liger-kernel-nightly 0.5.4.dev20250225015144__tar.gz → 0.5.4.dev20250227064037__tar.gz

{liger_kernel_nightly-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.4.dev20250225015144
+Version: 0.5.4.dev20250227064037
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -312,6 +312,7 @@ loss.backward()
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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        |
 | 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-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/README.md

@@ -264,6 +264,7 @@ loss.backward()
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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        |
 | 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-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.4.dev20250225015144"
+version = "0.5.4.dev20250227064037"
 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-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/src/liger_kernel/ops/cross_entropy.py

@@ -285,6 +285,10 @@ def cross_entropy_forward(
 
     target_mask = target != ignore_index
     n_non_ignore = target_mask.sum().item()
+    assert (target * target_mask).max() < _input.shape[-1], (
+        f"Target {target.max()} is out of bounds. Expected < {_input.shape[-1]}"
+    )
+    assert (target * target_mask).min() >= 0, f"Target {target.min()} is out of bounds. Expected >= 0"
     sum_non_ignore_weight = n_non_ignore
     weight_sum = 0.0
     if weight is not None:

{liger_kernel_nightly-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/src/liger_kernel/transformers/__init__.py

@@ -17,6 +17,7 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_olmo2  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_phi3  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2  # noqa: F401
+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.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401

liger_kernel_nightly-0.5.4.dev20250227064037/src/liger_kernel/transformers/model/qwen2_5_vl.py

@@ -0,0 +1,205 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import _CONFIG_FOR_DOC
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import QWEN2_5_VL_INPUTS_DOCSTRING
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLCausalLMOutputWithPast
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+
+
+@add_start_docstrings_to_model_forward(QWEN2_5_VL_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=Qwen2_5_VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+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,
+    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,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+    r"""
+    Copy paste Qwen2_5_VL's forward but replace torch cross entropy with liger fused linear cross entropy
+    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]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # 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]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+    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
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            image_mask = mask_expanded.to(inputs_embeds.device)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            video_mask = mask_expanded.to(inputs_embeds.device)
+
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (cache_position is not None and cache_position[0] == 0) or self.rope_deltas is None:
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                attention_mask,
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (cache_position[0] + self.rope_deltas).to(inputs_embeds.device) if cache_position is not None else 0
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+            position_ids = position_ids.add(delta)
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        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,
+    )
+
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    if self.training and (labels is not None):
+        shift_hidden_states = hidden_states[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+
+        # Flatten tokens
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1)
+
+        lce = LigerFusedLinearCrossEntropyLoss()
+        loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
+    else:
+        logits = self.lm_head(hidden_states)
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2_5_VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=rope_deltas,
+    )

{liger_kernel_nightly-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/src/liger_kernel/transformers/monkey_patch.py

@@ -745,6 +745,73 @@ def apply_liger_kernel_to_qwen2_vl(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_qwen2_5_vl(
+    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 Qwen2.5-VL models.
+    NOTE: Qwen2.5-VL is not available in transformers<4.48.2
+
+    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.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. 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.qwen2_5_vl import modeling_qwen2_5_vl
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLModel
+
+    from liger_kernel.transformers.model.qwen2_5_vl import lce_forward as qwen2_5_vl_lce_forward
+
+    if rope:
+        modeling_qwen2_5_vl.apply_multimodal_rotary_pos_emb = liger_multimodal_rotary_pos_emb
+    if rms_norm:
+        modeling_qwen2_5_vl.Qwen2RMSNorm = LigerRMSNorm
+    if cross_entropy:
+        modeling_qwen2_5_vl.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = qwen2_5_vl_lce_forward
+    if swiglu:
+        modeling_qwen2_5_vl.Qwen2MLP = LigerSwiGLUMLP
+
+    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: Qwen2_5_VLModel = getattr(model, model.base_model_prefix, model)
+
+        if hasattr(model, "visual"):
+            # Patch Qwen2_5_VisionTransformerPretrainedModel
+            for vision_block in model.visual.blocks:
+                if rms_norm:
+                    _patch_rms_norm_module(vision_block.norm1)
+                    _patch_rms_norm_module(vision_block.norm2)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _bind_method_to_module(decoder_layer.mlp, "forward", LigerSwiGLUMLP.forward)
+            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_phi3(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -890,6 +957,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
     "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-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.4.dev20250225015144
+Version: 0.5.4.dev20250227064037
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -312,6 +312,7 @@ loss.backward()
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Qwen2, Qwen2.5, & QwQ      | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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        |
 | 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-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -165,6 +165,7 @@ src/liger_kernel/transformers/model/mllama.py
 src/liger_kernel/transformers/model/olmo2.py
 src/liger_kernel/transformers/model/phi3.py
 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/trainer/__init__.py
 src/liger_kernel/transformers/trainer/orpo_trainer.py
@@ -197,6 +198,7 @@ test/convergence/fp32/test_mini_models_multimodal.py
 test/convergence/fp32/test_mini_models_with_logits.py
 test/resources/tiny_shakespeare.txt
 test/resources/fake_configs/Qwen/Qwen2-VL-7B-Instruct/tokenizer_config.json
+test/resources/fake_configs/Qwen/Qwen2.5-VL-7B-Instruct/tokenizer_config.json
 test/resources/fake_configs/meta-llama/Llama-3.2-11B-Vision-Instruct/tokenizer_config.json
 test/resources/scripts/generate_tokenized_dataset.py
 test/resources/tiny_shakespeare_tokenized/data-00000-of-00001.arrow

