pylint

docs/config.qmd

@@ -407,10 +407,7 @@ save_total_limit: # Checkpoints saved at a time
 max_steps:
 
 # bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.
-include_tokens_per_second: # Optional[bool]
-
-# whether to find batch size that fits in memory. Passed to underlying transformers Trainer
-auto_find_batch_size: # Optional[bool]
+include_tokens_per_second:
 
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128

docs/lora_optims.qmd

@@ -12,7 +12,6 @@ to leverage operator fusion and tensor re-use in order to improve speed and redu
 memory usage during the forward and backward passes of these calculations.
 
 We currently support several common model architectures, including (but not limited to):
-
 - `llama`
 - `mistral`
 - `qwen2`

requirements.txt

@@ -13,12 +13,12 @@ liger-kernel==0.5.2
 packaging==23.2
 
 peft==0.14.0
-transformers==4.49.0
+transformers==4.48.3
 tokenizers>=0.21.0
 accelerate==1.3.0
 datasets==3.2.0
 deepspeed==0.16.1
-trl==0.15.1
+trl==0.15.0
 
 optimum==1.16.2
 hf_transfer

src/axolotl/core/trainer_builder.py

@@ -831,9 +831,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             if "max_length" in kwargs:
                 kwargs.pop("max_length")
         elif use_batch_sampler_collator:
-            if self.cfg.flex_attention is True:
-                collator = V2BatchSamplerDataCollatorForSeq2Seq
-            elif self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
+            if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
             elif (
                 self.cfg.model_config_type in ["llama"]

src/axolotl/core/trainers/grpo/trainer.py

@@ -78,6 +78,7 @@ class AxolotlGRPOTrainer(SchedulerMixin, GRPOTrainer):
             if is_peft_model(unwrapped_model):
                 unwrapped_model.merge_adapter()
                 state_dict = unwrapped_model.state_dict()
+                unwrapped_model.unmerge_adapter()
                 # Remove base_model and base_layer prefixes
                 state_dict = {
                     k.removeprefix("base_model.model.")
@@ -99,10 +100,8 @@ class AxolotlGRPOTrainer(SchedulerMixin, GRPOTrainer):
                 }
             else:
                 state_dict = unwrapped_model.state_dict()
-            if self.accelerator.is_main_process:
-                llm_model = (
-                    self.llm.llm_engine.model_executor.driver_worker.model_runner.model
-                )
-                llm_model.load_weights(state_dict.items())
-            if is_peft_model(unwrapped_model):
-                unwrapped_model.unmerge_adapter()
+        if self.accelerator.is_main_process:
+            llm_model = (
+                self.llm.llm_engine.model_executor.driver_worker.model_runner.model
+            )
+            llm_model.load_weights(state_dict.items())

src/axolotl/monkeypatch/lora_kernels.py

@@ -4,13 +4,12 @@ import importlib
 import inspect
 import logging
 import types
-from typing import Type
 
 import torch
 from accelerate.logging import get_logger
 from peft import PeftModelForCausalLM
 from torch import nn
-from transformers import AutoConfig
+from transformers.modeling_utils import PreTrainedModel
 
 from axolotl.kernels.lora import (
     apply_lora_mlp_geglu,
@@ -96,108 +95,90 @@ def original_apply_o(self: nn.Module, hidden_states: torch.Tensor) -> torch.Tens
     return attn_output
 
