llama4 support (#2493)

* llama4 support

* add xet support [skip ci]

* be flexible on transformers version and skip test on version

* don't use deepspeed for the fix_untrained_tokens test

* reordering to trigger torch 2.6.0 tests first

* slightly smaller train set

* use 4.51.0 for now

* remove stray print, add llama4 chat template to schema, bump peft to 0.15.1

* patches to make llama4 performant

* add preliminary fp8 support

src/axolotl/core/trainers/base.py

@@ -562,6 +562,19 @@ class AxolotlTrainer(
 
         return res
 
+    def additional_accelerator_args(
+        self, fp8=None, **kwargs
+    ):  # pylint: disable=unused-argument
+        ret_kwargs = {}
+        if fp8:
+            from accelerate.utils import AORecipeKwargs
+
+            ret_kwargs["mixed_precision"] = "fp8"
+            ret_kwargs["kwargs_handlers"] = [AORecipeKwargs()]
+            os.environ["ACCELERATE_MIXED_PRECISION"] = "fp8"
+
+        return ret_kwargs
+
     def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
         """
         Log `logs` on the various objects watching training, including stored metrics.

src/axolotl/integrations/liger/__init__.py

@@ -173,5 +173,17 @@ class LigerPlugin(BasePlugin):
                 raise NotImplementedError(
                     "Fused linear cross entropy is not yet supported for Gemma3."
                 )
+        elif cfg.model_config_type == "llama4":
+            from axolotl.integrations.liger.models.llama4 import (
+                apply_liger_kernel_to_llama4,
+            )
+
+            apply_liger_kernel_to_llama4(
+                cross_entropy=cfg.liger_cross_entropy,
+                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
+                glu_activation=cfg.liger_glu_activation,
+                rms_norm=cfg.liger_rms_norm,
+                layer_norm=cfg.liger_layer_norm,
+            )
         elif cfg.model_config_type in ["deepseek_v3"]:
             raise ValueError(f"Unsupported model config type: {cfg.model_config_type}")

src/axolotl/integrations/liger/models/llama4.py

@@ -0,0 +1,171 @@
+"""
+Liger FLCE for llama4
+"""
+
+import sys
+from typing import List, Optional, Tuple, Union
+
+import torch
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[
+        Union["Cache", List[torch.FloatTensor]]  # noqa: F821
+    ] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+    """
+
+    # pylint: disable=duplicate-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,
+    )
+
+    hidden_states = outputs[0]
+
+    if hasattr(self.config, "pretraining_tp") and self.config.pretraining_tp > 1:
+        raise Exception(  # pylint: disable=broad-exception-raised
+            "Liger Kernel does not support pretraining_tp!!"
+        )
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
+
+    else:  # if in inference mode materialize logits
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def apply_liger_kernel_to_llama4(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = False,
+    rms_norm: bool = False,
+    glu_activation: bool = False,
+    layer_norm: bool = False,
+    **kwargs,  # pylint: disable=unused-argument
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Llama models (2 and 3)
+
+    Args:
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused linear cross entropy loss. Default is False.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be False.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is False.
+        glu_activation (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is False.
+    """
+
+    import transformers.models.llama4.modeling_llama4  # noqa: F401  # pylint: disable=unused-import
+    from liger_kernel.transformers.functional import liger_cross_entropy
+    from liger_kernel.transformers.layer_norm import LigerLayerNorm
+    from liger_kernel.transformers.rms_norm import LigerRMSNorm
+    from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
+
+    assert not (
+        cross_entropy and fused_linear_cross_entropy
+    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
+
+    modeling_llama4 = sys.modules["transformers.models.llama4.modeling_llama4"]
+
+    if rms_norm:
+        modeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
+    if glu_activation:
+        modeling_llama4.Llama4TextMLP = LigerSwiGLUMLP
+    if layer_norm:
+        modeling_llama4.nn.LayerNorm = LigerLayerNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_llama4.Llama4ForCausalLM.forward = lce_forward

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -162,7 +162,6 @@ def patch_flex_make_mask():
     for n in tuple(sys.modules):
         if ".modeling_" in n and "llama4" not in n:
             if hasattr(sys.modules[n], "make_flex_block_causal_mask"):
-                print(n)
                 sys.modules[n].make_flex_block_causal_mask = (
                     patched_make_flex_block_causal_mask
                 )

src/axolotl/monkeypatch/multipack.py

@@ -13,6 +13,7 @@ from axolotl.monkeypatch.utils import get_unpad_data
 SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "mllama_text_model",
     "llama",
+    "llama4",
     "mistral",
     "mixtral",
     "qwen2",

