liger-kernel-nightly 0.6.1.dev20250819172918__tar.gz → 0.6.1.dev20250819173444__tar.gz

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.1.dev20250819172918
+Version: 0.6.1.dev20250819173444
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.1.dev20250819172918"
+version = "0.6.1.dev20250819173444"
 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-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/src/liger_kernel/transformers/__init__.py

@@ -35,6 +35,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3  # noqa: F401
     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_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
@@ -93,6 +94,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_gemma3",
         "apply_liger_kernel_to_gemma3_text",
         "apply_liger_kernel_to_glm4",
+        "apply_liger_kernel_to_glm4v",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
@@ -156,6 +158,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_gemma3",
             "apply_liger_kernel_to_gemma3_text",
             "apply_liger_kernel_to_glm4",
+            "apply_liger_kernel_to_glm4v",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",

liger_kernel_nightly-0.6.1.dev20250819173444/src/liger_kernel/transformers/model/glm4v.py

@@ -0,0 +1,150 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+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,
+    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,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    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]`.
+
+        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 PIL import Image
+    >>> from transformers import AutoTokenizer, Glm4vForConditionalGeneration
+
+    >>> MODEL_PATH = "THUDM/GLM-4.1V-9B-Thinking"
+    >>> 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, use_fast=True)
+    >>> model = Glm4vForConditionalGeneration.from_pretrained(
+        pretrained_model_name_or_path=MODEL_PATH,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+    )
+    >>> inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt"
+    ).to(model.device)
+    >>> 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)
+    <think>Got it, let's describe the image. First, there's a vintage car, specifically a Volkswagen Beetle
+    ```"""
+    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=return_dict,
+        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 CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/src/liger_kernel/transformers/monkey_patch.py

@@ -1839,6 +1839,95 @@ def apply_liger_kernel_to_glm4(
                 _patch_rms_norm_module(decoder_layer.post_mlp_layernorm, in_place=False)
 
 
+def apply_liger_kernel_to_glm4v(
+    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 GLM-4v 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 SwiGLU Glm4MLP. 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 import modeling_glm4v
+    from transformers.models.glm4v.modeling_glm4v import Glm4vForConditionalGeneration
+    from transformers.models.glm4v.modeling_glm4v import Glm4vModel
+    from transformers.models.glm4v.modeling_glm4v import Glm4vTextModel
+    from transformers.models.glm4v.modeling_glm4v import Glm4vVisionModel
+
+    from liger_kernel.transformers.model.glm4v import lce_forward as glm4v_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.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_lce_forward, model)
+        else:
+            modeling_glm4v.Glm4vForConditionalGeneration.forward = glm4v_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, (Glm4vForConditionalGeneration, Glm4vModel)):
+            # 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/modeling_glm4v.py#L1305
+            text_model: Glm4vTextModel = model.language_model
+            vision_model: Glm4vVisionModel = model.visual
+        elif isinstance(model, Glm4vTextModel):
+            text_model: Glm4vTextModel = 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 glm4.1v model type. `model` must be `Glm4VLForConditionalGeneration`, `Glm4vVisionModel` or `Glm4vTextModel`. Got: {type(model)}"
+            )
+
+        if vision_model is not None:
+            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:
+                    _patch_swiglu_module(decoder_layer.mlp, LigerPhi3SwiGLUMLP)
+                if rms_norm:
+                    _patch_rms_norm_module(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_self_attn_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_mlp_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,
@@ -1846,6 +1935,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma3_text": apply_liger_kernel_to_gemma3_text,
     "gemma3": apply_liger_kernel_to_gemma3,
     "glm4": apply_liger_kernel_to_glm4,
+    "glm4v": apply_liger_kernel_to_glm4v,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,
     "llama4": apply_liger_kernel_to_llama4,

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.1.dev20250819172918
+Version: 0.6.1.dev20250819173444
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -190,6 +190,7 @@ src/liger_kernel/transformers/model/gemma.py
 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/llama.py
 src/liger_kernel/transformers/model/llama4.py
 src/liger_kernel/transformers/model/llava.py

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/test/convergence/bf16/test_mini_models.py

@@ -22,6 +22,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_gemma
 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_granite
 from liger_kernel.transformers import apply_liger_kernel_to_llama
 from liger_kernel.transformers import apply_liger_kernel_to_llama4
@@ -46,6 +47,7 @@ from test.utils import revert_liger_kernel_to_gemma
 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_granite
 from test.utils import revert_liger_kernel_to_llama
 from test.utils import revert_liger_kernel_to_llama4
@@ -151,6 +153,15 @@ try:
 except ImportError:
     GLM4_AVAILABLE = False
 
+try:
+    # Glm4v is only available in transformers>=4.51.3
+    from transformers.models.glm4v.configuration_glm4v import Glm4vConfig
+    from transformers.models.glm4v.modeling_glm4v import Glm4vForConditionalGeneration
+
+    GLM4V_AVAILABLE = True
+except ImportError:
+    GLM4V_AVAILABLE = False
+
 try:
     from transformers.models.gemma3.configuration_gemma3 import Gemma3TextConfig
     from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
@@ -822,6 +833,73 @@ if GLM4_AVAILABLE:
         ),
     )
 
+if GLM4V_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_glm4v"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_glm4v,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_glm4v,
+        model_class=Glm4vForConditionalGeneration,
+        mini_model_config=Glm4vConfig(
+            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,
+            },
+            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,
@@ -1207,6 +1285,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_glm4v",
+            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_AVAILABLE,
+                    reason="Glm4v not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_smollm3",
             32,

{liger_kernel_nightly-0.6.1.dev20250819172918 → liger_kernel_nightly-0.6.1.dev20250819173444}/test/convergence/bf16/test_mini_models_with_logits.py

@@ -22,6 +22,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_gemma
 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_granite
 from liger_kernel.transformers import apply_liger_kernel_to_llama
 from liger_kernel.transformers import apply_liger_kernel_to_llama4
@@ -46,6 +47,7 @@ from test.utils import revert_liger_kernel_to_gemma
 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_granite
 from test.utils import revert_liger_kernel_to_llama
 from test.utils import revert_liger_kernel_to_llama4
@@ -151,6 +153,15 @@ try:
 except ImportError:
     GLM4_AVAILABLE = False
 
+try:
+    # Glm4v is only available in transformers>=4.51.3
+    from transformers.models.glm4v.configuration_glm4v import Glm4vConfig
+    from transformers.models.glm4v.modeling_glm4v import Glm4vForConditionalGeneration
+
+    GLM4V_AVAILABLE = True
+except ImportError:
+    GLM4V_AVAILABLE = False
+
 try:
     from transformers.models.gemma3.configuration_gemma3 import Gemma3TextConfig
     from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
@@ -820,6 +831,73 @@ if GLM4_AVAILABLE:
             attn_implementation="sdpa",  # default value, pytorch native attention
         ),
     )
+if GLM4V_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_glm4v"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_glm4v,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_glm4v,
+        model_class=Glm4vForConditionalGeneration,
+        mini_model_config=Glm4vConfig(
+            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,
+            },
+            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,
@@ -1227,6 +1305,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_glm4v",
+            32,
+            1e-5,
+            torch.bfloat16,
+            1e-2,
+            2e-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_AVAILABLE,
+                    reason="Glm4v 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",