetRawDatasetSource(json_files, partiaion=True, shard_by_files=True) ```

TODO - [ ] make it work - [ ] causal mask for conformer (need update encoder forward args) - [x] subsequent tokens id for loss