Add support for Qwen3Next model with Liger kernels (#912)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->
Adding qwen3-next model support  #896 

https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct?library=transformers
Available in transformers>=4.57.0
<!---
## 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: <BLANK>
- [ ] run `make test` to ensure correctness
- [ ] run `make checkstyle` to ensure code style
- [ ] run `make test-convergence` to ensure convergence

---------

Signed-off-by: Tcc0403 <76503978+Tcc0403@users.noreply.github.com>
Co-authored-by: Tcc0403 <76503978+Tcc0403@users.noreply.github.com>

Author: vvvdwbvvv

src/liger_kernel/transformers/__init__.py

@@ -55,6 +55,7 @@
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_next  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
 
 
@@ -117,6 +118,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
+        "apply_liger_kernel_to_qwen3_next",
         "apply_liger_kernel_to_smollm3",
     }
 
@@ -185,6 +187,7 @@ def __getattr__(name: str):
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
+            "apply_liger_kernel_to_qwen3_next",
             "apply_liger_kernel_to_smollm3",
         ]
     )

src/liger_kernel/transformers/model/qwen3_next.py

@@ -0,0 +1,134 @@
+from typing import TYPE_CHECKING
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import MoeCausalLMOutputWithPast
+from transformers.modeling_outputs import MoeModelOutputWithPast
+
+if TYPE_CHECKING:
+    from transformers.models.qwen3_next.modeling_qwen3_next import load_balancing_loss_func
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+def lce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> MoeCausalLMOutputWithPast:
+    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]`.
+
+        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:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+    >>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-Next-80B-A3B-Instruct")
+
+    >>> prompt = "Give me a short introduction to large language model."
+    >>> 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_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = 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,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # 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 skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        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 model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -2180,6 +2180,97 @@ def apply_liger_kernel_to_falcon_h1(
                 _patch_rms_norm_module(decoder_layer.pre_ff_layernorm)
 
 
+def apply_liger_kernel_to_qwen3_next(
+    rope: bool = False,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace GLM4v_moe models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is False.
+        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 SwiGLUMLP. Default is True.
+        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.qwen3_next import modeling_qwen3_next
+    from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextForCausalLM
+    from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextMLP
+    from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextModel
+    from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextSparseMoeBlock
+
+    from liger_kernel.transformers.model.qwen3_next import lce_forward as qwen3_next_lce_forward
+    from liger_kernel.transformers.rms_norm import LigerRMSNormForQwen3Next
+    from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
+
+    if rope:
+        # It might enocunter nan issue
+        # modeling_qwen3_next.apply_rotary_pos_emb = liger_rotary_pos_emb
+        raise NotImplementedError("liger_rotary_pos_emb is not available for Qwen3Next models.")
+    if rms_norm:
+        modeling_qwen3_next.Qwen3NextRMSNorm = LigerRMSNormForQwen3Next
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+    if fused_linear_cross_entropy:
+        if model is not None:
+            if isinstance(model, Qwen3NextForCausalLM):
+                model.forward = MethodType(qwen3_next_lce_forward, model)
+            else:
+                raise TypeError(
+                    f" fused_linear_cross_entropy is only applicable on Qwen3NextForCausalLM. Got: {type(model)}"
+                )
+        else:
+            modeling_qwen3_next.Qwen3NextForCausalLM.forward = qwen3_next_lce_forward
+    if swiglu:
+        # Qwen3MoeMLP and Qwen3NextMLP are identical, hence we reuse LigerQwen3MoeSwiGLUMLP
+        modeling_qwen3_next.Qwen3NextMLP = LigerQwen3MoeSwiGLUMLP
+
+    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, (Qwen3NextForCausalLM, Qwen3NextModel)):
+            base_model: Qwen3NextForCausalLM = getattr(model, model.base_model_prefix, model)
+        else:
+            raise TypeError(
+                f"Unsupported qwen3_next model type. `model` must be `Qwen3NextForCausalLM`, `Qwen3NextModel`. Got: {type(model)}"
+            )
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+
+        for decoder_layer in base_model.layers:
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+            # Qwen3MoeMLP and Qwen3NextMLP are identical, hence we reuse LigerQwen3MoeSwiGLUMLP
+            if swiglu:
+                if isinstance(decoder_layer.mlp, Qwen3NextMLP):
+                    _patch_swiglu_module(decoder_layer.mlp, LigerQwen3MoeSwiGLUMLP)
+                if isinstance(decoder_layer.mlp, Qwen3NextSparseMoeBlock):
+                    _patch_swiglu_module(decoder_layer.mlp.shared_expert, LigerQwen3MoeSwiGLUMLP)
+                    experts = getattr(decoder_layer.mlp, "experts", None)
+                    if experts is not None:
+                        for expert in experts:
+                            _patch_swiglu_module(expert, LigerQwen3MoeSwiGLUMLP)
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
@@ -2207,6 +2298,7 @@ def apply_liger_kernel_to_falcon_h1(
     "qwen2_vl_text": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "qwen2_5_vl_text": apply_liger_kernel_to_qwen2_5_vl,
+    "qwen3_next": apply_liger_kernel_to_qwen3_next,
     "smollm3": apply_liger_kernel_to_smollm3,
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,

