liger-kernel-nightly 0.6.2.dev20251024142419__tar.gz → 0.6.3.dev20251027181634__tar.gz

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251024142419
+Version: 0.6.3.dev20251027181634
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.6.2.dev20251024142419"
+version = "0.6.3.dev20251027181634"
 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.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/src/liger_kernel/transformers/__init__.py

@@ -57,6 +57,7 @@ if TYPE_CHECKING:
     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
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smolvlm  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -120,6 +121,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen3_moe",
         "apply_liger_kernel_to_qwen3_next",
         "apply_liger_kernel_to_smollm3",
+        "apply_liger_kernel_to_smolvlm",
     }
 
     if name in monkey_patch_symbols:
@@ -189,5 +191,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen3_moe",
             "apply_liger_kernel_to_qwen3_next",
             "apply_liger_kernel_to_smollm3",
+            "apply_liger_kernel_to_smolvlm",
         ]
     )

liger_kernel_nightly-0.6.3.dev20251027181634/src/liger_kernel/transformers/model/smolvlm.py

@@ -0,0 +1,158 @@
+from typing import TYPE_CHECKING
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.models.smolvlm.modeling_smolvlm import SmolVLMCausalLMOutputWithPast
+from transformers.processing_utils import Unpack
+from transformers.utils.generic import can_return_tuple
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+if TYPE_CHECKING:
+    from transformers.cache_utils import Cache
+    from transformers.utils.generic import TransformersKwargs
+
+
+# Forward adapted to enable fused Linear + CE without materializing logits.
+# Mirrors the pattern used for other multimodal models (e.g., InternVL, LLaVA).
+@can_return_tuple
+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["Cache"] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    pixel_values: Optional[torch.FloatTensor] = None,
+    pixel_attention_mask: Optional[torch.BoolTensor] = None,
+    image_hidden_states: 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,
+    cache_position: Optional[torch.LongTensor] = None,
+    return_dict: Optional[bool] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,  # Added argument for liger-kernel
+    **lm_kwargs: Unpack["TransformersKwargs"],  # renamed from kwargs
+) -> Union[tuple, SmolVLMCausalLMOutputWithPast]:
+    r"""
+    pixel_attention_mask (`torch.Tensor` of shape `(batch_size, image_size, image_size)`, *optional*):
+        Mask to avoid performing attention on padding pixel indices.
+    image_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+        The hidden states of the image encoder after modality projection.
+    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 `model.image_token_id`. Tokens with indices set to `model.image_token_id` are
+        ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Example:
+
+    ```python
+    >>> import requests
+    >>> import torch
+    >>> from PIL import Image
+    >>> from io import BytesIO
+
+    >>> from transformers import AutoProcessor, AutoModelForImageTextToText
+    >>> from transformers.image_utils import load_image
+
+    >>> # Note that passing the image urls (instead of the actual pil images) to the processor is also possible
+    >>> image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
+    >>> image2 = load_image("https://cdn.britannica.com/59/94459-050-DBA42467/Skyline-Chicago.jpg")
+    >>> image3 = load_image("https://cdn.britannica.com/68/170868-050-8DDE8263/Golden-Gate-Bridge-San-Francisco.jpg")
+
+    >>> processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct")
+    >>> model = AutoModelForImageTextToText.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct", dtype=torch.bfloat16, device_map="auto")
+
+    >>> # Create inputs
+    >>> messages = [
+    ...     {
+    ...         "role": "user",
+    ...         "content": [
+    ...             {"type": "video", "path": path/to/video},
+    ...             {"type": "text", "text": "What is happening in this video?"},
+    ...         ]
+    ...     }
+    ... ]
+
+    >>> inputs = processor.apply_chat_template([messages], add_generation_prompt=True)
+
+    >>> # Generate
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=256)
+    >>> generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+    >>> print(generated_texts)
+    ```"""
+    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,
+        pixel_values=pixel_values,
+        pixel_attention_mask=pixel_attention_mask,
+        image_hidden_states=image_hidden_states,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        return_dict=True,
+        **lm_kwargs,
+    )
+
+    # Copied from llava.py
+    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 = lm_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.text_config.hidden_size,
+            **lm_kwargs,
+        )
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits, labels=labels, vocab_size=self.config.text_config.vocab_size, **lm_kwargs
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return SmolVLMCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=outputs.image_hidden_states,
+    )

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/src/liger_kernel/transformers/monkey_patch.py

@@ -2112,6 +2112,106 @@ def apply_liger_kernel_to_internvl(
             logger.warning(f"{vision_model_name} is not supported by Liger kernel.")
 