src/axolotl/monkeypatch/trainer_accelerator_args.py

@@ -0,0 +1,80 @@
+"""
+allow adding additional kwargs to Accelerator init
+"""
+
+import inspect
+import logging
+
+from transformers import Trainer
+
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_TRAINER_CODE = """
+    # create accelerator object
+    self.accelerator = Accelerator(**args)
+"""
+
+PATCHED_TRAINER_CODE = """
+    if hasattr(self, "additional_accelerator_args"):
+        additional_args = self.additional_accelerator_args(fp8=True, **args)
+        if additional_args:
+            args.update(additional_args)
+
+    # create accelerator object
+    self.accelerator = Accelerator(**args)
+"""
+
+
+def get_create_accelerate_code() -> str:
+    training_loop = inspect.getsource(Trainer.create_accelerator_and_postprocess)
+    return training_loop
+
+
+def check_create_accelerate_code_is_patchable() -> bool:
+    create_code = get_create_accelerate_code()
+    create_code, _ = detab_code(create_code)
+    return ORIGINAL_TRAINER_CODE in create_code
+
+
+def patch_create_accelerate_code_for_fp8():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        create_code = get_create_accelerate_code()
+    except OSError:
+        return
+    Trainer._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        create_code
+    )
+    create_code, _ = detab_code(create_code)
+    if ORIGINAL_TRAINER_CODE not in create_code:
+        return
+
+    create_code = create_code.replace(ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE)
+    create_code = create_code.replace(
+        "def create_accelerator_and_postprocess(",
+        "def fixed_create_accelerator_and_postprocess(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in create_code:
+            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(create_code, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching create_accelerator_and_postprocess to allow for overrides")
+    Trainer.create_accelerator_and_postprocess = fixed_create_accelerator_and_postprocess  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/utils/models.py

@@ -557,6 +557,14 @@ class ModelLoader:
         plugin_manager = PluginManager.get_instance()
         plugin_manager.pre_model_load(self.cfg)
 
+        # monkey patch to allow additional Accelerator init kwargs
+        if self.cfg.fp8:
+            from axolotl.monkeypatch.trainer_accelerator_args import (
+                patch_create_accelerate_code_for_fp8,
+            )
+
+            patch_create_accelerate_code_for_fp8()
+
         if self.cfg.adapter:
             from axolotl.monkeypatch.transformers_fa_utils import (
                 patch_fa_peft_integration,
@@ -988,10 +996,11 @@ class ModelLoader:
             )
             skip_move_to_device = True
         elif (
-            self.model_config.model_type == "llama"
+            self.model_config.model_type in ["llama", "llama4"]
             and not self.cfg.trust_remote_code
             and not self.cfg.gptq
         ):
+            # TODO do we need to open this up for all models?
             if self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading:
                 skip_move_to_device = True
                 if "device_map" in self.model_kwargs:

src/axolotl/utils/schemas/config.py

@@ -169,6 +169,7 @@ class AxolotlInputConfig(
 
     bf16: Literal["auto"] | bool | None = "auto"
     fp16: bool | None = None
+    fp8: bool | None = None
     bfloat16: bool | None = None  # for non-AMP cases
     float16: bool | None = None  # for non-AMP cases
     tf32: bool | None = None
@@ -464,9 +465,10 @@ class AxolotlInputConfig(
             data.get("sample_packing")
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
+            and not data.get("flex_attention")
         ):
             LOG.warning(
-                "sample_packing without flash_attention or sdp_attention does not handle cross-attention."
+                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
             )
 
         return data

src/axolotl/utils/schemas/enums.py

@@ -26,6 +26,7 @@ class ChatTemplate(str, Enum):
     gemma = "gemma"  # pylint: disable=invalid-name
     cohere = "cohere"  # pylint: disable=invalid-name
     llama3 = "llama3"  # pylint: disable=invalid-name
+    llama4 = "llama4"  # pylint: disable=invalid-name
     llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
     phi_3 = "phi_3"  # pylint: disable=invalid-name
     phi_35 = "phi_35"  # pylint: disable=invalid-name

src/axolotl/utils/trainer.py

@@ -582,7 +582,9 @@ def prepare_optim_env(cfg):
 
     setup_torch_compile_env(cfg)
 
-    if (cfg.bf16 == "auto" and is_torch_bf16_gpu_available()) or cfg.bf16 is True:
+    if cfg.fp8:
+        os.environ["ACCELERATE_MIXED_PRECISION"] = "fp8"
+    elif (cfg.bf16 == "auto" and is_torch_bf16_gpu_available()) or cfg.bf16 is True:
         os.environ["ACCELERATE_MIXED_PRECISION"] = "bf16"
     elif cfg.fp16:
         os.environ["ACCELERATE_MIXED_PRECISION"] = "fp16"