src/liger_kernel/transformers/rms_norm.py

@@ -77,3 +77,10 @@ def __init__(
         self, hidden_size, eps=1e-6, offset=0.0, casting_mode="llama", init_fn="ones", in_place=False, row_mode=None
     ):
         super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)
+
+
+class LigerRMSNormForQwen3Next(LigerRMSNorm):
+    def __init__(
+        self, hidden_size, eps=1e-6, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False, row_mode=None
+    ):
+        super().__init__(hidden_size, eps, offset, casting_mode, init_fn, in_place, row_mode)

test/convergence/bf16/test_mini_models.py

@@ -40,6 +40,7 @@
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
 from liger_kernel.transformers import apply_liger_kernel_to_qwen3
 from liger_kernel.transformers import apply_liger_kernel_to_qwen3_moe
+from liger_kernel.transformers import apply_liger_kernel_to_qwen3_next
 from liger_kernel.transformers import apply_liger_kernel_to_smollm3
 from test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
@@ -68,6 +69,7 @@
 from test.utils import revert_liger_kernel_to_qwen2_vl
 from test.utils import revert_liger_kernel_to_qwen3
 from test.utils import revert_liger_kernel_to_qwen3_moe
+from test.utils import revert_liger_kernel_to_qwen3_next
 from test.utils import revert_liger_kernel_to_smollm3
 from test.utils import set_seed
 from test.utils import simple_collate_fn
@@ -212,6 +214,15 @@
 except ImportError:
     FALCONH1_AVAILABLE = False
 
+try:
+    # Qwen3Next is only available in transformers>=4.57.0
+    from transformers.models.qwen3_next.configuration_qwen3_next import Qwen3NextConfig
+    from transformers.models.qwen3_next.modeling_qwen3_next import Qwen3NextForCausalLM
+
+    QWEN3NEXT_AVAILABLE = True
+except ImportError:
+    QWEN3NEXT_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -1106,6 +1117,43 @@
         ),
     )
 
+if QWEN3NEXT_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_qwen3_next"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen3_next,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen3_next,
+        model_class=Qwen3NextForCausalLM,
+        mini_model_config=Qwen3NextConfig(  # Copypaste Qwen3MoeConfig
+            vocab_size=32000,
+            hidden_size=896,
+            intermediate_size=4864,
+            num_hidden_layers=4,
+            num_attention_heads=8,
+            num_key_value_heads=2,
+            hidden_act="silu",
+            max_position_embeddings=32768,
+            initializer_range=0.02,
+            rms_norm_eps=1e-6,
+            use_cache=True,
+            tie_word_embeddings=False,
+            rope_theta=10000.0,
+            rope_scaling=None,
+            attention_bias=False,
+            use_sliding_window=False,
+            sliding_window=4096,
+            max_window_layers=28,
+            attention_dropout=0.0,
+            decoder_sparse_step=1,
+            moe_intermediate_size=768,
+            num_experts_per_tok=2,
+            num_experts=8,
+            norm_topk_prob=False,
+            output_router_logits=False,
+            router_aux_loss_coef=0.001,
+            # config.dtype must be set if fla installed since there's a bug in the original code (No torch.get_current_dtype())
+            dtype=torch.bfloat16,
+        ),
+    )
+
 
 def create_model(model_name="mini_llama4"):
     """
@@ -1141,7 +1189,7 @@ def run_mini_model(
             "rms_norm": True,
         }
 
-        if "glm4" in model_name:
+        if "glm4" in model_name or "qwen3_next" in model_name:
             kwargs["rope"] = False
 
         model_supports_layer_norm = "qwen2_vl" in model_name
@@ -1634,6 +1682,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_qwen3_next",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            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.mark.skipif(
+                    not QWEN3NEXT_AVAILABLE,
+                    reason="Qwen3Next not available in this version of transformers",
+                ),
+            ],
+        ),
     ],
 )
 def test_mini_model(