Monkeypatch for Qwen2-VL (linkedin#175)

## Summary
Monkeypatch for the recently-published
[Qwen2-VL](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).
HF `transformers` modeling code:
https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py

Feature Request: linkedin#165

## Details
Qwen2-VL in `transformers` is available on `transformers` main but is
yet to be published in a release.

## Testing Done
- Hardware Type: 4090
- [x] run `make test` to ensure correctness
- [x] run `make checkstyle` to ensure code style
- [x] run `make test-convergence` to ensure convergence

---------

Co-authored-by: Shao Tang <tangshao28@gmail.com>

.gitignore

@@ -12,5 +12,8 @@ site/
 build/
 dist/
 
+# Lockfiles
+uv.lock
+
 # Benchmark images
-benchmark/visualizations
+benchmark/visualizations

README.md

@@ -91,7 +91,7 @@ With one line of code, Liger Kernel can increase throughput by more than 20% and
 - **Exact:** Computation is exact—no approximations! Both forward and backward passes are implemented with rigorous unit tests and undergo convergence testing against training runs without Liger Kernel to ensure accuracy.
 - **Lightweight:** Liger Kernel has minimal dependencies, requiring only Torch and Triton—no extra libraries needed! Say goodbye to dependency headaches!
 - **Multi-GPU supported:** Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.).
-- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860) 
+- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860)
 
 ## Target Audiences
 
@@ -227,6 +227,7 @@ loss.backward()
 | Gemma1      | `liger_kernel.transformers.apply_liger_kernel_to_gemma`    | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 | Gemma2      | `liger_kernel.transformers.apply_liger_kernel_to_gemma2`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss         |
 | Qwen2       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2`    | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
+| Qwen2-VL       | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`    | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | Phi3        | `liger_kernel.transformers.apply_liger_kernel_to_phi3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
 
 

src/liger_kernel/transformers/__init__.py

@@ -15,6 +15,7 @@
     apply_liger_kernel_to_mixtral,
     apply_liger_kernel_to_phi3,
     apply_liger_kernel_to_qwen2,
+    apply_liger_kernel_to_qwen2_vl,
 )
 from liger_kernel.transformers.rms_norm import LigerRMSNorm  # noqa: F401
 from liger_kernel.transformers.rope import liger_rotary_pos_emb  # noqa: F401

src/liger_kernel/transformers/model/qwen2_vl.py

@@ -0,0 +1,172 @@
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+    _CONFIG_FOR_DOC,
+    QWEN2_VL_INPUTS_DOCSTRING,
+    Qwen2VLCausalLMOutputWithPast,
+)
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
+
+from liger_kernel.transformers.fused_linear_cross_entropy import (
+    LigerFusedLinearCrossEntropyLoss,
+)
+
+
+@add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=Qwen2VLCausalLMOutputWithPast, 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,
+) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
+    r"""
+    Copy paste Qwen2VL'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, Qwen2VLForConditionalGeneration
+
+    >>> model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2-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.get_dtype())
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw).to(
+                inputs_embeds.device
+            )
+            image_mask = input_ids == self.config.image_token_id
+            if self.training:
+                inputs_embeds = inputs_embeds.clone()
+            inputs_embeds[image_mask] = image_embeds
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.get_dtype())
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw).to(
+                inputs_embeds.device
+            )
+            video_mask = input_ids == self.config.video_token_id
+            inputs_embeds[video_mask] = video_embeds
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    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,
+    )
+
+    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)
+        logits = logits.float()
+        if labels is not None:
+            # 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 Qwen2VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=rope_deltas,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -4,6 +4,7 @@
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
 from liger_kernel.transformers.geglu import LigerGEGLUMLP
+from liger_kernel.transformers.layer_norm import LigerLayerNorm
 from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
 from liger_kernel.transformers.model.mistral import lce_forward as mistral_lce_forward
@@ -245,6 +246,54 @@ def apply_liger_kernel_to_qwen2(
         modeling_qwen2.Qwen2MLP = LigerSwiGLUMLP
 
 
+def apply_liger_kernel_to_qwen2_vl(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    layer_norm: bool = True,
+    swiglu: bool = True,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Qwen2-VL models.
+    NOTE: Qwen2-VL is not available in transformers<=4.44.2
+
+    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.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is True.
+        swiglu (bool): Whether to apply Liger's SwiGLU MLP. Default is True.
+    """
+    assert not (
+        cross_entropy and fused_linear_cross_entropy
+    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
+
+    from transformers.models.qwen2_vl import modeling_qwen2_vl
+
+    from liger_kernel.transformers.model.qwen2_vl import (
+        lce_forward as qwen2_vl_lce_forward,
+    )
+
+    # TODO: Support Qwen2-VL's multimodal RoPE implementation
+
+    if rms_norm:
+        # https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py#L439
+        modeling_qwen2_vl.Qwen2RMSNorm = partial(
+            LigerRMSNorm, init_fn="ones", casting_mode="gemma"
+        )
+    if layer_norm:
+        modeling_qwen2_vl.LayerNorm = LigerLayerNorm
+    if cross_entropy:
+        modeling_qwen2_vl.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_qwen2_vl.Qwen2VLForConditionalGeneration.forward = qwen2_vl_lce_forward
+    if swiglu:
+        modeling_qwen2_vl.Qwen2MLP = LigerSwiGLUMLP
+
+
 def apply_liger_kernel_to_phi3(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -291,6 +340,7 @@ def apply_liger_kernel_to_phi3(
     "mistral": apply_liger_kernel_to_mistral,
     "mixtral": apply_liger_kernel_to_mixtral,
     "qwen2": apply_liger_kernel_to_qwen2,
+    "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "phi3": apply_liger_kernel_to_phi3,
 }
 

test/convergence/test_mini_models.py

@@ -331,6 +331,7 @@ def run_mini_model(
 
     for i in range(num_steps):
         batch = next(loader_iter).to(model.device)
+        optimizer.zero_grad()
         output = model(**batch)
         output.loss.backward()
         optimizer.step()
@@ -343,7 +344,7 @@ def run_mini_model(
 @pytest.mark.parametrize(
     "model_name, num_steps, lr, dtype, loss_atol, loss_rtol, logits_atol, logits_rtol, param_atol, param_rtol",
     [
-        # Gemma 1.1 and 2 has more tolerance because currently, the kernel is not a perfect match (casts are not done the same way)
+        # Gemma 1 has more tolerance because currently, the kernel is not a perfect match (casts are not done the same way)
         ("mini_gemma1", 32, 1e-4, torch.float32, 1e-8, 6e-4, 5e-3, 1e-5, 5e-3, 1e-5),
         pytest.param(
             "mini_gemma1",