liger-kernel-nightly 0.6.2.dev20250903164435__tar.gz → 0.6.2.dev20250905160847__tar.gz

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20250903164435
+Version: 0.6.2.dev20250905160847
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20250903164435"
+version = "0.6.2.dev20250905160847"
 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-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/src/liger_kernel/transformers/__init__.py

@@ -36,6 +36,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3_text  # noqa: F401
     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_glm4v  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v_moe  # 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
@@ -95,6 +96,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_gemma3_text",
         "apply_liger_kernel_to_glm4",
         "apply_liger_kernel_to_glm4v",
+        "apply_liger_kernel_to_glm4v_moe",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
@@ -159,6 +161,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_gemma3_text",
             "apply_liger_kernel_to_glm4",
             "apply_liger_kernel_to_glm4v",
+            "apply_liger_kernel_to_glm4v_moe",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",

liger_kernel_nightly-0.6.2.dev20250905160847/src/liger_kernel/transformers/model/glm4v_moe.py

@@ -0,0 +1,152 @@
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeCausalLMOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+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[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, Glm4vMoeCausalLMOutputWithPast]:
+    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]`.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
+        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+            The rope index difference between sequence length and multimodal rope.
+
+
+        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).
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
+    >>> import torch
+
+    >>> MODEL_PATH = "zai-org/GLM-4.5V"
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image",
+                    "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+                },
+                {
+                    "type": "text",
+                    "text": "describe this image"
+                }
+            ],
+        }
+    ]
+    >>> processor = AutoProcessor.from_pretrained(MODEL_PATH)
+    >>> model = Glm4vMoeForConditionalGeneration.from_pretrained(
+        pretrained_model_name_or_path=MODEL_PATH,
+        torch_dtype="auto",
+        device_map="auto",
+    )
+    >>> inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt"
+    ).to(model.device)
+    >>> inputs.pop("token_type_ids", None)
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=8192)
+    >>> output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
+    ```
+    """
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        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 skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        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:
+        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 Glm4vMoeCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+    )

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/src/liger_kernel/transformers/monkey_patch.py

@@ -1928,6 +1928,107 @@ def apply_liger_kernel_to_glm4v(
                     _patch_rms_norm_module(decoder_layer.post_mlp_layernorm)
 
 
+def apply_liger_kernel_to_glm4v_moe(
+    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.glm4v_moe import modeling_glm4v_moe
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeForConditionalGeneration
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeModel
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextModel
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeVisionModel
+
+    from liger_kernel.transformers.model.glm4v_moe import lce_forward as glm4v_moe_lce_forward
+    from liger_kernel.transformers.rms_norm import LigerRMSNormForGlm4
+
+    if rope:
+        raise NotImplementedError("liger_rotary_pos_emb is not available for Glm4 models.")
+    if rms_norm:
+        modeling_glm4v_moe.Glm4vRMSNorm = LigerRMSNormForGlm4
+    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:
+            model.forward = MethodType(glm4v_moe_lce_forward, model)
+        else:
+            modeling_glm4v_moe.Glm4vMoeForConditionalGeneration.forward = glm4v_moe_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, (Glm4vMoeForConditionalGeneration, Glm4vMoeModel)):
+            # Note: language_model and visual properties can be accessed throught conditional class for BC.
+            # Not sure if it is subject to changes in the future.
+            # Reference: https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4v_moe/modeling_glm4v_moe.py#L337
+            text_model: Glm4vMoeTextModel = model.language_model
+            vision_model: Glm4vMoeVisionModel = model.visual
+            Glm4vMoeTextMoE = modeling_glm4v_moe.Glm4vMoeTextMoE
+        elif isinstance(model, Glm4vMoeTextModel):
+            text_model: Glm4vMoeTextModel = model
+            vision_model = None
+        else:
+            # Note: Currently there's no support for patching vision model only. Feel free to raise an issue if needed.