 
+def apply_liger_kernel_to_smolvlm(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    layer_norm: bool = True,
+    model: Optional[PreTrainedModel] = None,
+    **kwargs,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace SmolVLM models.
+    Due to the characteristics of SmolVLM, the model must be passed to apply Liger-Kernel's patch to other models connected to SmolVLM.
+    However, if an LM not supported by Liger-Kernel is connected to SmolVLM, unexpected side effects may occur.
+    NOTE: SmolVLM is not available in transformers<4.50.0
+
+    Args:
+        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.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. 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.smolvlm import modeling_smolvlm
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMEncoderLayer
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMForConditionalGeneration
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMModel
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMVisionTransformer
+
+    from liger_kernel.transformers.model.smolvlm import lce_forward as smolvlm_lce_forward
+
+    # Patch LayerNorm for vision model if model is not provided (pre-initialization)
+    if layer_norm and model is None:
+        modeling_smolvlm.nn.LayerNorm = LigerLayerNorm
+
+    if cross_entropy:
+        logger.info("Apply liger 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(smolvlm_lce_forward, model)
+        else:
+            modeling_smolvlm.SmolVLMForConditionalGeneration.forward = smolvlm_lce_forward
+    if rms_norm:
+        modeling_smolvlm.SmolVLMRMSNorm = LigerRMSNorm
+
+    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, SmolVLMForConditionalGeneration):
+            text_model = model.model.text_model
+            vision_model: SmolVLMVisionTransformer = model.model.vision_model
+        elif isinstance(model, SmolVLMModel):
+            text_model = model.text_model
+            vision_model: SmolVLMVisionTransformer = model.vision_model
+        else:
+            raise TypeError(
+                f"Unsupported smolvlm model type. `model` must be `SmolVLMForConditionalGeneration`, `SmolVLMModel`. Got: {type(model)}"
+            )
+
+        text_model_name = model.config.text_config.model_type
+        text_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN.get(text_model_name, None)
+
+        kwargs = {"cross_entropy": False, "fused_linear_cross_entropy": False, **kwargs} | {"rms_norm": rms_norm}
+        if text_liger_fn:
+            accept_params = inspect.signature(text_liger_fn).parameters
+            remain_params = set(kwargs) - (set(accept_params) & set(kwargs))
+            text_kwargs = {k: v for k, v in kwargs.items() if k not in remain_params}
+
+            if remain_params:
+                logger.warning(
+                    f"These parameters are not supported by {text_model_name}. Enter the remaining {list(text_kwargs.keys())} except for {list(remain_params)}\n"
+                    f"Parameters accepted by {text_model_name}: {list(accept_params.keys())}"
+                )
+            text_kwargs["model"] = text_model
+            text_liger_fn(**text_kwargs)
+        elif text_model_name not in MODEL_TYPE_TO_APPLY_LIGER_FN:
+            logger.warning(f"{text_model_name} is not supported by Liger kernel.")
+
+        # Patch vision model LayerNorm layers
+        if layer_norm:
+            # Patch post_layernorm
+            _patch_layer_norm_module(vision_model.post_layernorm)
+
+            # Patch encoder layers
+            for encoder_layer in vision_model.encoder.layers:
+                encoder_layer: SmolVLMEncoderLayer
+                _patch_layer_norm_module(encoder_layer.layer_norm1)
+                _patch_layer_norm_module(encoder_layer.layer_norm2)
+
+
 def apply_liger_kernel_to_falcon_h1(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -2304,6 +2404,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,
     "falcon_h1": apply_liger_kernel_to_falcon_h1,
+    "smolvlm": apply_liger_kernel_to_smolvlm,
 }
 
