refactor model_dtype, fix PPO trainer

src/llmtuner/extras/misc.py

@@ -3,6 +3,19 @@
 from typing import TYPE_CHECKING, Tuple
 from transformers import InfNanRemoveLogitsProcessor, LogitsProcessorList
 
+try:
+    from transformers.utils import (
+        is_torch_bf16_cpu_available,
+        is_torch_bf16_gpu_available,
+        is_torch_cuda_available,
+        is_torch_npu_available
+    )
+    _is_fp16_available = is_torch_npu_available() or is_torch_cuda_available()
+    _is_bf16_available = is_torch_bf16_gpu_available() or is_torch_bf16_cpu_available
+except ImportError:
+    _is_fp16_available = torch.cuda.is_available()
+    _is_bf16_available = torch.cuda.is_bf16_supported()
+
 if TYPE_CHECKING:
     from transformers.modeling_utils import PreTrainedModel
 
@@ -49,7 +62,22 @@ def count_parameters(model: torch.nn.Module) -> Tuple[int, int]:
     return trainable_params, all_param
 
 
+def infer_optim_dtype(model_dtype: torch.dtype) -> torch.dtype:
+    r"""
+    Infers the optimal dtype according to the model_dtype and device compatibility.
+    """
+    if _is_bf16_available and model_dtype == torch.bfloat16:
+        return torch.bfloat16
+    elif _is_fp16_available:
+        return torch.float16
+    else:
+        return torch.float32
+
+
 def get_logits_processor() -> LogitsProcessorList:
+    r"""
+    Gets logits processor that removes NaN and Inf logits.
+    """
     logits_processor = LogitsProcessorList()
     logits_processor.append(InfNanRemoveLogitsProcessor())
     return logits_processor

src/llmtuner/extras/patches/llama_patch.py

@@ -138,11 +138,11 @@ def forward(
         input_dtype = query_states.dtype
         if input_dtype == torch.float32:
             logger.warning_once("The input hidden states seems to be silently casted in float32.")
-            query_states = query_states.to(torch.float16)
-            key_states = key_states.to(torch.float16)
-            value_states = value_states.to(torch.float16)
+            query_states = query_states.to(self.config.torch_dtype)
+            key_states = key_states.to(self.config.torch_dtype)
+            value_states = value_states.to(self.config.torch_dtype)
 
-        if getattr(self, "num_key_value_groups"):
+        if getattr(self, "num_key_value_groups", None):
             key_states = repeat_kv(key_states, self.num_key_value_groups)
             value_states = repeat_kv(value_states, self.num_key_value_groups)
 

src/llmtuner/hparams/model_args.py

@@ -67,9 +67,9 @@ class ModelArguments:
         default=None,
         metadata={"help": "Auth token to log in with Hugging Face Hub."}
     )
-    layernorm_dtype: Optional[Literal["auto", "fp16", "bf16", "fp32"]] = field(
-        default="auto",
-        metadata={"help": "Data type of the layer norm weights."}
+    upcast_layernorm: Optional[bool] = field(
+        default=False,
+        metadata={"help": "Whether to upcast the layernorm weights in fp32."}
     )
 
     def __post_init__(self):

src/llmtuner/tuner/core/loader.py

@@ -24,7 +24,7 @@
     from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from llmtuner.extras.logging import reset_logging, get_logger
-from llmtuner.extras.misc import count_parameters
+from llmtuner.extras.misc import count_parameters, infer_optim_dtype
 from llmtuner.extras.patches import llama_patch as LlamaPatches
 from llmtuner.extras.save_and_load import load_valuehead_params
 from llmtuner.hparams import FinetuningArguments
@@ -86,11 +86,17 @@ def load_model_and_tokenizer(
     if getattr(config, "model_type", None) == "chatglm":
         tokenizer._pad = MethodType(PreTrainedTokenizerBase._pad, tokenizer)
 