 
-def get_attention_cls_from_config(cfg: DictDefault) -> Type[nn.Module]:
-    """
-    Get the appropriate attention class by inspecting the model config.
-    Uses dynamic import to support any model architecture that follows
-    the standard transformers naming convention.
-
-    Args:
-        cfg: Dictionary mapping `axolotl` config keys to values.
-
-    Returns:
-        The appropriate attention class for the model.
-
-    Raises:
-        ValueError: If `base_model` not specified or attention class cannot be imported
-        ImportError: If the model module or attention class doesn't exist
-    """
-    if "base_model" not in cfg:
-        raise ValueError("base_model must be specified in config")
-
-    # Get model config without loading the model
-    model_config = AutoConfig.from_pretrained(cfg["base_model"])
-    model_type = model_config.model_type
-
-    # Special case for model_type = "qwen2"
-    if model_type == "qwen2":
-        from transformers.models.qwen2.modeling_qwen2 import Qwen2Attention
-
-        return Qwen2Attention
-
-    try:
-        # Dynamically import the module and attention class
-        module_path = f"transformers.models.{model_type}.modeling_{model_type}"
-        module = __import__(
-            module_path, fromlist=[f"{model_type.capitalize()}Attention"]
-        )
-        attention_cls = getattr(module, f"{model_type.capitalize()}Attention")
-
-        return attention_cls
-    except (ImportError, AttributeError) as e:
-        raise ValueError(
-            f"Could not import attention class for model_type: {model_type}. "
-            f"Error: {str(e)}"
-        ) from e
-
-
 # pylint: disable=protected-access
-def patch_self_attn_lora(cfg: DictDefault):
+def patch_self_attn_lora(model: PreTrainedModel):
     """
-    Given an `axolotl` config, this method patches the inferred attention class forward
-    pass with optimized LoRA implementations.
+    Patches the attention classes in a transformer model with optimized LoRA implementations.
 
     It modifies the attention class to use optimized QKV and output projections. The
     original implementation is preserved and can be restored if needed.
 
     Args:
-        cfg: Dictionary mapping `axolotl` config keys to values.
+        model: A HuggingFace transformers model.
 
     Raises:
         AssertionError: If the required code blocks are not found in the attention
             implementation.
     """
-    attention_cls = get_attention_cls_from_config(cfg)
+    # Find all attention modules in the model
+    attention_modules = [
+        module
+        for module in model.modules()
+        if "attention" in module.__class__.__name__.lower()
+        and hasattr(module, "forward")
+    ]
 
-    # Check if already patched
-    if hasattr(attention_cls, "_original_forward"):
-        LOG.info(f"{attention_cls.__name__} already patched")
+    if not attention_modules:
+        LOG.warning("No attention modules found in model")
         return
 
-    self_attn_forward = inspect.getsource(attention_cls.forward)
-    attention_cls._original_forward = self_attn_forward
-    self_attn_forward, _ = detab_code(self_attn_forward)
+    attention_classes = {type(module) for module in attention_modules}
+    LOG.info(f"Found attention classes: {[cls.__name__ for cls in attention_classes]}")
 
-    assert ORIGINAL_QKV_CODE in self_attn_forward, "Original QKV code not found"
-    assert ORIGINAL_O_CODE in self_attn_forward, "Original O code not found"
+    for attention_cls in attention_classes:
+        # Skip if already patched
+        if hasattr(attention_cls, "_original_forward"):
+            LOG.info(f"{attention_cls.__name__} already patched")
+            continue
 
-    self_attn_forward = self_attn_forward.replace(ORIGINAL_QKV_CODE, PATCHED_QKV_CODE)
-    self_attn_forward = self_attn_forward.replace(ORIGINAL_O_CODE, PATCHED_O_CODE)
-    self_attn_forward = self_attn_forward.replace(
-        "def forward(",
-        "def axolotl_attn_forward(",
-        1,
-    )
+        # Get and store original forward implementation
+        self_attn_forward = inspect.getsource(attention_cls.forward)
+        attention_cls._original_forward = self_attn_forward
 
-    # Load necessary imports
-    module_name = attention_cls.__module__
-    module = importlib.import_module(module_name)
+        # Remove indentation
+        self_attn_forward, _ = detab_code(self_attn_forward)
 
-    items_to_import = []
-    for item in dir(module):
-        if item in self_attn_forward:
-            items_to_import.append(item)
+        # Verify required code blocks exist
+        assert (
+            ORIGINAL_QKV_CODE in self_attn_forward
+        ), f"Original QKV code not found in {attention_cls.__name__}"
+        assert (
+            ORIGINAL_O_CODE in self_attn_forward
+        ), f"Original O code not found in {attention_cls.__name__}"
 
-    exec(  # pylint: disable=exec-used  # nosec B102
-        f"from {module_name} import ({', '.join(items_to_import)})",
-        globals(),
-    )
-    exec(self_attn_forward, globals())  # pylint: disable=exec-used  # nosec B102
+        # Replace code blocks
+        self_attn_forward = self_attn_forward.replace(
+            ORIGINAL_QKV_CODE, PATCHED_QKV_CODE
+        )
+        self_attn_forward = self_attn_forward.replace(ORIGINAL_O_CODE, PATCHED_O_CODE)
+        self_attn_forward = self_attn_forward.replace(
+            "def forward(",
+            "def axolotl_attn_forward(",
+            1,
+        )
 
