liger-kernel-nightly 0.5.9.dev20250508211521__tar.gz → 0.5.9.dev20250512213150__tar.gz

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.9.dev20250508211521
+Version: 0.5.9.dev20250512213150
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -300,6 +300,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2.5-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_5_vl`    | RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
+| Qwen3 MoE | `liger_kernel_transformers.apply_liger_kernel_to_qwen3_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/README.md

@@ -252,6 +252,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2.5-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_5_vl`    | RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
+| Qwen3 MoE | `liger_kernel_transformers.apply_liger_kernel_to_qwen3_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.9.dev20250508211521"
+version = "0.5.9.dev20250512213150"
 description = "Efficient Triton kernels for LLM Training"
 urls = { "Homepage" = "https://github.com/linkedin/Liger-Kernel" }
 readme = { file = "README.md", content-type = "text/markdown" }

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/src/liger_kernel/transformers/__init__.py

@@ -14,6 +14,7 @@ from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerBlockSparseTop2MLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerPhi3SwiGLUMLP  # noqa: F401
+from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP  # noqa: F401
 from liger_kernel.transformers.tvd import LigerTVDLoss  # noqa: F401
 
@@ -40,6 +41,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_5_vl  # noqa: F401
     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
 
 
 # Check if 'transformers' is installed
@@ -95,6 +97,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_5_vl",
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
+        "apply_liger_kernel_to_qwen3_moe",
     }
 
     if name in monkey_patch_symbols:
@@ -118,6 +121,7 @@ __all__ = [
     "liger_rotary_pos_emb",
     "LigerBlockSparseTop2MLP",
     "LigerPhi3SwiGLUMLP",
+    "LigerQwen3MoeSwiGLUMLP",
     "LigerSwiGLUMLP",
     "LigerTVDLoss",
 ]
@@ -147,5 +151,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2_5_vl",
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
+            "apply_liger_kernel_to_qwen3_moe",
         ]
     )

liger_kernel_nightly-0.5.9.dev20250512213150/src/liger_kernel/transformers/model/qwen3_moe.py

@@ -0,0 +1,134 @@
+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
+from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
+from transformers.models.qwen3_moe.modeling_qwen3_moe import _CONFIG_FOR_DOC
+from transformers.models.qwen3_moe.modeling_qwen3_moe import QWEN3_MOE_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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,
+    **loss_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 AutoTokenizer, Qwen3MoeForCausalLM
+
+    >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+    >>> 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_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,
+    )
+
+    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 = loss_kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+
+    # if in training mode, do not materialize logits
+    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,
+            **loss_kwargs,
+        )
+    else:  # if in inference model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_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,
+    )

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/src/liger_kernel/transformers/monkey_patch.py

@@ -1102,6 +1102,61 @@ def apply_liger_kernel_to_qwen3(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_qwen3_moe(
+    rope: bool = True,
+    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 Qwen3 models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+    from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeModel
+
+    from liger_kernel.transformers.model.qwen3_moe import lce_forward as qwen3_lce_forward
+    from liger_kernel.transformers.swiglu import LigerQwen3MoeSwiGLUMLP
+
+    if rope:
+        modeling_qwen3_moe.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_qwen3_moe.Qwen3MoeRMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_qwen3_moe.Qwen3MoeForCausalLM.forward = qwen3_lce_forward
+
+    if swiglu:
+        modeling_qwen3_moe.Qwen3MoeMLP = 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
+
+        # get the base model from the model instance
+        base_model: Qwen3MoeModel = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerQwen3MoeSwiGLUMLP)
+            if rms_norm:
+                _patch_rms_norm_module(decoder_layer.input_layernorm)
+                _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 def apply_liger_kernel_to_qwen2_vl(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1455,6 +1510,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
     "qwen3": apply_liger_kernel_to_qwen3,
+    "qwen3_moe": apply_liger_kernel_to_qwen3_moe,
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "phi3": apply_liger_kernel_to_phi3,

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/src/liger_kernel/transformers/swiglu.py

@@ -56,3 +56,24 @@ class LigerPhi3SwiGLUMLP(nn.Module):
         up_states = self.gate_up_proj(x)
         gate, up_states = up_states.chunk(2, dim=-1)
         return self.down_proj(LigerSiLUMulFunction.apply(gate, up_states))
+
+
+class LigerQwen3MoeSwiGLUMLP(nn.Module):
+    """
+    Patch Qwen3MoeMLP to use LigerSiLUMulFunction.
+    https://github.com/huggingface/transformers/blob/v4.51.3/src/transformers/models/qwen3_moe/modular_qwen3_moe.py#L57
+    """
+
+    def __init__(self, config, intermediate_size=None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = intermediate_size if intermediate_size is not None else config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        if config.hidden_act not in ["silu", "swish"]:
+            raise ValueError(f"Activation function {config.hidden_act} not supported.")
+
+    def forward(self, x):
+        return self.down_proj(LigerSiLUMulFunction.apply(self.gate_proj(x), self.up_proj(x)))

