hyper-parallel-recompute

文件	最后提交记录	最后更新时间
callbacks	feat: add training monitoring function	3 个月前
utils	refactor: behavior-neutral code cleanup Behavior-neutral cleanups across the trainer and models — no functional change: relocate module __all__ above imports and move the entry-script deferred imports to the top; mark the instance-free StateDictAdapter load/save methods and _make_amp_context as @staticmethod; declare BaseTrainer attributes at class level (switching the sampler / _last_global_tokens reads to is-None / getattr fallbacks); use dict.get for guarded lookups; return a namedtuple from _build_grad_groups; extract _compute_shard_overrides; minor docstring/formatting. Callback lifecycle hooks stay instance methods (overridable interface — keeping them consistent with subclass overrides).	4 个月前
__init__.py	refactor: behavior-neutral code cleanup Behavior-neutral cleanups across the trainer and models — no functional change: relocate module __all__ above imports and move the entry-script deferred imports to the top; mark the instance-free StateDictAdapter load/save methods and _make_amp_context as @staticmethod; declare BaseTrainer attributes at class level (switching the sampler / _last_global_tokens reads to is-None / getattr fallbacks); use dict.get for guarded lookups; return a namedtuple from _build_grad_groups; extract _compute_shard_overrides; minor docstring/formatting. Callback lifecycle hooks stay instance methods (overridable interface — keeping them consistent with subclass overrides).	4 个月前
base.py	feat: add training monitoring function	3 个月前
config.py	feat: add training monitoring function	3 个月前
llm_trainer.py	feat: support Qwen3.5 and Qwen3-VL-MoE training Add dense, MoE, and vision-language model integration with checkpoint conversion, data loading, and unified trainer execution. Correct TP, CP, EP, PP, FSDP, and HSDP synchronization semantics so supported BF16 single-card and hybrid-parallel configurations follow the same training behavior.	3 个月前
parallel_dims.py	feat: support Qwen3.5 and Qwen3-VL-MoE training Add dense, MoE, and vision-language model integration with checkpoint conversion, data loading, and unified trainer execution. Correct TP, CP, EP, PP, FSDP, and HSDP synchronization semantics so supported BF16 single-card and hybrid-parallel configurations follow the same training behavior.	3 个月前
vl_trainer.py	feat: support Qwen3.5 and Qwen3-VL-MoE training Add dense, MoE, and vision-language model integration with checkpoint conversion, data loading, and unified trainer execution. Correct TP, CP, EP, PP, FSDP, and HSDP synchronization semantics so supported BF16 single-card and hybrid-parallel configurations follow the same training behavior.	3 个月前