-    LOG.info(f"Patched attention class with LoRA optims: {attention_cls.__name__}")
-    attention_cls.forward = (
-        axolotl_attn_forward  # pylint: disable=undefined-variable  # noqa: F821
-    )
+        # Import necessary symbols from the attention module
+        module_name = attention_cls.__module__
+        module = importlib.import_module(module_name)
+
+        items_to_import = []
+        for item in dir(module):
+            if item in self_attn_forward:
+                items_to_import.append(item)
+
+        if items_to_import:
+            exec(  # pylint: disable=exec-used  # nosec B102
+                f"from {module_name} import ({', '.join(items_to_import)})",
+                globals(),
+            )
+
+        # Execute the new implementation
+        exec(self_attn_forward, globals())  # pylint: disable=exec-used  # nosec B102
+
+        LOG.info(f"Patched attention class with LoRA optims: {attention_cls.__name__}")
+        attention_cls.forward = (
+            axolotl_attn_forward  # pylint: disable=undefined-variable  # noqa: F821
+        )
 
 
 def apply_lora_kernel_patches(

src/axolotl/monkeypatch/relora.py

@@ -127,8 +127,6 @@ class ReLoRACallback(TrainerCallback):
         optimizer: torch.optim.Optimizer,
         **_kwargs,
     ):
-        if not optimizer:
-            optimizer = state.optimizer
         if state.global_step > 0 and state.global_step % self.relora_steps == 0:
             checkpoint_folder = os.path.join(
                 args.output_dir,

src/axolotl/monkeypatch/utils.py

@@ -95,103 +95,6 @@ def get_cu_seqlens(attn_mask):
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
 
 
-def get_packed_mask_from_pos_ids(position_ids):
-    if len(position_ids.shape) == 1:
-        position_ids = position_ids.unsqueeze(0)
-
-    device = position_ids.device
-    results = []
-
-    for i, row in enumerate(position_ids):
-        # Count the number of consecutive zeros from the right side
-        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
-
-        # Adjust the row to exclude padding
-        adjusted_row = row[:-padding_length] if padding_length else row.clone()
-
-        # Find where the position resets to 0 (indicating a new sequence)
-        seq_starts = torch.cat(
-            [
-                torch.tensor([True], dtype=torch.bool, device=device),
-                adjusted_row[1:] == 0,
-            ]
-        )
-        # Get the indices where the sequence starts
-        start_indices = torch.cat(
-            [
-                torch.nonzero(seq_starts).unbind(dim=1)[0],
-                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
-            ]
-        )
-        # Calculate the sequence lengths
-        seq_lengths = start_indices[1:] - start_indices[:-1]
-        # Append the padding length to the sequence lengths
-        doc_mask = torch.ones(len(row), dtype=torch.int32, device=device)
-        for i, seq_len in enumerate(seq_lengths):
-            start_id = start_indices[i]
-            doc_mask[start_id : start_id + seq_len] = (
-                (i+1) * doc_mask[start_id : start_id + seq_len]
-            )
-        if padding_length:
-            doc_mask[len(adjusted_row) :] = 0 * doc_mask[len(adjusted_row) :]
-
-        results.append(doc_mask)
-
-    return torch.stack(results)
-
-
-def get_seqlens_from_pos_ids(position_ids):
-    """generate a sequence length set using pos ids for doc mask creation in flex attention"""
-    if len(position_ids.shape) == 1:
-        position_ids = position_ids.unsqueeze(0)
-    max_seq_len = position_ids.shape[1]
-
-    device = position_ids.device
-    results = []
-    totalseqlens = []
-
-    for row in position_ids:
-        # Count the number of consecutive zeros from the right side
-        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
-
-        # Adjust the row to exclude padding
-        adjusted_row = row[:-padding_length] if padding_length else row.clone()
-
-        # Find where the position resets to 0 (indicating a new sequence)
-        seq_starts = torch.cat(
-            [
-                torch.tensor([True], dtype=torch.bool, device=device),
-                adjusted_row[1:] == 0,
-            ]
-        )
-        # Get the indices where the sequence starts
-        start_indices = torch.cat(
-            [
-                torch.nonzero(seq_starts).unbind(dim=1)[0],
-                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
-            ]
-        )
-        # Calculate the sequence lengths
-        seq_lengths = start_indices[1:] - start_indices[:-1]
-        # Append the padding length to the sequence lengths
-        if padding_length:
-            seq_lengths = torch.cat(
-                [
-                    seq_lengths,
-                    torch.tensor(
-                        [len(row) - torch.sum(seq_lengths)],
-                        dtype=torch.int32,
-                        device=device,
-                    ),
-                ]
-            )
-
-        results.append(seq_lengths)
-        totalseqlens.append(len(adjusted_row))
-
-    return results, torch.tensor(totalseqlens, dtype=torch.int32, device=device)
-
-
 def get_cu_seqlens_from_pos_ids(position_ids):
     """generate a cumulative sequence length mask for flash attention using pos ids"""
     if len(position_ids.shape) == 1:
@@ -273,10 +176,7 @@ def mask_2d_to_4d(
     when they attend to each other within that sequence.
     This expansion transforms the mask to lower triangular form to prevent future peeking.
     """
-
-    if len(mask.size()) == 4:
-        return mask
-    bsz, src_len = int(mask.size()[0]), int(mask.size()[1])
+    bsz, src_len = mask.size()
     tgt_len = tgt_len if tgt_len is not None else src_len
 
