liger-kernel-nightly 0.5.10.dev20250624183504__py3-none-any.whl → 0.5.10.dev20250629005644__py3-none-any.whl

liger_kernel/transformers/__init__.py

@@ -30,6 +30,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_granite  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mistral  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
@@ -87,6 +88,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
+        "apply_liger_kernel_to_llama4",
         "apply_liger_kernel_to_mistral",
         "apply_liger_kernel_to_mixtral",
         "apply_liger_kernel_to_mllama",
@@ -141,6 +143,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",
+            "apply_liger_kernel_to_llama4",
             "apply_liger_kernel_to_mistral",
             "apply_liger_kernel_to_mixtral",
             "apply_liger_kernel_to_mllama",

liger_kernel/transformers/model/llama4.py

@@ -0,0 +1,108 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+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]]] = 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,
+    **kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    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]`.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Llama4ForCausalLM
+
+    >>> model = Llama4ForCausalLM.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama4/Llama4-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    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=True,
+        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
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+
+    if self.training and (labels is not None or shift_labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+
+    else:  # if in inference mode materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -363,6 +363,92 @@ def apply_liger_kernel_to_llava(
             logger.warning(f"{vision_model_name} is not supported by Liger kernel.")
 
 
+def apply_liger_kernel_to_llama4(
+    rope: bool = False,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+    layer_norm: bool = True,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Llama4 models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        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 True.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be True.
+            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 True.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+        loaded. Default is None.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.llama4 import modeling_llama4
+    from transformers.models.llama4.modeling_llama4 import Llama4ForCausalLM
+    from transformers.models.llama4.modeling_llama4 import Llama4ForConditionalGeneration
+    from transformers.models.llama4.modeling_llama4 import Llama4TextModel
+    from transformers.models.llama4.modeling_llama4 import Llama4VisionModel
+
+    from liger_kernel.transformers.model.llama4 import lce_forward as llama4_lce_forward
+
+    if rope:
+        raise NotImplementedError("liger_rotary_pos_emb is not available for Llama4 models.")
+    if rms_norm:
+        modeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
+    if swiglu:
+        modeling_llama4.Llama4TextMLP = LigerSwiGLUMLP
+
+    if cross_entropy:
+        modeling_llama4.CrossEntropyLoss = LigerCrossEntropyLoss
+
+    if fused_linear_cross_entropy:
+        modeling_llama4.Llama4ForCausalLM.forward = llama4_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+        if isinstance(model, Llama4ForConditionalGeneration):
+            language_model: Llama4ForCausalLM = model.language_model
+            vision_model: Llama4VisionModel = model.vision_model
+            text_model: Llama4TextModel = language_model.model
+        elif isinstance(model, Llama4ForCausalLM):
+            text_model = model.model
+            vision_model = None
+        elif isinstance(model, Llama4TextModel):
+            text_model = model
+            vision_model = None
+
+        else:
+            raise ValueError(f"Unsupported Llama4 model type: {type(model)}")
+
+        if text_model:
+            if rms_norm:
+                _patch_rms_norm_module(text_model.norm)
+            for decoder_layer in text_model.layers:
+                if swiglu:
+                    _patch_swiglu_module(decoder_layer.feed_forward, LigerSwiGLUMLP)
+                if rms_norm:
+                    _patch_rms_norm_module(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+        if vision_model:
+            _patch_layer_norm_module(vision_model.layernorm_pre)
+            _patch_layer_norm_module(vision_model.layernorm_post)
+
+            for layer in vision_model.model.layers:
+                if layer_norm:
+                    _patch_layer_norm_module(layer.input_layernorm)
+                    _patch_layer_norm_module(layer.post_attention_layernorm)
+
+
 def apply_liger_kernel_to_mllama(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1605,6 +1691,8 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma3": apply_liger_kernel_to_gemma3,
     "glm4": apply_liger_kernel_to_glm4,
     "llama": apply_liger_kernel_to_llama,
+    "llama4_text": apply_liger_kernel_to_llama4,
+    "llama4": apply_liger_kernel_to_llama4,
     "llava": apply_liger_kernel_to_llava,
     "granite": apply_liger_kernel_to_granite,
     "mllama": apply_liger_kernel_to_mllama,