+    # Set model dtype
+    if model_args.compute_dtype is not None:
+        setattr(config, "torch_dtype", model_args.compute_dtype)
+    else: # priority: bf16 > fp16 > fp32
+        optim_dtype = infer_optim_dtype(model_dtype=getattr(config, "torch_dtype", None))
+        setattr(config, "torch_dtype", optim_dtype)
+
     # Fix config (for Qwen)
     if getattr(config, "model_type", None) == "qwen":
-        setattr(config, "fp16", model_args.compute_dtype == torch.float16)
-        setattr(config, "bf16", model_args.compute_dtype == torch.bfloat16)
-        setattr(config, "fp32", model_args.compute_dtype == torch.float32)
+        for dtype_name, dtype in [("fp16", torch.float16), ("bf16", torch.bfloat16), ("fp32", torch.float32)]:
+            setattr(config, dtype_name, getattr(config, "torch_dtype", None) == dtype)
 
     # Set RoPE scaling
     if model_args.rope_scaling is not None:
@@ -131,9 +137,7 @@ def load_model_and_tokenizer(
     if model_args.flash_attn:
         if getattr(config, "model_type", None) == "llama":
             LlamaModule.LlamaAttention = LlamaPatches.LlamaFlashAttention2
-            LlamaModule.LlamaModel._prepare_decoder_attention_mask = (
-                LlamaPatches._prepare_decoder_attention_mask
-            )
+            LlamaModule.LlamaModel._prepare_decoder_attention_mask = LlamaPatches._prepare_decoder_attention_mask
             logger.info("Using FlashAttention-2 for faster training and inference.")
         elif getattr(config, "model_type", None) == "qwen":
             logger.info("Qwen models automatically enable FlashAttention if installed.")
@@ -180,7 +184,6 @@ def load_model_and_tokenizer(
     model = AutoModelForCausalLM.from_pretrained(
         model_to_load,
         config=config,
-        torch_dtype=model_args.compute_dtype,
         low_cpu_mem_usage=(not is_deepspeed_zero3_enabled()),
         **config_kwargs
     )
@@ -203,7 +206,7 @@ def load_model_and_tokenizer(
 
     # Initialize adapters
     if is_trainable:
-        model = prepare_model_for_training(model, model_args.layernorm_dtype, finetuning_args.finetuning_type)
+        model = prepare_model_for_training(model, model_args.upcast_layernorm, finetuning_args.finetuning_type)
     model = init_adapter(model, model_args, finetuning_args, is_trainable, is_mergeable)
     model = model.train() if is_trainable else model.eval()
 

src/llmtuner/tuner/core/parser.py

@@ -8,16 +8,6 @@
 from transformers.utils.versions import require_version
 from transformers.trainer_utils import get_last_checkpoint
 
-try:
-    from transformers.utils import is_torch_bf16_gpu_available, is_torch_npu_available, is_torch_cuda_available
-    is_fp16_available = is_torch_cuda_available()
-    is_bf16_available = is_torch_bf16_gpu_available()
-    is_npu_available = is_torch_npu_available()
-except ImportError:
-    is_fp16_available = torch.cuda.is_available()
-    is_bf16_available = torch.cuda.is_bf16_supported()
-    is_npu_available = False
-
 from llmtuner.extras.logging import get_logger
 from llmtuner.hparams import (
     ModelArguments,
@@ -31,17 +21,6 @@
 logger = get_logger(__name__)
 
 
-def _infer_dtype() -> torch.dtype:
-    if is_npu_available:
-        return torch.float16
-    elif is_bf16_available:
-        return torch.bfloat16
-    elif is_fp16_available:
-        return torch.float16
-    else:
-        return torch.float32
-
-
 def _parse_args(parser: HfArgumentParser, args: Optional[Dict[str, Any]] = None) -> Tuple[Any]:
     if args is not None:
         return parser.parse_dict(args)
@@ -178,12 +157,15 @@ def get_train_args(
             if not finetuning_args.resume_lora_training:
                 raise ValueError("Quantized model cannot create new LoRA weight. Merge them first.")
 
-    if model_args.quantization_bit is not None and (not training_args.do_train):
-        logger.warning("Evaluating model in 4/8-bit mode may cause lower scores.")
+    if training_args.do_train and model_args.quantization_bit is not None and (not model_args.upcast_layernorm):
+        logger.warning("We recommend enable `upcast_layernorm` in quantized training.")
 
     if training_args.do_train and (not training_args.fp16) and (not training_args.bf16):
         logger.warning("We recommend enable mixed precision training.")
 