     mask = mask.unsqueeze(1).unsqueeze(2)

src/axolotl/prompt_strategies/chat_template.py

@@ -272,7 +272,8 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
                 dict(zip(feature_names, row))
             )
             for key, val in tokenized_prompt.items():
-                res[key].append(val)
+                for i in range(0, len(val), self.sequence_len):
+                    res[key].append(val[i : i + self.sequence_len])
 
         # If there are no examples left, return an empty dictionary
         if not res:

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

@@ -342,7 +342,6 @@ class LoraConfig(BaseModel):
     peft_use_dora: Optional[bool] = None
     peft_use_rslora: Optional[bool] = None
     peft_layer_replication: Optional[List[Tuple[int, int]]] = None
-    peft_init_lora_weights: Optional[Union[bool, str]] = None
 
     qlora_sharded_model_loading: Optional[bool] = Field(
         default=False,
@@ -823,7 +822,6 @@ class AxolotlInputConfig(
     xformers_attention: Optional[bool] = None
     sdp_attention: Optional[bool] = None
     s2_attention: Optional[bool] = None
-    flex_attention: Optional[bool] = None
     flash_attention: Optional[bool] = None
     flash_attn_cross_entropy: Optional[bool] = None
     flash_attn_rms_norm: Optional[bool] = None
@@ -1790,26 +1788,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                 )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_flex_torch_version(cls, data):
-        if (data.get("flex_attention") is not None) and (
-            data.get("flex_attention") is True
-        ):
-            env_capabilities = data.get("env_capabilities", {})
-            torch_version = env_capabilities.get("torch_version")
-
-            if torch_version is None:
-                import torch
-
-                torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
-
-            if version.parse(torch_version) < version.parse("2.5.1"):
-                raise ValueError(
-                    "Flex attention is not supported on torch version < 2.5.1"
-                )
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_torch_compile_auto(cls, data):

src/axolotl/utils/data/utils.py

@@ -172,11 +172,10 @@ def drop_long_seq_in_dataset(dataset: Dataset, cfg: DictDefault):
     )
 
     try:
-        ds_lengths = get_dataset_lengths(dataset, from_arrow=True)
-        min_input_len = np.min(ds_lengths)
-        LOG.info(f"min_input_len: {min_input_len}")
-        max_input_len = np.max(ds_lengths)
-        LOG.info(f"max_input_len: {max_input_len}")
+        min_input_len = np.min(get_dataset_lengths(dataset))
+        LOG.debug(f"min_input_len: {min_input_len}")
+        max_input_len = np.max(get_dataset_lengths(dataset))
+        LOG.debug(f"max_input_len: {max_input_len}")
     except AttributeError:
         pass
 

src/axolotl/utils/models.py

@@ -403,7 +403,7 @@ class ModelLoader:
 
         if (
             self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
-            and (self.cfg.flash_attention or self.cfg.flex_attention)
+            and self.cfg.flash_attention
             and self.cfg.sample_packing
         ):
             if "auto_map" in self.model_config:
@@ -439,11 +439,6 @@ class ModelLoader:
 
             patch_mistral_cross_entropy()
 