{liger_kernel_nightly-0.5.10.dev20250624183504.dist-info → liger_kernel_nightly-0.5.10.dev20250629005644.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.10.dev20250624183504
+Version: 0.5.10.dev20250629005644
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -290,6 +290,7 @@ loss.backward()
 
 | **Model**   | **API**                                                      | **Supported Operations**                                                |
 |-------------|--------------------------------------------------------------|-------------------------------------------------------------------------|
+| Llama4 (Text) & (Multimodal)      | `liger_kernel.transformers.apply_liger_kernel_to_llama4`   | RMSNorm, LayerNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | LLaMA 2 & 3 | `liger_kernel.transformers.apply_liger_kernel_to_llama`   | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | LLaMA 3.2-Vision | `liger_kernel.transformers.apply_liger_kernel_to_mllama`   | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Mistral     | `liger_kernel.transformers.apply_liger_kernel_to_mistral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |

{liger_kernel_nightly-0.5.10.dev20250624183504.dist-info → liger_kernel_nightly-0.5.10.dev20250629005644.dist-info}/RECORD

@@ -38,7 +38,7 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=0KX0rxyy0E_uNWVE0PSTzEVzKqc5KdFHtvdHhJm23Kk,7077
+liger_kernel/transformers/__init__.py,sha256=mWMEhOabqUkPimMOmkg9DawnO-vL9u_u-N4iIqfNZeg,7259
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -53,7 +53,7 @@ liger_kernel/transformers/grpo_loss.py,sha256=uAkUNKSnUGEOqa82L9w2e6AI1kcmG8K45-
 liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCcScY,2979
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
-liger_kernel/transformers/monkey_patch.py,sha256=FacsJGO8MUbn1I7HttRAlDGhLtD8jjRTXSAwvI0iPrg,77754
+liger_kernel/transformers/monkey_patch.py,sha256=3KqEl_-WlXgUoEAEYgGs-SPolASshGem2ISFemzQAIc,81705
 liger_kernel/transformers/multi_token_attention.py,sha256=l9VDICK0dfmifUDW668hGscP8AHq2rYcM2oGUa3baRQ,1751
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -70,6 +70,7 @@ liger_kernel/transformers/model/gemma2.py,sha256=ORmzklEAMpk93nToRo4d_ZJbM4ScVE2
 liger_kernel/transformers/model/gemma3.py,sha256=JI4jj9K660HeRsofB6cpkCHBQ0OsazElArRtKUehUmw,15945
 liger_kernel/transformers/model/glm4.py,sha256=GlnEhdGJuDIqp2R9qC54biY3HwV1tWmfpJm6ijoAsrM,5257
 liger_kernel/transformers/model/llama.py,sha256=LcIxVfF0PXXWHBVJa6Ody_5fAtIpxQcI4jC_j-o51fU,12503
+liger_kernel/transformers/model/llama4.py,sha256=IgbB8sTh3dlETQnaNNy1bZLuXy-Nt7qmeAjF27ydGpg,4210
 liger_kernel/transformers/model/llava.py,sha256=bLCioday_SOm69ogMDBhy_4UsVkH2-BSl93-EXY6-7I,15076
 liger_kernel/transformers/model/loss_utils.py,sha256=WWAMdiONPaXpIvxyOim_0igLrYh0yyOok5Q9_L9xvZw,1787
 liger_kernel/transformers/model/mistral.py,sha256=okKkyashfFLfhjIT--f3JY6JHOslOtDI8U1dlpBC2Zs,5565
@@ -87,9 +88,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/METADATA,sha256=QSqkcNekBuJTkV8QZdt5x9iGR-5Z4G_BIbqhCRDvkIs,24358
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.5.10.dev20250624183504.dist-info/RECORD,,
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/METADATA,sha256=FMeKbXVH-02gQ_G0kVMIc6ftN9rv5WeQZ94Br45A9ek,24536
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.5.10.dev20250629005644.dist-info/RECORD,,