import os import torch from torch.utils.data import Sampler from transformers import Trainer from transformers.trainer import ( has_length, ) from typing import List, Optional def maybe_zero_3(param, ignore_status=False, name=None): from deepspeed import zero from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus if hasattr(param, "ds_id"): if param.ds_status == ZeroParamStatus.NOT_AVAILABLE: if not ignore_status: print(name, 'no ignore status') with zero.GatheredParameters([param]): param = param.data.detach().cpu().clone() else: param = param.detach().cpu().clone() return param def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match): to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)} to_return = {k: maybe_zero_3(v, ignore_status=True, name=k).cpu() for k, v in to_return.items()} return to_return def split_to_even_chunks(indices, lengths, num_chunks): """ Split a list of indices into `chunks` chunks of roughly equal lengths. """ if len(indices) % num_chunks != 0: return [indices[i::num_chunks] for i in range(num_chunks)] num_indices_per_chunk = len(indices) // num_chunks chunks = [[] for _ in range(num_chunks)] chunks_lengths = [0 for _ in range(num_chunks)] for index in indices: shortest_chunk = chunks_lengths.index(min(chunks_lengths)) chunks[shortest_chunk].append(index) chunks_lengths[shortest_chunk] += lengths[index] if len(chunks[shortest_chunk]) == num_indices_per_chunk: chunks_lengths[shortest_chunk] = float("inf") return chunks def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None): # We need to use torch for the random part as a distributed sampler will set the random seed for torch. assert all(l != 0 for l in lengths), "Should not have zero length." mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0]) lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0]) assert len(mm_indices) > 0, "Should have at least one multimodal sample." assert len(lang_indices) > 0, "Should have at least one language sample." mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)] lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)] megabatch_size = world_size * batch_size mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)] lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)] last_mm = mm_megabatches[-1] last_lang = lang_megabatches[-1] additional_batch = last_mm + last_lang megabatches = mm_megabatches[:-1] + lang_megabatches[:-1] megabatch_indices = torch.randperm(len(megabatches), generator=generator) megabatches = [megabatches[i] for i in megabatch_indices] if len(additional_batch) >= megabatch_size: megabatches = [additional_batch[:megabatch_size]] + megabatches additional_batch = additional_batch[megabatch_size:] if len(additional_batch) > 0: megabatches.append(additional_batch) return [i for megabatch in megabatches for i in megabatch] def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True): # We need to use torch for the random part as a distributed sampler will set the random seed for torch. indices = torch.randperm(len(lengths), generator=generator) megabatch_size = world_size * batch_size megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)] megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches] megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches] return [i for megabatch in megabatches for batch in megabatch for i in batch] class LengthGroupedSampler(Sampler): r""" Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while keeping a bit of randomness. """ def __init__( self, batch_size: int, world_size: int, lengths: Optional[List[int]] = None, generator=None, group_by_modality: bool = False, ): if lengths is None: raise ValueError("Lengths must be provided.") self.batch_size = batch_size self.world_size = world_size self.lengths = lengths self.generator = generator self.group_by_modality = group_by_modality def __len__(self): return len(self.lengths) def __iter__(self): if self.group_by_modality: indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator) else: indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator) return iter(indices) class LLaVATrainer(Trainer): def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]: if self.train_dataset is None or not has_length(self.train_dataset): return None if self.args.group_by_modality_length: lengths = self.train_dataset.modality_lengths return LengthGroupedSampler( # self.args.train_batch_size * self.args.gradient_accumulation_steps, # TODO: seems that we should not have gradient_accumulation_steps self.args.train_batch_size, world_size=self.args.world_size, lengths=lengths, group_by_modality=True, ) else: return super()._get_train_sampler() def _save_checkpoint(self, model, trial, metrics=None): if getattr(self.args, 'tune_mm_mlp_adapter', False): from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}" run_dir = self._get_output_dir(trial=trial) output_dir = os.path.join(run_dir, checkpoint_folder) # Only save Adapter keys_to_match = ['mm_projector', 'vision_resampler'] if getattr(self.args, "use_im_start_end", False): keys_to_match.extend(['embed_tokens', 'embed_in']) weight_to_save = get_mm_adapter_state_maybe_zero_3(self.model.named_parameters(), keys_to_match) if self.args.local_rank == 0 or self.args.local_rank == -1: self.model.config.save_pretrained(output_dir) torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin')) else: super(LLaVATrainer, self)._save_checkpoint(model, trial, metrics) def _save(self, output_dir: Optional[str] = None, state_dict=None): if getattr(self.args, 'tune_mm_mlp_adapter', False): pass else: super(LLaVATrainer, self)._save(output_dir, state_dict)