-        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
-            from axolotl.monkeypatch.lora_kernels import patch_self_attn_lora
-
-            patch_self_attn_lora(self.cfg)
-
     def patch_attention(self) -> None:
         if hasattr(self.model_config, "model_type"):
             if self.model_config.model_type == "mllama" and self.cfg.flash_attention:
@@ -707,13 +702,7 @@ class ModelLoader:
         """
         sample packing uses custom FA2 patch
         """
-
-        if self.cfg.flex_attention:
-            self.model_kwargs["attn_implementation"] = "flex_attention"
-            self.model_config._attn_implementation = (  # pylint: disable=protected-access
-                "flex_attention"
-            )
-        elif self.cfg.flash_attention:
+        if self.cfg.flash_attention:
             if not self.cfg.sample_packing and self.cfg.s2_attention:
                 pass
             self.model_kwargs["attn_implementation"] = "flash_attention_2"
@@ -1034,6 +1023,12 @@ class ModelLoader:
             integrate_rope_embeddings()
 
     def apply_lora_patch(self) -> None:
+        """Applies patching relevant to LoRA Triton kernels if enabled."""
+        if self.cfg.lora_qkv_kernel or self.cfg.lora_o_kernel:
+            from axolotl.monkeypatch.lora_kernels import patch_self_attn_lora
+
+            patch_self_attn_lora(self.model)
+
         if (
             self.cfg.lora_mlp_kernel
             or self.cfg.lora_qkv_kernel
@@ -1119,7 +1114,7 @@ class ModelLoader:
         should_convert = (
             # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
             # convert them back to fp16/bf16 for flash-attn compatibility.
-            ((needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention) and not qlora_fsdp)
+            ((needs_fa2_dtype or self.cfg.flash_attention) and not qlora_fsdp)
             or self.cfg.cut_cross_entropy  # Cut cross entropy requires embedding layers to be in fp16/bf16 for backward pass
         )
 
@@ -1187,6 +1182,7 @@ class ModelLoader:
         if self.cfg.adapter is not None:
             log_gpu_memory_usage(LOG, "after adapters", self.model.device)
 
+        # TODO: Deprecate this.
         self.apply_unsloth_lora_patch()
         self.apply_lora_patch()
 
@@ -1207,9 +1203,7 @@ def load_model(
     reference_model: bool = False,
     **kwargs,  # pylint: disable=unused-argument
 ) -> Tuple[PreTrainedModel, Optional[PeftConfig]]:
-    """
-    Load a model for a given configuration and tokenizer.
-    """
+    """Load a model for a given configuration and tokenizer."""
     loader = ModelLoader(
         cfg,
         tokenizer,
@@ -1327,8 +1321,6 @@ def load_lora(model, cfg, inference=False, config_only=False):
     if loftq_bits:
         lora_config_kwargs["loftq_config"] = LoftQConfig(loftq_bits=loftq_bits)
         lora_config_kwargs["init_lora_weights"] = "loftq"
-    if cfg.peft_init_lora_weights:
-        lora_config_kwargs["init_lora_weights"] = cfg.peft_init_lora_weights
     if cfg.peft_use_dora:
         lora_config_kwargs["use_dora"] = cfg.peft_use_dora
         LOG.info("Initializing LoRA weights using dora. This might take longer.")

src/axolotl/utils/samplers/utils.py

@@ -4,17 +4,13 @@ helper util to calculate dataset lengths
 import numpy as np
 
 
-def get_dataset_lengths(dataset, from_arrow=False):
-    if "length" in dataset.column_names:
-        lengths = np.array(dataset["length"])
-    elif "position_ids" in dataset.column_names:
-        position_ids = dataset["position_ids"]
+def get_dataset_lengths(dataset):
+    if "length" in dataset.data.column_names:
+        lengths = np.array(dataset.data.column("length"))
+    elif "position_ids" in dataset.data.column_names:
+        position_ids = dataset.data.column("position_ids")
         lengths = np.array([x[-1] + 1 for x in position_ids])
     else:
-        if from_arrow:
-            input_ids = dataset.data.column("input_ids")
-            lengths = np.vectorize(len)(np.array(input_ids, dtype=object))
-        else:
-            input_ids = dataset["input_ids"]
-            lengths = np.array([len(seq) for seq in input_ids])
+        input_ids = dataset.data.column("input_ids")
+        lengths = np.vectorize(len)(np.array(input_ids, dtype=object))
     return lengths