+    if (not training_args.do_train) and model_args.quantization_bit is not None:
+        logger.warning("Evaluating model in 4/8-bit mode may cause lower scores.")
+
     # postprocess data_args
     if data_args.max_samples is not None and data_args.streaming:
         logger.warning("`max_samples` is incompatible with `streaming`. Disabling max_samples.")
@@ -206,10 +188,9 @@ def get_train_args(
         and os.path.isdir(training_args.output_dir)
         and not training_args.overwrite_output_dir
     ):
-        require_version("transformers>=4.31.0", "Resuming training requires transformers>=4.31.0.")
         last_checkpoint = get_last_checkpoint(training_args.output_dir)
         if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
-            raise ValueError("Output directory already exists and is not empty. Use `overwrite_output_dir`.")
+            raise ValueError("Output directory already exists and is not empty. Please set `overwrite_output_dir`.")
 
         if last_checkpoint is not None:
             training_args_dict = training_args.to_dict()
@@ -220,26 +201,7 @@ def get_train_args(
             )
 
     # postprocess model_args
-    if training_args.bf16:
-        if not is_bf16_available:
-            raise ValueError("Current device does not support bf16 training.")
-        model_args.compute_dtype = torch.bfloat16
-    elif training_args.fp16:
-        model_args.compute_dtype = torch.float16
-    else:
-        model_args.compute_dtype = _infer_dtype()
-
-    if model_args.layernorm_dtype == "bf16":
-        if not is_bf16_available:
-            raise ValueError("Current device does not support bf16 type.")
-        model_args.layernorm_dtype = torch.bfloat16
-    elif model_args.layernorm_dtype == "fp16":
-        model_args.layernorm_dtype = torch.float16
-    elif model_args.layernorm_dtype == "fp32":
-        model_args.layernorm_dtype = torch.float32
-    else:
-        model_args.layernorm_dtype = model_args.compute_dtype
-
+    model_args.compute_dtype = torch.bfloat16 if training_args.bf16 else (torch.float16 if training_args.fp16 else None)
     model_args.model_max_length = data_args.cutoff_len
 
     # Log on each process the small summary:
@@ -278,7 +240,4 @@ def get_infer_args(
         if model_args.quantization_bit is not None and len(model_args.checkpoint_dir) != 1:
             raise ValueError("Quantized model only accepts a single checkpoint. Merge them first.")
 
-    # auto-detect cuda capability
-    model_args.compute_dtype = _infer_dtype()
-
     return model_args, data_args, finetuning_args, generating_args

src/llmtuner/tuner/core/utils.py

@@ -31,11 +31,11 @@ def find_all_linear_modules(
 
 def prepare_model_for_training(
     model: "PreTrainedModel",
-    layernorm_dtype: torch.dtype,
+    upcast_layernorm: bool,
     finetuning_type: str,
     output_layer_name: Optional[str] = "lm_head",
     use_gradient_checkpointing: Optional[bool] = True,
-    layer_norm_names: Optional[List[str]] = LAYERNORM_NAMES
+    layernorm_names: Optional[List[str]] = LAYERNORM_NAMES
 ) -> "PreTrainedModel":
     r"""
     Includes:
@@ -44,9 +44,10 @@ def prepare_model_for_training(
         (3) upcast the lm_head to fp32
     Inspired by: https://github.com/huggingface/peft/blob/v0.2.0/src/peft/utils/other.py#L33
     """
-    for name, param in model.named_parameters():
-        if param.ndim == 1 and any(layer_norm_name in name for layer_norm_name in layer_norm_names):
-            param.data = param.data.to(layernorm_dtype)
+    if upcast_layernorm:
+        for name, param in model.named_parameters():
+            if param.ndim == 1 and any(ln_name in name for ln_name in layernorm_names):
+                param.data = param.data.to(torch.float32)
 
     if use_gradient_checkpointing:
         if hasattr(model, "enable_input_require_grads"):

src/llmtuner/tuner/ppo/trainer.py

@@ -10,14 +10,15 @@
 from trl import PPOTrainer
 from trl.core import PPODecorators, logprobs_from_logits
 
