use the extracted field_messages to parse the role fields

cicd/cicd.sh

@@ -6,6 +6,5 @@ python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
 # pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
-pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+pytest -v --durations=10 --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

docs/config.qmd

@@ -244,8 +244,6 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
-# whether to concatenate samples during pretraining
-pretraining_sample_concatenation:
 
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:

requirements.txt

@@ -13,9 +13,9 @@ liger-kernel==0.5.2
 packaging==23.2
 
 peft==0.14.0
-transformers @ git+https://github.com/huggingface/transformers.git@mueller-trainer-refactor
+transformers==4.47.1
 tokenizers>=0.21.0
-accelerate==1.3.0
+accelerate==1.2.1
 datasets==3.2.0
 deepspeed==0.16.1
 trl==0.13.0

scripts/chat_datasets.py

@@ -30,7 +30,7 @@ def parse_dataset(dataset=None, split="train"):
         )
     ds_cfg["field_messages"] = field_messages
 
-    message_fields = features["conversations"][0].keys()
+    message_fields = features[field_messages][0].keys()
     message_field_role = None
     for key in ["from", "role"]:
         if key in message_fields:

src/axolotl/core/trainer_builder.py

@@ -1877,8 +1877,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
     ):
         if training_args.pretraining:
-            if self.cfg.pretraining_sample_concatenation is False:
-                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/monkeypatch/trainer_grad_accum.py

@@ -14,85 +14,15 @@ LOG = logging.getLogger("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, num_items_in_batch=num_items_in_batch)
-            loss = self.compute_loss(model, inputs)
-        else:
-            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
-
-    del inputs
-    if (
-        self.args.torch_empty_cache_steps is not None
-        and self.state.global_step % self.args.torch_empty_cache_steps == 0
-    ):
-        if is_torch_xpu_available():
-            torch.xpu.empty_cache()
-        elif is_torch_mlu_available():
-            torch.mlu.empty_cache()
-        elif is_torch_musa_available():
-            torch.musa.empty_cache()
-        elif is_torch_npu_available():
-            torch.npu.empty_cache()
-        elif is_torch_mps_available(min_version="2.0"):
-            torch.mps.empty_cache()
-        else:
-            torch.cuda.empty_cache()
-
-    kwargs = {}
-
-    # For LOMO optimizers you need to explicitly use the learnign rate
-    if self.args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
-        kwargs["learning_rate"] = self._get_learning_rate()
-
-    if self.args.n_gpu > 1:
-        loss = loss.mean()  # mean() to average on multi-gpu parallel training
-
-    if self.use_apex:
-        with amp.scale_loss(loss, self.optimizer) as scaled_loss:
-            scaled_loss.backward()
-    else:
-        # Finally we need to normalize the loss for reporting
-        if num_items_in_batch is None:
-            loss = loss / self.args.gradient_accumulation_steps
 """
 
 PATCHED_CONTEXT_CODE = """
     with self.compute_loss_context_manager():
-        loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
+        if self.model_accepts_loss_kwargs:
-
+            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
-    del inputs
-    if (
-        self.args.torch_empty_cache_steps is not None
-        and self.state.global_step % self.args.torch_empty_cache_steps == 0
-    ):
-        if is_torch_xpu_available():
-            torch.xpu.empty_cache()
-        elif is_torch_mlu_available():
-            torch.mlu.empty_cache()
-        elif is_torch_musa_available():
-            torch.musa.empty_cache()
-        elif is_torch_npu_available():
-            torch.npu.empty_cache()
-        elif is_torch_mps_available(min_version="2.0"):
-            torch.mps.empty_cache()
         else:
-            torch.cuda.empty_cache()
+            loss = self.compute_loss(model, inputs)
-
-    kwargs = {}
-
-    # For LOMO optimizers you need to explicitly use the learnign rate
-    if self.args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
-        kwargs["learning_rate"] = self._get_learning_rate()
-
-    if self.args.n_gpu > 1:
-        loss = loss.mean()  # mean() to average on multi-gpu parallel training
-
-    if self.use_apex:
-        with amp.scale_loss(loss, self.optimizer) as scaled_loss:
-            scaled_loss.backward()
-    else:
-        # Finally we need to normalize the loss for reporting
-        if not self.model_accepts_loss_kwargs and self.compute_loss_func is None:
-            loss = loss / self.args.gradient_accumulation_steps
 """
 
 ORIGINAL_LLAMA_FCLM_CODE = """