+            raise TypeError(
+                f"Unsupported glm4v_moe model type. `model` must be `Glm4vMoeForConditionalGeneration`, `Glm4vMoeVisionModel` or `Glm4vMoeTextModel`. Got: {type(model)}"
+            )
+
+        if vision_model is not None:
+            _patch_rms_norm_module(vision_model.post_conv_layernorm)
+            _patch_rms_norm_module(vision_model.post_layernorm)
+            for vision_block in vision_model.blocks:
+                if rms_norm:
+                    _patch_rms_norm_module(vision_block.norm1)
+                    _patch_rms_norm_module(vision_block.norm2)
+                if swiglu:
+                    _patch_swiglu_module(vision_block.mlp, LigerSwiGLUMLP)
+
+        if text_model is not None:
+            if rms_norm:
+                _patch_rms_norm_module(text_model.norm)
+            for decoder_layer in text_model.layers:
+                if swiglu:
+                    decoder_layer.mlp = _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+                if rms_norm:
+                    _patch_rms_norm_module(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+        if isinstance(Glm4vMoeTextMoE, type) and isinstance(decoder_layer.mlp, Glm4vMoeTextMoE):
+            experts = getattr(decoder_layer.mlp, "experts", None)
+            if experts is not None:
+                for expert in experts:
+                    _patch_swiglu_module(expert, LigerSwiGLUMLP)
+            if decoder_layer.mlp.shared_experts is not None:
+                _patch_swiglu_module(decoder_layer.mlp.shared_experts, LigerSwiGLUMLP)
+            for decoder_layer in text_model.layers:
+                if rms_norm:
+                    _patch_rms_norm_module(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 # 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,
@@ -1936,6 +2037,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma3": apply_liger_kernel_to_gemma3,
     "glm4": apply_liger_kernel_to_glm4,
     "glm4v": apply_liger_kernel_to_glm4v,
+    "glm4v_moe": apply_liger_kernel_to_glm4v_moe,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,
     "llama4": apply_liger_kernel_to_llama4,

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20250903164435
+Version: 0.6.2.dev20250905160847
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -192,6 +192,7 @@ src/liger_kernel/transformers/model/gemma2.py
 src/liger_kernel/transformers/model/gemma3.py
 src/liger_kernel/transformers/model/glm4.py
 src/liger_kernel/transformers/model/glm4v.py
+src/liger_kernel/transformers/model/glm4v_moe.py
 src/liger_kernel/transformers/model/llama.py
 src/liger_kernel/transformers/model/llama4.py
 src/liger_kernel/transformers/model/llava.py

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/test/convergence/bf16/test_mini_models.py

@@ -23,6 +23,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
 from liger_kernel.transformers import apply_liger_kernel_to_glm4
 from liger_kernel.transformers import apply_liger_kernel_to_glm4v
+from liger_kernel.transformers import apply_liger_kernel_to_glm4v_moe
 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
@@ -48,6 +49,7 @@ from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
 from test.utils import revert_liger_kernel_to_glm4
 from test.utils import revert_liger_kernel_to_glm4v
+from test.utils import revert_liger_kernel_to_glm4v_moe
 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
@@ -162,6 +164,15 @@ try:
 except ImportError:
     GLM4V_AVAILABLE = False
 
+try:
+    # Glm4v_moe is only available in transformers>=4.51.3
+    from transformers.models.glm4v_moe.configuration_glm4v_moe import Glm4vMoeConfig
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeForConditionalGeneration
+
+    GLM4V_MOE_AVAILABLE = True
+except ImportError:
+    GLM4V_MOE_AVAILABLE = False
+
 try:
     from transformers.models.gemma3.configuration_gemma3 import Gemma3TextConfig
     from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
@@ -900,6 +911,83 @@ if GLM4V_AVAILABLE:
         ),
     )
 
+if GLM4V_MOE_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_glm4v_moe"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_glm4v_moe,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_glm4v_moe,
+        model_class=Glm4vMoeForConditionalGeneration,
+        mini_model_config=Glm4vMoeConfig(
+            bos_token_id=1,  # None
+            eos_token_id=2,  # 151329, 151336, 151338
+            pad_token_id=2,  # 151329
+            image_token_id=151343,
+            video_token_id=151344,
+            image_start_token_id=151339,
+            image_end_token_id=151340,
+            video_start_token_id=151341,
+            video_end_token_id=151342,
+            partial_rotary_factor=0.5,
+            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=500_000,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 151552
+            attention_bias=True,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+            text_config={
+                "partial_rotary_factor": 0.5,
+                "hidden_act": "silu",
+                "hidden_size": 1024,
+                "intermediate_size": 2048,
+                "max_position_embeddings": 4096,
+                "num_attention_heads": 8,
+                "num_hidden_layers": 4,
+                "num_key_value_heads": 2,
+                "rms_norm_eps": 1e-5,
+                "rope_scaling": {
+                    "type": "default",
+                    "mrope_section": [8, 12, 12],  # (temporal, height, width)
+                },
+                "rope_theta": 500_000,
+                "vocab_size": 32000,
+                "attention_bias": True,
+                "attention_dropout": 0.0,
+                "moe_intermediate_size": 1408,
+                "num_experts_per_tok": 2,
+                "n_shared_experts": 1,
+                "n_routed_experts": 128,
+                "routed_scaling_factor": 1.0,
+                "n_group": 1,
+                "topk_group": 1,
+                "first_k_dense_replace": 1,
