ViperEkura ef8783b7e3 fix: separate attn_mask and loss_mask in get_logprobs, compose causal masking in strategy ...

- add loss_mask parameter to get_logprobs to decouple attention from loss masking
- DPO/GRPO strategies compose key-padding + causal mask before model forward
- prevents prompt tokens from being masked out of attention and missing causal masking

2026-07-21 23:47:28 +08:00

..

config

refactor : align config field names with Hugging Face

2026-07-20 22:05:31 +08:00

dataset

feat: support SFT directly from JSONL without dataset_config.json

2026-07-20 17:25:09 +08:00

extension

feat: add attention dimension dispatch

2026-07-21 12:52:45 +08:00

inference

perf: eliminate GPU syncs in contiguous cache write/gather hot paths

2026-07-21 16:40:56 +08:00

model

refactor : align config field names with Hugging Face

2026-07-20 22:05:31 +08:00

parallel

feat: add FSDP2Executor using torch.distributed.fsdp.fully_shard API

2026-07-20 01:46:25 +08:00

preprocessing

refactor: map instruction/input/output to chat roles

2026-07-20 13:55:25 +08:00

serialization

fix: memmap mode=r, tool parser json.loads, greedy decode

2026-07-19 16:38:28 +08:00

tokenize

refactor: unify tokenizer encode and apply_chat_template for batch support

2026-07-20 17:49:26 +08:00

trainer

fix: separate attn_mask and loss_mask in get_logprobs, compose causal masking in strategy

2026-07-21 23:47:28 +08:00

__init__.py

chore: bump to 1.3.10

2026-07-19 16:40:27 +08:00

factory.py

fix: ForwardRef._evaluate Python 3.12 compatibility

2026-07-03 18:41:19 +08:00

protocols.py

refactor: 重构训练后端为 Executor 模式

2026-05-24 20:35:44 +08:00