 

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20251024142419
+Version: 0.6.3.dev20251027181634
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -215,6 +215,7 @@ src/liger_kernel/transformers/model/qwen3.py
 src/liger_kernel/transformers/model/qwen3_moe.py
 src/liger_kernel/transformers/model/qwen3_next.py
 src/liger_kernel/transformers/model/smollm3.py
+src/liger_kernel/transformers/model/smolvlm.py
 src/liger_kernel/transformers/trainer/__init__.py
 src/liger_kernel/transformers/trainer/orpo_trainer.py
 src/liger_kernel/triton/__init__.py
@@ -248,6 +249,7 @@ test/convergence/fp32/test_mini_models_with_logits.py
 test/resources/tiny_shakespeare.txt
 test/resources/fake_configs/Google/Gemma3/gemma-3-4b-it/tokenizer_config.json
 test/resources/fake_configs/Google/Paligemma/paligemma-3b-pt-224/tokenizer_config.json
+test/resources/fake_configs/HuggingFaceTB/SmolVLM2-256M-Video-Instruct/tokenizer_config.json
 test/resources/fake_configs/Llava/llava-1.5-7b-hf/preprocessor_config.json
 test/resources/fake_configs/Llava/llava-1.5-7b-hf/processor_config.json
 test/resources/fake_configs/Llava/llava-1.5-7b-hf/tokenizer_config.json

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -19,6 +19,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_paligemma
 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_smolvlm
 from test.utils import FAKE_CONFIGS_PATH
 from test.utils import UNTOKENIZED_DATASET_PATH
 from test.utils import MiniModelConfig
@@ -39,6 +40,7 @@ from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_Paligemma
 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_smolvlm2
 from test.utils import set_seed
 from test.utils import supports_bfloat16
 from test.utils import train_bpe_tokenizer
@@ -159,6 +161,26 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # SmolVLM2 is only available in transformers>=4.50.0
+    from transformers.models.gpt2.tokenization_gpt2_fast import GPT2TokenizerFast
+    from transformers.models.smolvlm.configuration_smolvlm import SmolVLMConfig
+    from transformers.models.smolvlm.image_processing_smolvlm import SmolVLMImageProcessor
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMForConditionalGeneration
+    from transformers.models.smolvlm.processing_smolvlm import SmolVLMProcessor
+    from transformers.models.smolvlm.video_processing_smolvlm import SmolVLMVideoProcessor
+
+    SMOLVLM2_AVAILABLE = True
+except ImportError:
+    SMOLVLM2_AVAILABLE = False
+
+try:
+    from num2words import num2words  # noqa: F401
+
+    NUM2WORDS_AVAILABLE = True
+except ImportError:
+    NUM2WORDS_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -566,6 +588,44 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if SMOLVLM2_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_smolvlm2"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_smolvlm,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_smolvlm2,
+        model_class=SmolVLMForConditionalGeneration,
+        mini_model_config=SmolVLMConfig(
+            text_config=LlamaConfig(
+                attention_bias=False,
+                attention_dropout=0.0,
+                bos_token_id=1,
+                eos_token_id=2,
+                pad_token_id=2,
+                hidden_act="silu",
+                hidden_size=576,  # 576 for 256M model
+                initializer_range=0.041666666666666664,
+                intermediate_size=1536,  # 1536 for 256M model
+                max_position_embeddings=8192,
+                num_attention_heads=9,  # 9 for 256M model
+                num_hidden_layers=4,  # 30 -> reduced to 4 for testing
+                num_key_value_heads=3,  # 3 for 256M model
+                rms_norm_eps=1e-5,
+                rope_theta=100000,
+                tie_word_embeddings=False,
+                vocab_size=49280,
+            ),
+            vision_config={
+                "hidden_size": 768,
+                "intermediate_size": 3072,
+                "num_hidden_layers": 4,  # 12 -> reduced to 4 for testing
+                "num_attention_heads": 12,
+                "image_size": 512,
+                "patch_size": 16,
+            },
+            image_token_id=49190,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    )
+
 if QWEN2_5_VL_AVAILABLE:
     MINI_MODEL_SETUPS["mini_qwen2_5_vl"] = MiniModelConfig(
         liger_kernel_patch_func=functools.partial(apply_liger_kernel_to_qwen2_5_vl, fused_linear_cross_entropy=False),
@@ -718,6 +778,28 @@ def create_processor(model_name: str):
             image_processor=image_processor, tokenizer=qwen_tokenizer, video_processor=video_processor
         )
 