{liger_kernel_nightly-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/test/convergence/bf16/test_mini_models.py

@@ -28,6 +28,7 @@ from liger_kernel.transformers import apply_liger_kernel_to_mllama
 from liger_kernel.transformers import apply_liger_kernel_to_olmo2
 from liger_kernel.transformers import apply_liger_kernel_to_phi3
 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 test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
@@ -42,6 +43,7 @@ from test.utils import revert_liger_kernel_to_mllama
 from test.utils import revert_liger_kernel_to_olmo2
 from test.utils import revert_liger_kernel_to_phi3
 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 set_seed
 from test.utils import simple_collate_fn
@@ -65,6 +67,15 @@ try:
 except ImportError:
     QWEN2_VL_AVAILABLE = False
 
+try:
+    # Qwen2.5-VL is only available in transformers>4.48.2
+    from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLConfig
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
+
+    QWEN2_5_VL_AVAILABLE = True
+except ImportError:
+    QWEN2_5_VL_AVAILABLE = False
+
 try:
     from transformers.models.granite import GraniteConfig
     from transformers.models.granite import GraniteForCausalLM
@@ -402,6 +413,62 @@ if QWEN2_VL_AVAILABLE:
         ),
     )
 
+if QWEN2_5_VL_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_qwen2_5_vl"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen2_5_vl,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen2_5_vl,
+        model_class=Qwen2_5_VLForConditionalGeneration,
+        mini_model_config=Qwen2_5_VLConfig(
+            attention_dropout=0.0,
+            # bos and eos set to match the Mistral-7B tokenizer used to create the test dataset
+            # https://huggingface.co/mistralai/Mistral-7B-v0.1/blob/main/config.json
+            bos_token_id=1,  # 151643
+            eos_token_id=2,  # 151645
+            vision_start_token_id=32765,  # vocab_size - 5
+            vision_end_token_id=32766,  # vocab_size - 4
+            vision_token_id=32767,  # vocab_size - 3
+            image_token_id=32768,  # vocab_size - 2
+            video_token_id=32769,  # vocab_size - 1
+            hidden_act="silu",
+            hidden_size=1536,  # 8192
+            initializer_range=0.02,
+            intermediate_size=4864,  # 29568
+            max_position_embeddings=32768,
+            max_window_layers=4,  # 80
+            num_attention_heads=12,  # 64
+            num_hidden_layers=4,  # 80
+            num_key_value_heads=2,  # 8
+            rms_norm_eps=1e-6,  # 1e-5
+            rope_theta=1000000.0,
+            rope_scaling=dict(
+                type="mrope",
+                mrope_section=[16, 24, 24],  # (temporal, height, width)
+            ),
+            sliding_window=4096,
+            tie_word_embeddings=False,
+            use_cache=True,
+            vocab_size=32768,  # 152064  # >32k, Mistral-7B tokenizer vocab size
+            use_sliding_window=False,
+            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,
+                "spatial_patch_size": 14,
+                "window_size": 112,
+                "fullatt_block_indexes": [7, 15, 23, 31],
+                "tokens_per_second": 2,
+                "temporal_patch_size": 2,
+            },
+            attn_implementation="sdpa",
+        ),
+    )
+
 if GRANITE_AVAILABLE:
     MINI_MODEL_SETUPS["mini_granite3"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_granite,
@@ -626,6 +693,25 @@ def run_mini_model(
                 ),
             ],
         ),
+        pytest.param(
+            "mini_qwen2_5_vl",
+            32,
+            1e-4,
+            torch.bfloat16,
+            1e-3,
+            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 QWEN2_5_VL_AVAILABLE,
+                    reason="Qwen2.5-VL not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_phi3",
             32,

{liger_kernel_nightly-0.5.4.dev20250225015144 → liger_kernel_nightly-0.5.4.dev20250227064037}/test/convergence/bf16/test_mini_models_multimodal.py