+                "norm_topk_prob": True,
+            },
+            vision_config={
+                "depth": 4,  # 32
+                "hidden_act": "silu",
+                "hidden_size": 128,  # 1280
+                "intermediate_size": 256,  # 3420
+                "num_heads": 16,
+                "in_chans": 3,
+                "out_hidden_size": 128,  # 3584
+                "patch_size": 14,
+                "spatial_merge_size": 2,
+                "temporal_patch_size": 2,
+            },
+        ),
+    )
+
 if SMOLLM3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_smollm3"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_smollm3,
@@ -1304,6 +1392,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_glm4v_moe",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            2e-1,
+            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 GLM4V_MOE_AVAILABLE,
+                    reason="Glm4v_moe not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_smollm3",
             32,

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -71,6 +71,7 @@ try:
 except ImportError:
     QWEN2_5_VL_AVAILABLE = False
 
+
 try:
     # Mllama is only available in transformers>=4.45.0
     from transformers.models.mllama.configuration_mllama import MllamaConfig

{liger_kernel_nightly-0.6.2.dev20250903164435 → liger_kernel_nightly-0.6.2.dev20250905160847}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -23,6 +23,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_gemma2
 from liger_kernel.transformers import apply_liger_kernel_to_gemma3_text
 from liger_kernel.transformers import apply_liger_kernel_to_glm4
 from liger_kernel.transformers import apply_liger_kernel_to_glm4v
+from liger_kernel.transformers import apply_liger_kernel_to_glm4v_moe
 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
@@ -48,6 +49,7 @@ from test.utils import revert_liger_kernel_to_gemma2
 from test.utils import revert_liger_kernel_to_gemma3_text
 from test.utils import revert_liger_kernel_to_glm4
 from test.utils import revert_liger_kernel_to_glm4v
+from test.utils import revert_liger_kernel_to_glm4v_moe
 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
@@ -162,6 +164,15 @@ try:
 except ImportError:
     GLM4V_AVAILABLE = False
 
+try:
+    # Glm4v_moe is only available in transformers>=4.51.3
+    from transformers.models.glm4v_moe.configuration_glm4v_moe import Glm4vMoeConfig
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeForConditionalGeneration
+
+    GLM4V_MOE_AVAILABLE = version.parse(transformers.__version__) >= version.parse("4.53.1")
+except ImportError:
+    GLM4V_MOE_AVAILABLE = False
+
 try:
     from transformers.models.gemma3.configuration_gemma3 import Gemma3TextConfig
     from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
@@ -898,6 +909,83 @@ if GLM4V_AVAILABLE:
         ),
     )
 
+if GLM4V_MOE_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_glm4v_moe"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_glm4v_moe,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_glm4v_moe,
+        model_class=Glm4vMoeForConditionalGeneration,
+        mini_model_config=Glm4vMoeConfig(
+            bos_token_id=1,  # None
+            eos_token_id=2,  # 151329, 151336, 151338
+            pad_token_id=2,  # 151329
+            image_token_id=151343,
+            video_token_id=151344,
+            image_start_token_id=151339,
+            image_end_token_id=151340,
+            video_start_token_id=151341,
+            video_end_token_id=151342,
+            partial_rotary_factor=0.5,
+            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=500_000,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32000,  # 151552
+            attention_bias=True,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+            text_config={
+                "partial_rotary_factor": 0.5,
+                "hidden_act": "silu",
+                "hidden_size": 1024,
+                "intermediate_size": 2048,
+                "max_position_embeddings": 4096,
+                "num_attention_heads": 8,
+                "num_hidden_layers": 4,
+                "num_key_value_heads": 2,
+                "rms_norm_eps": 1e-5,
+                "rope_scaling": {
+                    "type": "default",
+                    "mrope_section": [8, 12, 12],  # (temporal, height, width)
+                },
+                "rope_theta": 500_000,
+                "vocab_size": 32000,
+                "attention_bias": True,
+                "attention_dropout": 0.0,
+                "moe_intermediate_size": 1408,
+                "num_experts_per_tok": 2,
+                "n_shared_experts": 1,
+                "n_routed_experts": 128,
+                "routed_scaling_factor": 1.0,
+                "n_group": 1,
+                "topk_group": 1,
+                "first_k_dense_replace": 1,
+                "norm_topk_prob": True,
+            },
+            vision_config={
+                "depth": 4,  # 32
+                "hidden_act": "silu",
+                "hidden_size": 128,  # 1280
+                "intermediate_size": 256,  # 3420
+                "num_heads": 16,
+                "in_chans": 3,
+                "out_hidden_size": 128,  # 3584
+                "patch_size": 14,
+                "spatial_merge_size": 2,
+                "temporal_patch_size": 2,
+            },
+        ),
+    )
+
 if SMOLLM3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_smollm3"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_smollm3,
@@ -1324,6 +1412,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_glm4v_moe",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            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 GLM4V_MOE_AVAILABLE,
+                    reason="Glm4v_moe not available in this version of transformers",
+                ),
+            ],
+        ),
         # TODO: Gemma2 test for bf16 is not passing within the tolerance range, might be casting issue, need to investigate
         # pytest.param(
         #     "mini_gemma2",