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -706,12 +706,6 @@ class AxolotlInputConfig(
     pad_to_sequence_len: Optional[bool] = None
     curriculum_sampling: Optional[bool] = None
     multipack_real_batches: Optional[bool] = None
-    pretraining_sample_concatenation: Optional[bool] = Field(
-        default=None,
-        json_schema_extra={
-            "description": "whether to soft pack/concatenate samples during pretraining",
-        },
-    )
 
     batch_flattening: Optional[Union[Literal["auto"], bool]] = None
 

src/axolotl/utils/data/pretraining.py

@@ -22,7 +22,6 @@ def encode_pretraining(
     max_tokens: int,
     examples: Dict[str, List],
     text_column: str = "text",
-    concatenate: bool = True,
 ) -> Dict[str, List]:
     res = tokenizer(
         examples[text_column],
@@ -34,13 +33,6 @@ def encode_pretraining(
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
     targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
-    if not concatenate:
-        return {
-            "input_ids": [seq.tolist() for seq in input_ids],
-            "labels": [seq.tolist() for seq in targets],
-            "attention_mask": [seq.tolist() for seq in attention_mask],
-        }
-
     new_input_ids = []
     new_labels = []
     new_attention_mask = []
@@ -212,7 +204,6 @@ def wrap_pretraining_dataset(
             tokenizer,
             max_tokens,
             text_column=cfg.pretraining_dataset[0].text_column or "text",
-            concatenate=cfg.pretraining_sample_concatenation is True,
         )
 
     if cfg.shuffle_merged_datasets:

src/axolotl/utils/models.py

@@ -386,15 +386,16 @@ class ModelLoader:
         if self.cfg.flash_attention:
             self.patch_attention()
 
-        # if self.cfg.model_config_type == "llama":
+        if self.cfg.model_config_type == "llama":
-        #     from axolotl.monkeypatch.trainer_grad_accum import (  # patch_forward_for_ga,
+            from axolotl.monkeypatch.trainer_grad_accum import (
-        #         patch_flash_attention_forward,
+                patch_flash_attention_forward,
-        #         patch_training_step_for_ga,
+                patch_forward_for_ga,
-        #     )
+                patch_training_step_for_ga,
-        #
+            )
-        #     patch_flash_attention_forward()
+
-        #     # patch_forward_for_ga()
+            patch_flash_attention_forward()
-        #     patch_training_step_for_ga()
+            patch_forward_for_ga()
+            patch_training_step_for_ga()
 
         if self.cfg.sample_packing and self.cfg.s2_attention:
             raise ValueError(

tests/e2e/patched/test_mixtral_samplepack.py

@@ -102,5 +102,9 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, dataset_meta=dataset_meta)
+        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+        assert (
+            "MixtralFlashAttention2"
+            in model.model.layers[0].self_attn.__class__.__name__
+        )
         check_model_output_exists(temp_dir, cfg)

tests/e2e/patched/test_model_patches.py

@@ -49,7 +49,12 @@ class TestModelPatches(unittest.TestCase):
         )
         normalize_config(cfg)
         tokenizer = load_tokenizer(cfg)
-        load_model(cfg, tokenizer, inference=False)
+        model, _ = load_model(cfg, tokenizer, inference=False)
+
+        assert (
+            "MixtralFlashAttention2"
+            in model.model.layers[0].self_attn.__class__.__name__
+        )
 
     @with_temp_dir
     def test_mistral_multipack(self, temp_dir):

tests/e2e/patched/test_unsloth_integration.py

@@ -3,6 +3,8 @@ import unittest
 
 import pytest
 
+from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
+
 
 @pytest.mark.skip(
     reason="Unsloth integration will be broken going into latest transformers"
@@ -11,8 +13,6 @@ class TestUnslothIntegration(unittest.TestCase):
     """Unsloth monkeypatch integration tests."""
 
     def test_is_self_attn_patchable(self):
-        from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
-
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(
             check_self_attn_is_patchable(),

tests/e2e/test_relora_llama.py

@@ -13,7 +13,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, check_tensorboard, with_temp_dir
+from .utils import check_model_output_exists, check_tensorboard, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"

tests/patched/test_llama_trainer_ga.py

@@ -1,8 +1,6 @@
 """"Test module for checking whether the Hugging Face Transformers is working as expected."""
 import unittest
 
-import pytest
-
 from axolotl.monkeypatch.trainer_grad_accum import (
     check_forward_is_patchable,
     check_training_step_is_patchable,
@@ -12,7 +10,6 @@ from axolotl.monkeypatch.trainer_grad_accum import (
 class TestTrainerGAIntegration(unittest.TestCase):
     """llama monkeypatch integration tests."""
 
-    @pytest.mark.skip("may not be needed for latest transformers version")
     def test_train_step_patchable(self):
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(
@@ -20,7 +17,6 @@ class TestTrainerGAIntegration(unittest.TestCase):
             "HF transformers Trainer.training_step has changed and isn't patchable",
         )
 
-    @pytest.mark.skip("may not be needed for latest transformers version")
     def test_model_forward_patchable(self):
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(