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.9.dev20250508211521
+Version: 0.5.9.dev20250512213150
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -300,6 +300,7 @@ loss.backward()
 | Qwen2-VL, & QVQ       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen2.5-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_5_vl`    | RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
+| Qwen3 MoE | `liger_kernel_transformers.apply_liger_kernel_to_qwen3_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 | Phi3 & Phi3.5       | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Granite 3.0 & 3.1   | `liger_kernel.transformers.apply_liger_kernel_to_granite`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss |
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -178,6 +178,7 @@ src/liger_kernel/transformers/model/qwen2.py
 src/liger_kernel/transformers/model/qwen2_5_vl.py
 src/liger_kernel/transformers/model/qwen2_vl.py
 src/liger_kernel/transformers/model/qwen3.py
+src/liger_kernel/transformers/model/qwen3_moe.py
 src/liger_kernel/transformers/trainer/__init__.py
 src/liger_kernel/transformers/trainer/orpo_trainer.py
 src/liger_kernel/triton/__init__.py

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/test/convergence/bf16/test_mini_models.py

@@ -34,6 +34,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
 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 test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
@@ -53,6 +54,7 @@ from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
 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 set_seed
 from test.utils import simple_collate_fn
 from test.utils import supports_bfloat16
@@ -87,6 +89,8 @@ except ImportError:
 try:
     from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
     from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
+    from transformers.models.qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig
+    from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeForCausalLM
 
     QWEN3_AVAILABLE = True
 except ImportError:
@@ -395,6 +399,41 @@ if QWEN3_AVAILABLE:
         ),
     )
 
+    MINI_MODEL_SETUPS["mini_qwen3_moe"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen3_moe,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen3_moe,
+        model_class=Qwen3MoeForCausalLM,
+        mini_model_config=Qwen3MoeConfig(
+            vocab_size=32000,  # 151936
+            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,
+            mlp_only_layers=None,
+        ),
+    )
+
 if GEMMA3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_gemma3_text"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_gemma3_text,
@@ -907,6 +946,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_qwen3_moe",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            1e-2,
+            1e-1,
+            1e-2,
+            1e-2,
+            1e-2,
+            marks=[
+                pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
+                pytest.mark.skipif(
+                    not QWEN3_AVAILABLE,
+                    reason="Qwen3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_qwen2_vl",
             32,

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -34,6 +34,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
 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 test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
@@ -53,6 +54,7 @@ from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
 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 set_seed
 from test.utils import simple_collate_fn
 from test.utils import supports_bfloat16
@@ -87,6 +89,8 @@ except ImportError:
 try:
     from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
     from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
+    from transformers.models.qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig
+    from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeForCausalLM
 
     QWEN3_AVAILABLE = True
 except ImportError:
@@ -395,6 +399,41 @@ if QWEN3_AVAILABLE:
         ),
     )
 
+    MINI_MODEL_SETUPS["mini_qwen3_moe"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen3_moe,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen3_moe,
+        model_class=Qwen3MoeForCausalLM,
+        mini_model_config=Qwen3MoeConfig(
+            vocab_size=32000,  # 151936
+            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,
+            mlp_only_layers=None,
+        ),
+    )
+
 if GEMMA3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_gemma3_text"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_gemma3_text,
@@ -906,6 +945,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_qwen3_moe",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            1e-2,
+            2e-1,
+            1e-2,
+            1e-2,
+            1e-2,
+            marks=[
+                pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
+                pytest.mark.skipif(
+                    not QWEN3_AVAILABLE,
+                    reason="Qwen3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_qwen2_vl",
             32,

{liger_kernel_nightly-0.5.9.dev20250508211521 → liger_kernel_nightly-0.5.9.dev20250512213150}/test/convergence/fp32/test_mini_models.py

@@ -34,6 +34,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_qwen2
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
 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 test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
@@ -53,6 +54,7 @@ from test.utils import revert_liger_kernel_to_qwen2
 from test.utils import revert_liger_kernel_to_qwen2_5_vl
 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 set_seed
 from test.utils import simple_collate_fn
 
@@ -129,6 +131,8 @@ except ImportError:
 try:
     from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
     from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
+    from transformers.models.qwen3_moe.configuration_qwen3_moe import Qwen3MoeConfig
+    from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeForCausalLM
 
     QWEN3_AVAILABLE = True
 except ImportError:
@@ -394,6 +398,41 @@ if QWEN3_AVAILABLE:
         ),
     )
 
+    MINI_MODEL_SETUPS["mini_qwen3_moe"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen3_moe,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen3_moe,
+        model_class=Qwen3MoeForCausalLM,
+        mini_model_config=Qwen3MoeConfig(
+            vocab_size=32000,  # 151936
+            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,
+            mlp_only_layers=None,
+        ),
+    )
+
 if GEMMA3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_gemma3_text"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_gemma3_text,
@@ -869,6 +908,22 @@ def run_mini_model(
                 reason="Qwen3 not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_qwen3_moe",
+            32,
+            1e-4,
+            torch.float32,
+            1e-8,
+            1e-5,
+            5e-3,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=pytest.mark.skipif(
+                not QWEN3_AVAILABLE,
+                reason="Qwen3 not available in this version of transformers",
+            ),
+        ),
         pytest.param(  # qwen2_vl requires slightly larger tolerances to pass this test after bug fix to qwen2_vl in transformers v4.47.0
             "mini_qwen2_vl",
             32,