import os import torch from torch.utils.data import Sampler from torch.cuda import synchronize from transformers import Trainer from transformers.trainer import ( is_sagemaker_mp_enabled, get_parameter_names, has_length, ALL_LAYERNORM_LAYERS, logger, ) 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." if all(l > 0 for l in lengths) or all(l < 0 for l in lengths): # all samples are in the same modality return get_length_grouped_indices(lengths, batch_size, world_size, generator=generator) 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]) 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) > 0: megabatches.append(sorted(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) # indices = torch.arange(len(lengths)) 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 # gather self lengths from all processes # if self.world_size > 1: # # gather the size of lengths from all processes # sizes = torch.tensor([len(lengths)], device=torch.device("cuda")) # # take minimum length # torch.distributed.all_reduce(sizes, op=torch.distributed.ReduceOp.MIN) # min_size = sizes.item() # # trim lengths to the minimum size # lengths = lengths[:min_size] # # append lengths from all processes # all_lengths = [torch.zeros_like(lengths) for _ in range(self.world_size)] # torch.distributed.all_gather(all_lengths, torch.tensor(lengths, device=torch.device("cuda"))) # lengths = torch.cat(all_lengths, dim=0).tolist() # if torch.distributed.get_rank() == 0: # import pdb; pdb.set_trace() 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 MagmaTrainer(Trainer): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.loss_moving_average = None 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, world_size=self.args.world_size * self.args.gradient_accumulation_steps, lengths=lengths, group_by_modality=True, ) else: return super()._get_train_sampler() def evaluation_loop(self, dataloader, description: str, prediction_loss_only: Optional[bool] = None, ignore_keys=None, metric_key_prefix: str = "eval"): """ Override the `evaluation_loop` method for custom evaluation. """ # Custom logic before evaluation loop starts print(f"Starting custom evaluation loop: {description}") pass synchronize() # This ensures all GPU operations are completed # Initialize containers for predictions and labels all_preds = [] all_labels = [] # Iterate over the evaluation data loader for step, inputs in enumerate(dataloader): # Optionally, apply the data collator manually here if needed # inputs = self.data_collator(inputs) # Move batch to the appropriate device inputs = self._prepare_inputs(inputs) # Disable gradient calculation during evaluation with torch.no_grad(): outputs = self.model(**inputs) # Extract logits (predictions) and labels logits = outputs.logits labels = inputs['labels'] # Collect predictions and labels for this batch preds = logits.argmax(dim=-1) # Assuming classification task all_preds.append(preds.cpu().numpy()) all_labels.append(labels.cpu().numpy()) # Concatenate all batches to get complete predictions and labels all_preds = np.concatenate(all_preds, axis=0) all_labels = np.concatenate(all_labels, axis=0) # Custom metric computation logic can be applied here metrics = self.compute_metrics((all_preds, all_labels)) print(f"Finished custom evaluation loop. Metrics: {metrics}") # Return the results as expected by the trainer (you can customize this) return metrics def create_optimizer(self): """ Setup the optimizer. We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer's init through `optimizers`, or subclass and override this method in a subclass. """ opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model if self.optimizer is None: decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS) decay_parameters = [name for name in decay_parameters if "bias" not in name] optimizer_grouped_parameters = [] for n, p in opt_model.named_parameters(): if p.requires_grad: if "mm_projector" in n and self.args.mm_projector_lr is not None: optimizer_grouped_parameters.append( { "params": [p], "weight_decay": self.args.weight_decay if n in decay_parameters else 0.0, "lr": self.args.mm_projector_lr, } ) elif "vision_tower" in n and self.args.vision_tokenizer_lr is not None: optimizer_grouped_parameters.append( { "params": [p], "weight_decay": self.args.weight_decay if n in decay_parameters else 0.0, "lr": self.args.vision_tokenizer_lr, } ) else: optimizer_grouped_parameters.append( { "params": [p], "weight_decay": self.args.weight_decay if n in decay_parameters else 0.0, } ) optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args) self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs) if optimizer_cls.__name__ == "Adam8bit": import bitsandbytes manager = bitsandbytes.optim.GlobalOptimManager.get_instance() skipped = 0 for module in opt_model.modules(): if isinstance(module, nn.Embedding): skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values()) logger.info(f"skipped {module}: {skipped/2**20}M params") manager.register_module_override(module, "weight", {"optim_bits": 32}) logger.debug(f"bitsandbytes: will optimize {module} in fp32") logger.info(f"skipped: {skipped/2**20}M params") return self.optimizer 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', 'segtok_'] if getattr(self.args, "use_im_start_end", False): keys_to_match.extend(['embed_tokens', 'embed_in']) if getattr(self.args, "tune_vision_tokenizer", 'none') == "posembed": keys_to_match.extend(['posembed']) elif getattr(self.args, "tune_vision_tokenizer", 'none') == "decoder": keys_to_match.extend(['sem_seg_head.predictor']) elif getattr(self.args, "tune_vision_tokenizer", 'none') == "all": keys_to_match.extend(['vision_tower']) 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) print(f"keys to match: {keys_to_match}") print(f"save checkpoint to {os.path.join(output_dir, f'mm_projector.bin')}") torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin')) else: super(MagmaTrainer, 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(MagmaTrainer, self)._save(output_dir, state_dict)