liger-kernel-nightly 0.5.5.dev20250314002525__tar.gz → 0.5.5.dev20250315175408__tar.gz

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250314002525
+Version: 0.5.5.dev20250315175408
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -310,6 +310,7 @@ loss.backward()
 | Mixtral     | `liger_kernel.transformers.apply_liger_kernel_to_mixtral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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, FusedLinearCrossEntropy         |
+| Paligemma, Paligemma2, & Paligemma2 Mix      | `liger_kernel.transformers.apply_liger_kernel_to_paligemma`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/README.md

@@ -262,6 +262,7 @@ loss.backward()
 | Mixtral     | `liger_kernel.transformers.apply_liger_kernel_to_mixtral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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, FusedLinearCrossEntropy         |
+| Paligemma, Paligemma2, & Paligemma2 Mix      | `liger_kernel.transformers.apply_liger_kernel_to_paligemma`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.5.dev20250314002525"
+version = "0.5.5.dev20250315175408"
 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-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel/chunked_loss/fused_linear_distillation.py

@@ -117,7 +117,7 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
 
         hard_loss /= full_target.shape[0]
 
-        soft_loss = distillation_loss_fn(student_logits_chunk, teacher_logits_chunk)
+        soft_loss = distillation_loss_fn(student_logits_chunk, teacher_logits_chunk, **loss_kwargs)
         soft_loss /= full_target.shape[0]
 
         loss = weight_hard_loss * hard_loss + weight_soft_loss * soft_loss
@@ -180,9 +180,9 @@ class LigerFusedLinearDistillationBase(torch.autograd.Function):
             ignore_index=ignore_index,
             weight_hard_loss=weight_hard_loss,
             weight_soft_loss=weight_soft_loss,
-            beta=beta,
             compute_ce_loss=compute_ce_loss,
             temperature=temperature,
+            beta=beta,
             **loss_kwargs,
         )
 

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel/chunked_loss/jsd_loss.py

@@ -19,15 +19,20 @@ class LigerFusedLinearJSDFunction(LigerFusedLinearDistillationBase):
         student_log_probs = F.log_softmax(student_logits, dim=-1)
         teacher_log_probs = F.log_softmax(teacher_logits, dim=-1)
 
-        # Compute probabilities (only required for mean calculation)
-        mean_probs = beta * student_log_probs.exp() + (1 - beta) * teacher_log_probs.exp()
-        log_mean_probs = mean_probs.log()
+        if beta == 0:
+            jsd_loss = F.kl_div(student_log_probs, teacher_log_probs, reduction="sum", log_target=True)
+        elif beta == 1:
+            jsd_loss = F.kl_div(teacher_log_probs, student_log_probs, reduction="sum", log_target=True)
+        else:
+            # Compute probabilities (only required for mean calculation)
+            mean_probs = (1 - beta) * student_log_probs.exp() + beta * teacher_log_probs.exp()
+            log_mean_probs = mean_probs.log()
 
-        student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
-        teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
+            student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="sum", log_target=True)
+            teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="sum", log_target=True)
 
-        # JSD is the weighted average of the KL divergences
-        jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
+            # JSD is the weighted average of the KL divergences
+            jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
         return jsd_loss
 
     @classmethod

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel/ops/jsd.py

@@ -51,24 +51,43 @@ def _jsd_kernel(
         Y = tl.load(Y_ptr + offsets, mask=mask, other=float("-inf")).to(tl.float32)
 
         if beta == 0.0:  # forward KL
-            Y_prob = tl.exp(Y)
+            Y_max = tl.max(Y, axis=0)
+            Y_shifted = Y - Y_max
+            Y_prob = tl.exp(Y_shifted) * tl.exp(Y_max)  # Compensate for the shift
             loss = Y_prob * (Y - X)
             dX = -Y_prob
-        elif beta == 1.0:
-            X_prob = tl.exp(X)
+        elif beta == 1.0:  # reverse KL
+            X_max = tl.max(X, axis=0)
+            X_shifted = X - X_max
+            X_prob = tl.exp(X_shifted) * tl.exp(X_max)  # Compensate for the shift
             loss = X_prob * (X - Y)
             dX = loss + X_prob
         else:
-            Q = tl.exp(X)
-            P = tl.exp(Y)
-            M = beta * P + (1 - beta) * Q
-            log_M = tl.log(M)
+            max_val = tl.maximum(tl.max(X, axis=0), tl.max(Y, axis=0))
+            X_shifted = X - max_val
+            Y_shifted = Y - max_val
 
