Liger support for Llama4 (#740)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->
Adding liger support to llama4 (text and multimodal). Liger rope is not
supported yet for llama4 as it has a different implementation.
The tolerance are set higher than usual to pass the tests. 

<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->

## Testing Done
<!--- This is a required section; please describe how this change was
tested. --->

<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: NVIDIA H100 80GB HBM3
- [ ] run `make test` to ensure correctness
- [x] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

---------

Co-authored-by: Shao Tang <[email protected]>
Co-authored-by: Manan Shah <[email protected]>
Co-authored-by: Vaibhav Jindal <[email protected]>

Author: 4 people

README.md

@@ -241,6 +241,7 @@ 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        |

src/liger_kernel/transformers/__init__.py

@@ -30,6 +30,7 @@
     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 @@ 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",

src/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,
+    )

src/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 @@ def apply_liger_kernel_to_glm4(
     "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,

test/convergence/bf16/test_mini_models.py

@@ -9,6 +9,8 @@
 from transformers.models.gemma2 import Gemma2ForCausalLM
 from transformers.models.llama import LlamaConfig
 from transformers.models.llama import LlamaForCausalLM
+from transformers.models.llama4 import Llama4ForCausalLM
+from transformers.models.llama4.configuration_llama4 import Llama4TextConfig
 from transformers.models.mistral import MistralConfig
 from transformers.models.mistral import MistralForCausalLM
 from transformers.models.mixtral import MixtralConfig
@@ -24,6 +26,7 @@
 from liger_kernel.transformers import apply_liger_kernel_to_glm4
 from liger_kernel.transformers import apply_liger_kernel_to_granite
 from liger_kernel.transformers import apply_liger_kernel_to_llama
+from liger_kernel.transformers import apply_liger_kernel_to_llama4
 from liger_kernel.transformers import apply_liger_kernel_to_llava
 from liger_kernel.transformers import apply_liger_kernel_to_mistral
 from liger_kernel.transformers import apply_liger_kernel_to_mixtral
@@ -46,6 +49,7 @@
 from test.utils import revert_liger_kernel_to_glm4
 from test.utils import revert_liger_kernel_to_granite
 from test.utils import revert_liger_kernel_to_llama
+from test.utils import revert_liger_kernel_to_llama4
 from test.utils import revert_liger_kernel_to_llava
 from test.utils import revert_liger_kernel_to_mistral
 from test.utils import revert_liger_kernel_to_mixtral
@@ -152,6 +156,35 @@
 device = infer_device()
 
 MINI_MODEL_SETUPS = {
+    "mini_llama4": MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_llama4,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_llama4,
+        model_class=Llama4ForCausalLM,
+        mini_model_config=Llama4TextConfig(
+            bos_token_id=1,  # None
+            eos_token_id=2,  # 151329, 151336, 151338
+            pad_token_id=2,  # 151329
+            partial_rotary_factor=1.0,
+            cross_attention_layers=None,
+            dropout=0,
+            hidden_act="silu",
+            hidden_size=1024,  # 6144
+            initializer_range=0.02,
+            intermediate_size=2048,  # 14336
+            max_position_embeddings=4096,  # 32768
+            num_attention_heads=8,  # 48
+            num_hidden_layers=4,  # 61
+            num_key_value_heads=2,
+            rms_norm_eps=1e-5,
+            rope_scaling=None,
+            rope_theta=10000.0,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 151552
+            attention_bias=True,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    ),
     "mini_llama3": MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_llama,
         liger_kernel_patch_revert_func=revert_liger_kernel_to_llama,
@@ -380,6 +413,7 @@
     ),
 }
 
+
 if QWEN3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_qwen3"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_qwen3,
@@ -770,7 +804,7 @@
     )
 
 
-def create_model(model_name="mini_llama3"):
+def create_model(model_name="mini_llama4"):
     """
     Create a mini version model
     The commented values are the original values
@@ -781,7 +815,7 @@ def create_model(model_name="mini_llama3"):
 
 
 def run_mini_model(
-    model_name="mini_llama3",
+    model_name="mini_llama4",
     num_steps=100,
     dtype=torch.bfloat16,
     lr=1e-5,
@@ -804,7 +838,7 @@ def run_mini_model(
             "rms_norm": True,
         }
 
-        if "glm4" in model_name:
+        if "glm4" in model_name or "llama4" in model_name:
             kwargs["rope"] = False
 
         model_supports_layer_norm = "qwen2_vl" in model_name
@@ -865,6 +899,19 @@ def run_mini_model(
 @pytest.mark.parametrize(
     "model_name, num_steps, lr, dtype, loss_atol, loss_rtol, logprobs_atol, logprobs_rtol, param_atol, param_rtol",
     [
+        pytest.param(
+            "mini_llama4",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            1e-2,
+            1e-1,
+            1e-1,
+            1e-2,
+            1e-2,
+            marks=pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
+        ),
         pytest.param(
             "mini_llama3",
             32,