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -9,16 +9,14 @@ from transformers import AutoModelForCausalLM, LlamaForCausalLM
 from transformers.models.llama.configuration_llama import LlamaConfig
 from transformers.models.llama.modeling_llama import LlamaAttention
 
+from axolotl.cli.utils import load_model_and_tokenizer
 from axolotl.kernels.lora import (
     apply_lora_mlp_geglu,
     apply_lora_mlp_swiglu,
     apply_lora_o,
     apply_lora_qkv,
 )
-from axolotl.monkeypatch.lora_kernels import (
-    apply_lora_kernel_patches,
-    patch_self_attn_lora,
-)
+from axolotl.monkeypatch.lora_kernels import apply_lora_kernel_patches
 from axolotl.utils.dict import DictDefault
 
 MODEL_CONFIGS = [
@@ -65,15 +63,45 @@ def small_llama_model():
     return LlamaForCausalLM(LlamaConfig(**config))
 
 
-def test_attention_patching_integration():
-    """Test attention patching in integration context."""
-    cfg = {"base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0"}
+# pylint: disable=duplicate-code
+@pytest.fixture
+def minimal_cfg():
+    "Config of real HuggingFace Hub model"
+    cfg = DictDefault(
+        {
+            "base_model": "HuggingFaceTB/SmolLM2-135M",
+            "tokenizer_config": "HuggingFaceTB/SmolLM2-135M",
+            "learning_rate": 0.000001,
+            "datasets": [
+                {
+                    "path": "mhenrichsen/alpaca_2k_test",
+                    "type": "alpaca",
+                }
+            ],
+            "micro_batch_size": 1,
+            "gradient_accumulation_steps": 1,
+            "adapter": "lora",
+            "lora_r": 8,
+            "lora_alpha": 16,
+            "lora_dropout": 0.0,
+            "lora_target_linear": True,
+            "sequence_len": 1024,
+            "lora_mlp_kernel": True,
+            "lora_qkv_kernel": True,
+            "lora_o_kernel": True,
+        }
+    )
 
+    return cfg
+
+
+def test_attention_patching_integration(minimal_cfg):
+    """Test attention patching in integration context."""
     # Store the original implementation
     original_forward = getattr(LlamaAttention, "forward")
 
-    # Apply patch
-    patch_self_attn_lora(cfg)
+    # Load model
+    _, _ = load_model_and_tokenizer(cfg=minimal_cfg)
 
     # Get the new forward method
     patched_forward = LlamaAttention.forward
@@ -376,38 +404,10 @@ def test_model_architecture(model_config):
 
 
 # pylint: disable=duplicate-code
-def test_kernel_training_integration():
+def test_kernel_training_integration(minimal_cfg):
     """Test model loading with kernel patches enabled."""
-    from axolotl.cli.utils import load_model_and_tokenizer
-
-    # Create minimal config
-    cfg = DictDefault(
-        {
-            "base_model": "HuggingFaceTB/SmolLM2-135M",
-            "tokenizer_config": "HuggingFaceTB/SmolLM2-135M",
-            "learning_rate": 0.000001,
-            "datasets": [
-                {
-                    "path": "mhenrichsen/alpaca_2k_test",
-                    "type": "alpaca",
-                }
-            ],
-            "micro_batch_size": 1,
-            "gradient_accumulation_steps": 1,
-            "adapter": "lora",
-            "lora_r": 8,
-            "lora_alpha": 16,
-            "lora_dropout": 0.0,
-            "lora_target_linear": True,
-            "sequence_len": 1024,
-            "lora_mlp_kernel": True,
-            "lora_qkv_kernel": True,
-            "lora_o_kernel": True,
-        }
-    )
-
     # Load model
-    model, _ = load_model_and_tokenizer(cfg=cfg)
+    model, _ = load_model_and_tokenizer(cfg=minimal_cfg)
 