-            loss = beta * P * Y + (1 - beta) * Q * X - M * log_M
-            dX = (1 - beta) * Q * (X - log_M)
+            # Pre-compute exp(max_val) since it's used twice
+            exp_max = tl.exp(max_val)
+
+            # Compute exp terms with compensation
+            Q = tl.exp(X_shifted) * exp_max  # = exp(X)
+            P = tl.exp(Y_shifted) * exp_max  # = exp(Y)
+
+            # Pre-compute common terms
+            beta_P = beta * P
+            one_minus_beta_Q = (1 - beta) * Q
+            M = beta_P + one_minus_beta_Q
+            log_M = tl.log(M)  # No need to compensate as M is already in original scale
+
+            loss = beta_P * Y + one_minus_beta_Q * X - M * log_M
+            dX = one_minus_beta_Q * (X - log_M)
+
+        # Pre-compute scaling factor
+        scale = 1.0 / n_non_ignore
+        loss = loss * scale
+        dX = dX * scale
 
-        loss = loss / n_non_ignore
-        dX = dX / n_non_ignore
         tl.store(loss_ptr + offsets, loss, mask=mask)
         tl.store(dX_ptr + offsets, dX, mask=mask)
 

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel/transformers/__init__.py

@@ -15,6 +15,7 @@ from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mistral
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mixtral  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_mllama  # noqa: F401
 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_paligemma  # 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

liger_kernel_nightly-0.5.5.dev20250315175408/src/liger_kernel/transformers/model/paligemma.py

