DachengLi1
commented
1 year ago @brewswang Thanks for trying out the training code! In this release, the code has only been tested on 8xA100 for a 7B model, because of the very long sequence length causes high memory consumption. To run on V100 16GB, first change the monkey_patch here from flash attention to xformer.
There are several things to try:
(1) use FSDP cpu offloading. (2) Try lower sequence length. (remember also change the ratio to lower ones, e.g. 2 for 4K, 4 for 8K).
Let me know if it works for you!
brewswang
ChaoyuHuang
I have 9 V100 16G GPUs,but training CUDA out of memory. The specific errors are as follows: Formatting inputs...Skip in lazy mode /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/_init_utils.py:295: UserWarning: FSDP is switching to use │
│ │
│ 12 from longchat.train.fine_tune.train import train │
│ 13 │
│ 14 if name == "main": │
│ ❱ 15 │ train() │
│ 16 │
│ │
│ /nvme/soft/brewswang/chatgpt/LongChat/longchat/train/fine_tune/train.py:262 in train │
│ │
│ 259 │ if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")): │
│ 260 │ │ trainer.train(resume_from_checkpoint=True) │
│ 261 │ else: │
│ ❱ 262 │ │ trainer.train() │
│ 263 │ trainer.save_state() │
│ 264 │ safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir) │
│ 265 │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/transformers/trainer.py:16 │
│ 62 in train │
│ │
│ 1659 │ │ inner_training_loop = find_executable_batch_size( │
│ 1660 │ │ │ self._inner_training_loop, self._train_batch_size, args.auto_find_batch_size │
│ 1661 │ │ ) │
│ ❱ 1662 │ │ return inner_training_loop( │
│ 1663 │ │ │ args=args, │
│ 1664 │ │ │ resume_from_checkpoint=resume_from_checkpoint, │
│ 1665 │ │ │ trial=trial, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/transformers/trainer.py:17 │
│ 49 in _inner_training_loop │
│ │
│ 1746 │ │ if args.gradient_checkpointing: │
│ 1747 │ │ │ self.model.gradient_checkpointing_enable() │
│ 1748 │ │ │
│ ❱ 1749 │ │ model = self._wrap_model(self.model_wrapped) │
│ 1750 │ │ │
│ 1751 │ │ if is_sagemaker_mp_enabled() and resume_from_checkpoint is not None: │
│ 1752 │ │ │ self._load_from_checkpoint(resume_from_checkpoint, model) │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/transformers/trainer.py:14 │
│ 89 in _wrap_model │
│ │
│ 1486 │ │ │ │ │ for arg in ["limit_all_gathers", "forward_prefetch", "backward_prefe │
│ 1487 │ │ │ │ │ │ if arg in signature: │
│ 1488 │ │ │ │ │ │ │ kwargs[arg] = getattr(self, arg) │
│ ❱ 1489 │ │ │ │ │ self.model = model = FSDP( │
│ 1490 │ │ │ │ │ │ model, │
│ 1491 │ │ │ │ │ │ sharding_strategy=self.fsdp, │
│ 1492 │ │ │ │ │ │ cpu_offload=cpu_offload, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/ful │
│ ly_sharded_data_parallel.py:391 in init │
│ │
│ 388 │ │ │ │ # process groups. │
│ 389 │ │ │ │ fsdp_kwargs["process_group"] = (self.process_group, self._inter_node_pg) │
│ 390 │ │ │ │
│ ❱ 391 │ │ │ _auto_wrap(auto_wrap_kwargs, fsdp_kwargs, FullyShardedDataParallel) │
│ 392 │ │ │
│ 393 │ │ backward_prefetch_limit = 1 │
│ 394 │ │ forward_prefetch_limit = 1 │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/_wr │
│ ap_utils.py:73 in _auto_wrap │
│ │
│ 70 │ │ │ "kernels do not support low precision." │
│ 71 │ │ ) │
│ 72 │ auto_wrap_kwargs["auto_wrap_policy"] = auto_wrap_policy │
│ ❱ 73 │ _recursive_wrap(auto_wrap_kwargs, fsdp_kwargs) │
│ 74 │
│ 75 │
│ 76 def _get_fully_sharded_module_to_states( │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/wra │
│ p.py:370 in _recursive_wrap │
│ │
│ 367 │ │ for name, child in module.named_children(): │
│ 368 │ │ │ if child in ignored_modules: │
│ 369 │ │ │ │ continue │
│ ❱ 370 │ │ │ wrapped_child, num_wrapped_params = _recursive_wrap( │
│ 371 │ │ │ │ module=child, │
│ 372 │ │ │ │ auto_wrap_policy=auto_wrap_policy, │
│ 373 │ │ │ │ wrapper_cls=wrapper_cls, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/wra │
│ p.py:370 in _recursive_wrap │
│ │
│ 367 │ │ for name, child in module.named_children(): │
│ 368 │ │ │ if child in ignored_modules: │
│ 369 │ │ │ │ continue │
│ ❱ 370 │ │ │ wrapped_child, num_wrapped_params = _recursive_wrap( │
│ 371 │ │ │ │ module=child, │
│ 372 │ │ │ │ auto_wrap_policy=auto_wrap_policy, │
│ 373 │ │ │ │ wrapper_cls=wrapper_cls, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/wra │
│ p.py:370 in _recursive_wrap │
│ │
│ 367 │ │ for name, child in module.named_children(): │
│ 368 │ │ │ if child in ignored_modules: │
│ 369 │ │ │ │ continue │
│ ❱ 370 │ │ │ wrapped_child, num_wrapped_params = _recursive_wrap( │
│ 371 │ │ │ │ module=child, │
│ 372 │ │ │ │ auto_wrap_policy=auto_wrap_policy, │
│ 373 │ │ │ │ wrapper_cls=wrapper_cls, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/wra │
│ p.py:388 in _recursive_wrap │