@@ -9,6 +9,7 @@ from torch.utils.data import DataLoader
 from transformers import PreTrainedTokenizerFast
 
 from liger_kernel.transformers import apply_liger_kernel_to_mllama
+from liger_kernel.transformers import apply_liger_kernel_to_qwen2_5_vl
 from liger_kernel.transformers import apply_liger_kernel_to_qwen2_vl
 from test.utils import FAKE_CONFIGS_PATH
 from test.utils import UNTOKENIZED_DATASET_PATH
@@ -17,6 +18,7 @@ from test.utils import assert_verbose_allclose
 from test.utils import load_tokenizer_config
 from test.utils import multimodal_collate_fn
 from test.utils import revert_liger_kernel_to_mllama
+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 set_seed
 from test.utils import supports_bfloat16
@@ -34,6 +36,18 @@ try:
 except ImportError:
     QWEN2_VL_AVAILABLE = False
 
+try:
+    # Qwen2.5-VL is only available in transformers>4.48.2
+    from transformers.models.qwen2.tokenization_qwen2_fast import Qwen2TokenizerFast
+    from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import Qwen2_5_VLConfig
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLForConditionalGeneration
+    from transformers.models.qwen2_5_vl.processing_qwen2_5_vl import Qwen2_5_VLProcessor
+    from transformers.models.qwen2_vl.image_processing_qwen2_vl import Qwen2VLImageProcessor
+
+    QWEN2_5_VL_AVAILABLE = True
+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
@@ -169,6 +183,52 @@ if QWEN2_VL_AVAILABLE:
         ),
     )
 
+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),
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen2_5_vl,
+        model_class=Qwen2_5_VLForConditionalGeneration,
+        mini_model_config=Qwen2_5_VLConfig(
+            attention_dropout=0.0,
+            # Token Ids and vocab size must match those in the tokenizer/processor
+            # test/resources/fake_configs/Qwen/Qwen2-VL-7B-Instruct/tokenizer_config.json
+            bos_token_id=0,
+            eos_token_id=0,
+            vision_start_token_id=1,
+            vision_end_token_id=2,
+            vision_token_id=3,
+            image_token_id=4,
+            video_token_id=5,
+            hidden_act="silu",
+            hidden_size=1024,  # 8192
+            initializer_range=0.02,
+            intermediate_size=1024,  # 29568
+            max_position_embeddings=32768,
+            max_window_layers=4,  # 80
+            num_attention_heads=8,  # 64
+            num_hidden_layers=4,  # 80
+            num_key_value_heads=2,  # 8
+            rms_norm_eps=1e-6,  # 1e-5
+            rope_theta=1000000.0,
+            rope_scaling=dict(
+                type="mrope",
+                mrope_section=[16, 24, 24],  # (temporal, height, width)
+            ),
+            sliding_window=4096,
+            tie_word_embeddings=True,
+            use_cache=False,  # True
+            vocab_size=32000,  # 152064,
+            use_sliding_window=False,
+            vision_config={
+                "depth": 4,  # 32
+                "hidden_size": 128,  # 1280
+                "num_heads": 16,
+                "in_chans": 3,
+            },
+            attn_implementation="sdpa",
+        ),
+    )
+
 
 def create_processor(model_name):
     if model_name == "mini_qwen2_vl":
@@ -188,6 +248,23 @@ def create_processor(model_name):
         image_processor = Qwen2VLImageProcessor()
         return Qwen2VLProcessor(image_processor=image_processor, tokenizer=qwen_tokenizer)
 
+    elif model_name == "mini_qwen2_5_vl":
+        tokenizer_config = load_tokenizer_config(
+            os.path.join(FAKE_CONFIGS_PATH, "Qwen/Qwen2.5-VL-7B-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]),
+                )
+            ]
+        )
+        qwen_tokenizer = Qwen2TokenizerFast(tokenizer_object=tokenizer_base, **tokenizer_config)
+        image_processor = Qwen2VLImageProcessor()
+        return Qwen2_5_VLProcessor(image_processor=image_processor, tokenizer=qwen_tokenizer)
+
     elif model_name == "mini_mllama":
         tokenizer_config = load_tokenizer_config(
             os.path.join(
@@ -297,9 +374,11 @@ def run_mini_model_multimodal(
             "rope": True,
             "rms_norm": True,
             "cross_entropy": False,
-            "layer_norm": True,
         }
 
+        if "qwen2_5_vl" not in model_name:
+            kwargs["layer_norm"] = True
+
         if "gemma" in model_name:
             kwargs["geglu"] = True
         else:
@@ -355,6 +434,26 @@ def run_mini_model_multimodal(
                 pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
             ],
         ),
+        pytest.param(
+            "mini_qwen2_5_vl",
+            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 QWEN2_5_VL_AVAILABLE,
+                    reason="Qwen2.5-VL not available in this version of transformers",
+                ),
+                pytest.mark.skipif(device == "xpu", reason="skip for XPU"),
+            ],
+        ),
         pytest.param(
             "mini_mllama",
             32,