+from llmtuner.extras.callbacks import LogCallback, SavePeftModelCallback
 from llmtuner.extras.logging import get_logger
 from llmtuner.extras.misc import AverageMeter, count_parameters, get_logits_processor
-from llmtuner.tuner.ppo.utils import cast_layernorm_dtype, replace_model
+from llmtuner.tuner.ppo.utils import dump_layernorm, restore_layernorm, replace_model
 
 if TYPE_CHECKING:
     from transformers import Seq2SeqTrainingArguments, TrainerCallback
     from trl import AutoModelForCausalLMWithValueHead
-    from llmtuner.hparams import GeneratingArguments
+    from llmtuner.hparams import ModelArguments, GeneratingArguments
 
 
 logger = get_logger(__name__)
@@ -30,22 +31,27 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
 
     def __init__(
         self,
+        model_args: "ModelArguments",
         training_args: "Seq2SeqTrainingArguments",
         generating_args: "GeneratingArguments",
         callbacks: List["TrainerCallback"],
-        compute_dtype: torch.dtype,
         **kwargs
     ):
         PPOTrainer.__init__(self, **kwargs)
         if getattr(self.accelerator.state, "deepspeed_plugin", None) is not None:
             raise ValueError("PPOTrainer is incompatible with DeepSpeed.")
 
         self.args = training_args
-        self.generating_args = generating_args
-        self.log_callback, self.save_callback = callbacks[0], callbacks[1]
-        self.compute_dtype = compute_dtype
+        self.model_args = model_args
+        self.generation_config = GenerationConfig(
+            pad_token_id=self.tokenizer.pad_token_id,
+            eos_token_id=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
+            **generating_args.to_dict()
+        )
         self.state = TrainerState()
         self.control = TrainerControl()
+        self.log_callback, self.save_callback = callbacks[0], callbacks[1]
+        assert isinstance(self.log_callback, LogCallback) and isinstance(self.save_callback, SavePeftModelCallback)
 
     def ppo_train(self) -> None:
         r"""
@@ -74,13 +80,6 @@ def ppo_train(self) -> None:
             logger.info(f"  Total optimization steps = {max_steps}")
             logger.info(f"  Number of trainable parameters = {count_parameters(self.model)[0]}")
 
-        # Keyword arguments for `model.generate`
-        generating_args = self.generating_args.to_dict()
-        generating_args.update(dict(
-            eos_token_id=[self.tokenizer.eos_token_id] + self.tokenizer.additional_special_tokens_ids,
-            pad_token_id=self.tokenizer.pad_token_id
-        ))
-
         unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
         dataiter = iter(self.dataloader)
         steps_trained = 0
@@ -98,7 +97,7 @@ def ppo_train(self) -> None:
             self.model.eval()
 
             # Get inputs
-            queries, responses = self.get_inputs(batch, generating_args)
+            queries, responses = self.get_inputs(batch)
             self.tokenizer.padding_side = "right" # change padding side
             rewards = self.get_rewards(queries, responses, unwrapped_model)
 
@@ -159,27 +158,24 @@ def ppo_train(self) -> None:
         )
 
     @torch.no_grad()
-    def get_inputs(
-        self,
-        batch: Dict[str, torch.Tensor],
-        generating_args: Dict[str, Any]
-    ) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
+    def get_inputs(self, batch: Dict[str, torch.Tensor]) -> Tuple[List[torch.Tensor], List[torch.Tensor]]:
         r"""
         Generates model's responses given queries.
         """
-        gen_kwargs = dict(
-            generation_config=GenerationConfig(**generating_args),
+        if self.model_args.upcast_layernorm:
+            layernorm_params = dump_layernorm(self.model)
+
+        unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
+        response: torch.Tensor = unwrapped_model.generate(
+            generation_config=self.generation_config,
             logits_processor=get_logits_processor(),
             **batch
         )
 
-        input_ids = batch["input_ids"]
-        self.model, layer_norm_params = cast_layernorm_dtype(self.model, self.compute_dtype)
-        unwrapped_model: "AutoModelForCausalLMWithValueHead" = self.accelerator.unwrap_model(self.model)
-        response: torch.Tensor = unwrapped_model.generate(**gen_kwargs)
-        self.model, _ = cast_layernorm_dtype(self.model, self.compute_dtype, layer_norm_params)
-        query, response = input_ids.detach().cpu(), response[:, input_ids.size(-1):].detach().cpu()
+        if self.model_args.upcast_layernorm:
+            restore_layernorm(self.model, layernorm_params)
 
+        query, response = batch["input_ids"].detach().cpu(), response[:, batch["input_ids"].size(-1):].detach().cpu()
         queries, responses = [], []
         for i in range(len(query)):
             query_length = (query[i] != self.tokenizer.pad_token_id).nonzero()[0]