+    elif model_name == "mini_smolvlm2":
+        tokenizer_config = load_tokenizer_config(
+            os.path.join(FAKE_CONFIGS_PATH, "HuggingFaceTB/SmolVLM2-256M-Video-Instruct/tokenizer_config.json")
+        )
+        tokenizer_base = train_bpe_tokenizer(
+            [
+                token.content
+                for key, token in sorted(
+                    tokenizer_config["added_tokens_decoder"].items(),
+                    key=lambda x: int(x[0]),
+                )
+            ]
+        )
+        gpt2_tokenizer = GPT2TokenizerFast(tokenizer_object=tokenizer_base, **tokenizer_config)
+        image_processor = SmolVLMImageProcessor(size={"longest_edge": 512})
+        video_processor = SmolVLMVideoProcessor()
+
+        # Return proper SmolVLM processor
+        return SmolVLMProcessor(
+            image_processor=image_processor, tokenizer=gpt2_tokenizer, video_processor=video_processor
+        )
+
     elif model_name.startswith("mini_llama4"):
         tokenizer_config = load_tokenizer_config(
             os.path.join(
@@ -1032,6 +1114,29 @@ def run_mini_model_multimodal(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_smolvlm2",
+            32,
+            1e-5,
+            torch.bfloat16,
+            5e-2,
+            5e-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 SMOLVLM2_AVAILABLE,
+                    reason="SmolVLM2 not available in this version of transformers",
+                ),
+                pytest.mark.skipif(
+                    not NUM2WORDS_AVAILABLE,
+                    reason="num2words must be present to run SmolVLMProcessor",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_qwen2_5_vl",
             32,

{liger_kernel_nightly-0.6.2.dev20251024142419 → liger_kernel_nightly-0.6.3.dev20251027181634}/test/convergence/fp32/test_mini_models_multimodal.py

@@ -20,6 +20,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_paligemma
 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_smolvlm
 from test.utils import FAKE_CONFIGS_PATH
 from test.utils import UNTOKENIZED_DATASET_PATH
 from test.utils import MiniModelConfig
@@ -40,6 +41,7 @@ from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_Paligemma
 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_smolvlm2
 from test.utils import set_seed
 from test.utils import train_bpe_tokenizer
 
@@ -155,6 +157,26 @@ try:
 except ImportError:
     INTERNVL_AVAILABLE = False
 
+try:
+    # SmolVLM2 is only available in transformers>=4.50.0
+    from transformers.models.gpt2.tokenization_gpt2_fast import GPT2TokenizerFast
+    from transformers.models.smolvlm.configuration_smolvlm import SmolVLMConfig
+    from transformers.models.smolvlm.image_processing_smolvlm import SmolVLMImageProcessor
+    from transformers.models.smolvlm.modeling_smolvlm import SmolVLMForConditionalGeneration
+    from transformers.models.smolvlm.processing_smolvlm import SmolVLMProcessor
+    from transformers.models.smolvlm.video_processing_smolvlm import SmolVLMVideoProcessor
+
+    SMOLVLM2_AVAILABLE = True
+except ImportError:
+    SMOLVLM2_AVAILABLE = False
+
+try:
+    from num2words import num2words  # noqa: F401
+
+    NUM2WORDS_AVAILABLE = True
+except ImportError:
+    NUM2WORDS_AVAILABLE = False
+
 from liger_kernel.utils import infer_device
 
 device = infer_device()
@@ -564,6 +586,44 @@ if INTERNVL_AVAILABLE:
         ),
     )
 