│ │
│ 385 │ │ │ module=module, recurse=False, nonwrapped_numel=remainder │
│ 386 │ │ ): │
│ 387 │ │ │ # Leaf node or final wrapping of the remainder both happen here. │
│ ❱ 388 │ │ │ return _wrap(module, wrapper_cls, kwargs), nonwrapped_numel │
│ 389 │ │ else: │
│ 390 │ │ │ return module, total_wrapped_numel │
│ 391 │ return module, 0 │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/wra │
│ p.py:317 in _wrap │
│ │
│ 314 │ │ overrides = {kwargs, module._wrap_overrides} # type: ignore[arg-type] │
│ 315 │ │ return wrapper_cls(module, overrides) │
│ 316 │ │
│ ❱ 317 │ return wrapper_cls(module, **kwargs) │
│ 318 │
│ 319 │
│ 320 def _recursive_wrap( │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/ful │
│ ly_sharded_data_parallel.py:408 in init │
│ │
│ 405 │ │ _init_runtime_state(self) │
│ 406 │ │ _init_prefetching_state(self, backward_prefetch, forward_prefetch) │
│ 407 │ │ _init_buffer_state(self, module) │
│ ❱ 408 │ │ _init_param_handle_from_module( │
│ 409 │ │ │ self, │
│ 410 │ │ │ module, │
│ 411 │ │ │ device_id, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/_in │
│ it_utils.py:429 in _init_param_handle_from_module │
│ │
│ 426 │ │ _sync_module_params_and_buffers( │
│ 427 │ │ │ fully_sharded_module, managed_params, state.process_group │
│ 428 │ │ ) │
│ ❱ 429 │ _init_param_handle_from_params(state, managed_params, fully_sharded_module) │
│ 430 │ return state │
│ 431 │
│ 432 │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/_in │
│ it_utils.py:525 in _init_param_handle_from_params │
│ │
│ 522 ): │
│ 523 │ if len(params) == 0: │
│ 524 │ │ return │
│ ❱ 525 │ handle = FlatParamHandle( │
│ 526 │ │ params, │
│ 527 │ │ fully_sharded_module, │
│ 528 │ │ state.compute_device, │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/fla │
│ t_param.py:366 in init │
│ │
│ 363 │ │ self._training_state = HandleTrainingState.IDLE │
│ 364 │ │ self._debug_level = dist.get_debug_level() │
│ 365 │ │ self._fully_sharded_module = fully_sharded_module │
│ ❱ 366 │ │ self._init_flat_param(params, fully_sharded_module, use_orig_params) │
│ 367 │ │ self._orig_param_dtype = self.flat_param.dtype │
│ 368 │ │ self._use_unsharded_views(as_params=False) │
│ 369 │ │ self._init_param_reduce_dtypes(mp_param_dtype, mp_reduce_dtype) │
│ │
│ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/fla │
│ t_param.py:462 in _init_flat_param │
│ │
│ 459 │ │ │ "Passed-in
main()
File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/init.py", line 346, in wrapper
return f(*args, **kwargs)
File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/run.py", line 794, in main
run(args)
File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 134, in call
return launch_agent(self._config, self._entrypoint, list(args))
File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
NO_SHARDinstead of ShardingStrategy.FULL_SHARD since the world size is 1. warnings.warn( ╭─────────────────────────────── Traceback (most recent call last) ────────────────────────────────╮ │ /nvme/soft/brewswang/chatgpt/LongChat/longchat/train/fine_tune/train_condense_16K.py:15 in │ │paramswere not found in the module tree\n" │ │ 460 │ │ │ f"params: {params}\nmodule: {module}" │ │ 461 │ │ ) │ │ ❱ 462 │ │ self.flat_param = FlatParamHandle.flatten_params( │ │ 463 │ │ │ params_to_flatten, requires_grad │ │ 464 │ │ ) │ │ 465 │ │ # Foruse_orig_params=True, ensure that the logical parameters are │ │ │ │ /home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/fsdp/fla │ │ t_param.py:505 in flatten_params │ │ │ │ 502 │ │ │ │ p.detach().reshape(-1) if isinstance(p, nn.Parameter) else p.reshape(-1) │ │ 503 │ │ │ │ for p in params │ │ 504 │ │ │ ] │ │ ❱ 505 │ │ │ flat_param_data = torch.cat(flat_params, dim=0) │ │ 506 │ │ flat_param = FlatParameter(flat_param_data, requires_grad=requires_grad) │ │ 507 │ │ return flat_param │ │ 508 │ ╰──────────────────────────────────────────────────────────────────────────────────────────────────╯ OutOfMemoryError: CUDA out of memory. Tried to allocate 774.00 MiB (GPU 0; 15.78 GiB total capacity; 14.62 GiB already allocated; 369.69 MiB free; 14.72 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: 1) local_rank: 0 (pid: 23015) of binary: /home/chat_glm6b/anaconda3/envs/longeval/bin/python Traceback (most recent call last): File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/runpy.py", line 196, in _run_module_as_main return _run_code(code, main_globals, None, File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/runpy.py", line 86, in _run_code exec(code, run_globals) File "/home/chat_glm6b/anaconda3/envs/longeval/lib/python3.10/site-packages/torch/distributed/run.py", line 798, inlongchat/train/fine_tune/train_condense_16K.py FAILED
Failures: