Transformers 4.47.0 (axolotl-ai-cloud#2138)

* bump transformers and trl

* fix: update trainer.log signature

* fix trl trainer.log interfaces

* broken 🦥 with latest transformers

* skip parent, call grandparent - yeah, super janky

* update HF HUB env var and fix reward trainer log since it doesn't directly override log

* also bump accelerate

* patches for llama ga

* detab the code to check

* fix whitespace for patch check

* play nicely with CI tests since we patch everytime

* fix pop default in case it doesn't exist

* more tweaks to make patches nicer in CI

* fix detab for when there are possibly multiple patches

---------

Co-authored-by: NanoCode012 <[email protected]>

docker/Dockerfile-cloud

@@ -2,7 +2,7 @@ ARG BASE_TAG=main
 FROM axolotlai/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
-ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV HF_HOME="/workspace/data/huggingface-cache/hub"
 ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 

docker/Dockerfile-cloud-no-tmux

@@ -2,7 +2,7 @@ ARG BASE_TAG=main
 FROM axolotlai/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
-ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV HF_HOME="/workspace/data/huggingface-cache/hub"
 ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 

requirements.txt

@@ -1,10 +1,10 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.14.0
-transformers==4.46.3
+transformers==4.47.0
 tokenizers>=0.20.1
 bitsandbytes==0.45.0
-accelerate==1.1.0
+accelerate==1.2.0
 datasets==3.1.0
 deepspeed==0.15.4
 pydantic==2.6.3
@@ -42,7 +42,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl==0.12.0
+trl==0.12.1
 zstandard==0.22.0
 fastcore
 

src/axolotl/cli/__init__.py

@@ -442,7 +442,7 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
             "compute_capability": gpu_version,
         },
         env_capabilities={
-            "torch_version": str(torch.__version__).split("+", maxsplit=1)[0]
+            "torch_version": str(torch.__version__).split("+", maxsplit=1)[0],
         },
     )
 

src/axolotl/core/trainer_builder.py

@@ -957,21 +957,23 @@ def create_accelerator_and_postprocess(self):
 
         return res
 
-    def log(self, logs: Dict[str, float]) -> None:
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
         """
         Log `logs` on the various objects watching training, including stored metrics.
 
         Args:
             logs (`Dict[str, float]`):
                 The values to log.
+            start_time (`Optional[float]`):
+                The start of training.
         """
         # logs either has 'loss' or 'eval_loss'
         train_eval = "train" if "loss" in logs else "eval"
         # Add averaged stored metrics to logs
         for key, metrics in self._stored_metrics[train_eval].items():
             logs[key] = torch.tensor(metrics).mean().item()
         del self._stored_metrics[train_eval]
-        return super().log(logs)
+        return super().log(logs, start_time)
 
     def store_metrics(
         self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
@@ -1155,6 +1157,18 @@ def training_step(
         torch.cuda.empty_cache()
         return loss
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super(DPOTrainer, self).log(  # pylint: disable=bad-super-call
+            logs, start_time
+        )
+
 
 class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
     """
@@ -1163,6 +1177,18 @@ class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
 
     tag_names = ["axolotl", "orpo"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super(ORPOTrainer, self).log(  # pylint: disable=bad-super-call
+            logs, start_time
+        )
+
 
 class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
     """
@@ -1171,6 +1197,45 @@ class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
 
     tag_names = ["axolotl", "kto"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # train metrics should have no prefix, eval should have 'eval_'
+        prefix = "eval_" if train_eval == "eval" else ""
+        # accumulate average metrics from sums and lengths
+        for split in ["chosen", "rejected"]:
+            if f"count/{split}" in self._stored_metrics[train_eval]:
+                count_sum = (
+                    torch.Tensor(self._stored_metrics[train_eval][f"count/{split}"])
+                    .sum()
+                    .item()
+                )
+                for metric in ["rewards", "logps", "logits"]:
+                    logs[f"{prefix}{metric}/{split}"] = (
+                        torch.Tensor(
+                            self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
+                        )
+                        .sum()
+                        .item()
+                        / count_sum
+                    )
+                    # delete obsolete metric
+                    del self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
+                del self._stored_metrics[train_eval][f"count/{split}"]
+        # calculate reward margin
+        if f"{prefix}rewards/chosen" in logs and f"{prefix}rewards/rejected" in logs:
+            logs[f"{prefix}rewards/margins"] = (
+                logs[f"{prefix}rewards/chosen"] - logs[f"{prefix}rewards/rejected"]
+            )
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[f"{prefix}{key}"] = torch.Tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super(KTOTrainer, self).log(  # pylint: disable=bad-super-call
+            logs, start_time
+        )
+
 
 class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
     """
@@ -1179,6 +1244,18 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
 
     tag_names = ["axolotl", "cpo"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super(CPOTrainer, self).log(  # pylint: disable=bad-super-call
+            logs, start_time
+        )
+
 
 class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
     """
@@ -1187,6 +1264,12 @@ class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
 
     tag_names = ["axolotl", "reward"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        return super(RewardTrainer, self).log(  # pylint: disable=bad-super-call
+            logs, start_time
+        )
+
 
 class TrainerBuilderBase(abc.ABC):
     """

src/axolotl/monkeypatch/trainer_grad_accum.py

@@ -0,0 +1,207 @@
+"""
+fix for FSDP gradient accumulation
+see https://github.com/huggingface/transformers/pull/35128
+"""
+import inspect
+
+from accelerate.logging import get_logger
+from transformers import LlamaForCausalLM
+from transformers.trainer import Trainer
+
+from axolotl.monkeypatch.unsloth_ import detab_code
+
+LOG = get_logger("axolotl.monkeypatch.trainer_grad_accum")
+
+ORIGINAL_CONTEXT_CODE = """
+    with self.compute_loss_context_manager():
+        if self.model_accepts_loss_kwargs:
+            loss = self.compute_loss(model, inputs)
+        else:
+            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
+"""
+
+PATCHED_CONTEXT_CODE = """
+    with self.compute_loss_context_manager():
+        if self.model_accepts_loss_kwargs:
+            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
+        else:
+            loss = self.compute_loss(model, inputs)
+"""
+
+ORIGINAL_LLAMA_FCLM_CODE = """
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+    loss = None
+    if labels is not None:
+        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+"""
+
+PATCHED_LLAMA_FCLM_CODE = """
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # remove num_items_in_batch otherwise self.model attempts to pass it to flash_attention
+    num_items_in_batch = kwargs.pop("num_items_in_batch", None)
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+    loss = None
+    if labels is not None:
+        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, num_items_in_batch=num_items_in_batch, **kwargs)
+"""
+
+
+def get_training_step_code() -> str:
+    training_step = inspect.getsource(
+        Trainer.training_step  # pylint: disable=protected-access
+    )
+    return training_step
+
+
+def check_training_step_is_patchable() -> bool:
+    training_step = get_training_step_code()
+    training_step, _ = detab_code(training_step)
+    return ORIGINAL_CONTEXT_CODE in training_step
+
+
+def patch_training_step_for_ga():
+    """
+    monkeypatch for fixing the training loop for gradient accumulation
+    """
+
+    try:
+        training_step = get_training_step_code()
+    except OSError:
+        return
+    Trainer._original_training_step = training_step  # pylint: disable=protected-access
+    training_step, _ = detab_code(training_step)
+    if ORIGINAL_CONTEXT_CODE not in training_step:
+        return
+    # assert (
+    #     ORIGINAL_CONTEXT_CODE in training_step
+    # ), "Original training_step code not found"
+
+    training_step = training_step.replace(ORIGINAL_CONTEXT_CODE, PATCHED_CONTEXT_CODE)
+    training_step = training_step.replace(
+        "def training_step(",
+        "def _fixed_training_step(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in training_step:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(training_step, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching training_step", main_process_only=True)
+    Trainer.training_step = (  # pylint: disable=protected-access
+        _fixed_training_step  # pylint: disable=undefined-variable  # noqa: F821
+    )
+
+
+def get_model_forward_code() -> str:
+    forward = inspect.getsource(
+        LlamaForCausalLM.forward  # pylint: disable=protected-access
+    )
+    return forward
+
+
+def check_forward_is_patchable() -> bool:
+    forward = get_model_forward_code()
+    forward, _ = detab_code(forward)
+    return ORIGINAL_LLAMA_FCLM_CODE in forward
+
+
+def patch_forward_for_ga():
+    """
+    monkeypatch for fixing the training loop for gradient accumulation
+    """
+
+    try:
+        forward = get_model_forward_code()
+    except OSError:
+        return
+    LlamaForCausalLM._original_forward = forward  # pylint: disable=protected-access
+    forward, _ = detab_code(forward)
+    if ORIGINAL_LLAMA_FCLM_CODE not in forward:
+        return
+    # assert ORIGINAL_LLAMA_FCLM_CODE in forward, "Original forward code not found"
+
+    forward = forward.replace(ORIGINAL_LLAMA_FCLM_CODE, PATCHED_LLAMA_FCLM_CODE)
+    forward = forward.replace(
+        "def forward(",
+        "def _fixed_forward(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.models.llama.modeling_llama
+
+    items_to_import = []
+    for item in dir(transformers.models.llama.modeling_llama):
+        if item in forward:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.models.llama.modeling_llama import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(forward, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching forward", main_process_only=True)
+    LlamaForCausalLM.forward = (  # pylint: disable=protected-access
+        _fixed_forward  # pylint: disable=undefined-variable  # noqa: F821
+    )