+if SMOLVLM2_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_smolvlm2"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_smolvlm,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_smolvlm2,
+        model_class=SmolVLMForConditionalGeneration,
+        mini_model_config=SmolVLMConfig(
+            text_config=LlamaConfig(
+                attention_bias=False,
+                attention_dropout=0.0,
+                bos_token_id=1,
+                eos_token_id=2,
+                pad_token_id=2,
+                hidden_act="silu",
+                hidden_size=576,  # 576 for 256M model
+                initializer_range=0.041666666666666664,
+                intermediate_size=1536,  # 1536 for 256M model
+                max_position_embeddings=8192,
+                num_attention_heads=9,  # 9 for 256M model
+                num_hidden_layers=4,  # 30 -> reduced to 4 for testing
+                num_key_value_heads=3,  # 3 for 256M model
+                rms_norm_eps=1e-5,
+                rope_theta=100000,
+                tie_word_embeddings=False,
+                vocab_size=49280,
+            ),
+            vision_config={
+                "hidden_size": 768,
+                "intermediate_size": 3072,
+                "num_hidden_layers": 4,  # 12 -> reduced to 4 for testing
+                "num_attention_heads": 12,
+                "image_size": 512,
+                "patch_size": 16,
+            },
+            image_token_id=49190,
+            attn_implementation="sdpa",  # default value, pytorch native attention
+        ),
+    )
+
 if QWEN2_5_VL_AVAILABLE:
     MINI_MODEL_SETUPS["mini_qwen2_5_vl"] = MiniModelConfig(
         liger_kernel_patch_func=functools.partial(apply_liger_kernel_to_qwen2_5_vl, fused_linear_cross_entropy=False),
@@ -716,6 +776,28 @@ def create_processor(model_name: str):
             image_processor=image_processor, tokenizer=qwen_tokenizer, video_processor=video_processor
         )
 
+    elif model_name == "mini_smolvlm2":
+        tokenizer_config = load_tokenizer_config(
+            os.path.join(FAKE_CONFIGS_PATH, "HuggingFaceTB/SmolVLM2-256M-Video-Instruct/tokenizer_config.json")
+        )
+        tokenizer_base = train_bpe_tokenizer(
+            [
+                token.content
+                for key, token in sorted(
+                    tokenizer_config["added_tokens_decoder"].items(),
+                    key=lambda x: int(x[0]),
+                )
+            ]
+        )
+        gpt2_tokenizer = GPT2TokenizerFast(tokenizer_object=tokenizer_base, **tokenizer_config)
+        image_processor = SmolVLMImageProcessor(size={"longest_edge": 512})
+        video_processor = SmolVLMVideoProcessor()
+
+        # Return proper SmolVLM processor
+        return SmolVLMProcessor(
+            image_processor=image_processor, tokenizer=gpt2_tokenizer, video_processor=video_processor
+        )
+
     elif model_name.startswith("mini_llama4"):
         tokenizer_config = load_tokenizer_config(
             os.path.join(
@@ -1035,6 +1117,28 @@ def run_mini_model_multimodal(
                 reason="InternVL not available in this version of transformers",
             ),
         ),
+        pytest.param(
+            "mini_smolvlm2",
+            32,
+            1e-4,
+            torch.float32,
+            1e-8,
+            1e-5,
+            5e-3,
+            1e-5,
+            5e-3,
+            1e-5,
+            marks=[
+                pytest.mark.skipif(
+                    not SMOLVLM2_AVAILABLE,
+                    reason="SmolVLM2 not available in this version of transformers",
+                ),
+                pytest.mark.skipif(
+                    not NUM2WORDS_AVAILABLE,
+                    reason="num2words must be present to run SmolVLMProcessor",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_qwen2_5_vl",
             32,