     # Verify correct activation function
     layer = model.model.model.layers[0]

tests/prompt_strategies/conftest.py

@@ -125,12 +125,6 @@ def fixture_llama3_tokenizer():
     return tokenizer
 
 
-@pytest.fixture(name="smollm2_tokenizer", scope="session", autouse=True)
-def fixture_smollm2_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM2-135M")
-    return tokenizer
-
-
 @pytest.fixture(name="mistralv03_tokenizer", scope="session", autouse=True)
 def fixture_mistralv03_tokenizer():
     tokenizer = AutoTokenizer.from_pretrained(

tests/prompt_strategies/test_dpo_chatml.py

@@ -1,61 +0,0 @@
-"""
-Tests for loading DPO preference datasets with chatml formatting
-"""
-import unittest
-
-import pytest
-
-from axolotl.prompt_strategies.dpo import load as load_dpo
-from axolotl.utils.data.rl import load_prepare_preference_datasets
-from axolotl.utils.dict import DictDefault
-
-
-@pytest.fixture(name="minimal_dpo_cfg")
-def fixture_cfg():
-    return DictDefault(
-        {
-            "base_model": "HuggingFaceTB/SmolLM2-135M",
-            "tokenizer_config": "HuggingFaceTB/SmolLM2-135M",
-            "rl": "dpo",
-            "learning_rate": 0.000001,
-            "micro_batch_size": 1,
-            "gradient_accumulation_steps": 1,
-            "special_tokens": {
-                "pad_token": "<|endoftext|>",
-            },
-            "sequence_len": 2048,
-        }
-    )
-
-
-class TestDPOChatml:
-    """
-    Test loading DPO preference datasets with chatml formatting
-    """
-
-    def test_default(self, minimal_dpo_cfg):
-        cfg = DictDefault(
-            {
-                "datasets": [
-                    {
-                        "path": "argilla/distilabel-intel-orca-dpo-pairs",
-                        "type": "chatml",
-                        "split": "train[:1%]",
-                    }
-                ]
-            }
-            | minimal_dpo_cfg
-        )
-
-        # test that dpo.load works
-        load_dpo("chatml", cfg)
-        # now actually load the datasets with the strategy
-        train_ds, _ = load_prepare_preference_datasets(cfg)
-        assert train_ds[0]["prompt"].startswith("<|im_start|>")
-        assert train_ds[0]["prompt"].endswith("<|im_start|>assistant\n")
-        assert "chosen" in train_ds[0]
-        assert "rejected" in train_ds[0]
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/test_packed_batch_sampler.py

@@ -7,7 +7,6 @@ from transformers import AutoTokenizer
 from axolotl.datasets import TokenizedPromptDataset
 from axolotl.prompt_strategies.completion import load
 from axolotl.utils.collators import V2BatchSamplerDataCollatorForSeq2Seq
-from axolotl.utils.data.utils import drop_long_seq_in_dataset
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
@@ -19,6 +18,11 @@ def fixture_tokenizer():
     return tokenizer
 
 
+@pytest.fixture(name="max_seq_length")
+def fixture_max_seq_length():
+    return 4096
+
+
 class TestBatchedSamplerPacking:
     """
     Test class for packing streaming dataset sequences
@@ -33,7 +37,6 @@ class TestBatchedSamplerPacking:
             (2, 2),
         ],
     )
-    @pytest.mark.parametrize("max_seq_length", [4096, 512])
     def test_packing(self, batch_size, num_workers, tokenizer, max_seq_length):
         import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 
@@ -59,9 +62,6 @@ class TestBatchedSamplerPacking:
             dataset,
         )
         train_dataset = concatenate_datasets([dataset_wrapper])
-
-        train_dataset = drop_long_seq_in_dataset(train_dataset, cfg)
-
         lengths = get_dataset_lengths(train_dataset)
         batch_sampler = MultipackBatchSampler(
             sampler=RandomSampler(train_dataset),
@@ -90,7 +90,7 @@ class TestBatchedSamplerPacking:
                 batch_idxs.extend(pack)
 
         for batch in loader:
-            assert batch["input_ids"].numel() <= batch_size * max_seq_length
+            assert len(batch["input_ids"]) <= batch_size * max_seq_length
             assert batch["input_ids"].shape[1] == max_seq_length
 
         original_idxs = set(range(len(train_dataset)))