@@ -0,0 +1,213 @@
+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.cache_utils import Cache
+from transformers.models.paligemma.modeling_paligemma import _CONFIG_FOR_DOC
+from transformers.models.paligemma.modeling_paligemma import PALIGEMMA_INPUTS_DOCSTRING
+from transformers.models.paligemma.modeling_paligemma import PaliGemmaCausalLMOutputWithPast
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import is_torchdynamo_compiling
+from transformers.utils import logging
+from transformers.utils import replace_return_docstrings
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.fused_linear_cross_entropy import LigerFusedLinearCrossEntropyLoss
+
+logger = logging.get_logger(__name__)
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(PALIGEMMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=PaliGemmaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
+    token_type_ids: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = 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,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **lm_kwargs,
+) -> Union[Tuple, PaliGemmaCausalLMOutputWithPast]:
+    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.text_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.text_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
+    >>> import requests
+    >>> from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
+
+    >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/PaliGemma-test-224px-hf")
+    >>> processor = AutoProcessor.from_pretrained("google/PaliGemma-test-224px-hf")
+
+    >>> prompt = "answer en Where is the cow standing?"
+    >>> url = "https://huggingface.co/gv-hf/PaliGemma-test-224px-hf/resolve/main/cow_beach_1.png"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_length=30)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "answer en Where is the cow standing?\nbeach"
+    ```"""
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    if pixel_values is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both pixel_values and inputs_embeds at the same time, and must specify either one"
+        )
+
+    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
+
+    is_training = token_type_ids is not None and labels is not None
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(input_ids)
+
+    if cache_position is None:
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        cache_position = torch.arange(
+            past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+        )
+
+    if position_ids is None:
+        position_ids = cache_position.unsqueeze(0) + 1  # Paligemma positions are 1-indexed
+
+    # Merge text and images
+    if pixel_values is not None:
+        image_features = self.get_image_features(pixel_values)
+
+        special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(-1)
+        special_image_mask = special_image_mask.expand_as(inputs_embeds).to(inputs_embeds.device)
+        if not is_torchdynamo_compiling() and inputs_embeds[special_image_mask].numel() != image_features.numel():
+            image_tokens_in_text = torch.sum(input_ids == self.config.image_token_index)
+            raise ValueError(
+                f"Number of images does not match number of special image tokens in the input text. "
+                f"Got {image_tokens_in_text} image tokens in the text but {image_features.shape[0] * image_features.shape[1]} "
+                "tokens from image embeddings."
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+    # mask out pad-token-ids in labels for BC
+    if labels is not None and self.pad_token_id in labels:
+        logger.warning_once(
+            "`labels` contains `pad_token_id` which will be masked with `config.ignore_index`. "
+            "You have to mask out `pad_token_id` when preparing `labels`, this behavior will be removed in v.4.46.",
+        )
+        labels = torch.where(input_ids == self.pad_token_id, self.config.ignore_index, labels)
+
+    causal_mask = self._update_causal_mask(
+        attention_mask, token_type_ids, past_key_values, cache_position, inputs_embeds, is_training
+    )
+
+    outputs = self.language_model.model(
+        attention_mask=causal_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,
+        logits_to_keep=logits_to_keep,
+        **lm_kwargs,
+    )
+
+    hidden_states = outputs[0]
+
+    loss = None
+    logits = None
+
+    if self.training and (labels is not None):
+        shift_hidden_states = hidden_states[..., :-1, :]
+        shift_labels = labels[..., 1:]
+
+        hidden_device = shift_hidden_states.device
+
+        if attention_mask is not None:
+            # we use the input attention mask to shift the hidden_states and labels, because it is 2D.
+            # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+            shift_attention_mask = attention_mask[:, -shift_hidden_states.shape[1] :].to(hidden_device)
+            shift_hidden_states = shift_hidden_states[shift_attention_mask.to(hidden_device) != 0].contiguous()
+            shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+        else:
+            shift_hidden_states = shift_hidden_states.contiguous()
+            shift_labels = shift_labels.contiguous()
+
+        # Flatten hidden state
+        shift_hidden_states = shift_hidden_states.view(-1, self.config.hidden_size)
+        shift_labels = shift_labels.view(-1).to(hidden_device)
+
+        lce = LigerFusedLinearCrossEntropyLoss()
+        loss = lce(self.language_model.lm_head.weight, shift_hidden_states, shift_labels)
+    else:
+        logits = self.language_model.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_logits = logits[..., :-1, :]
+            shift_labels = labels[..., 1:]
+            if attention_mask is not None:
+                # we use the input attention mask to shift the logits and labels, because it is 2D.
+                # we also crop attn mask in case it is longer, which happens in PrefixTuning with peft
+                shift_attention_mask = attention_mask[:, -shift_logits.shape[1] :].to(logits.device)
+                shift_logits = shift_logits[shift_attention_mask.to(logits.device) != 0].contiguous()
+                shift_labels = shift_labels[shift_attention_mask.to(shift_labels.device) != 0].contiguous()
+            else:
+                shift_logits = shift_logits.contiguous()
+                shift_labels = shift_labels.contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+
+            flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
+            flat_labels = shift_labels.view(-1).to(shift_logits.device)
+            loss = loss_fct(flat_logits, flat_labels)
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return PaliGemmaCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel/transformers/monkey_patch.py

@@ -600,6 +600,90 @@ def apply_liger_kernel_to_gemma2(
                 _patch_rms_norm_module_for_gemma2(decoder_layer.post_feedforward_layernorm)
 
 
+def apply_liger_kernel_to_paligemma(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    layer_norm: bool = True,
+    rms_norm: bool = True,
+    geglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace PaliGemma
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is True.
+        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.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is True.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is True.
+        geglu (bool): Whether to apply Liger's GeGLU 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."
+    )
+
+    # PaliGemma submodules are ['vision_tower', 'multi_modal_projector', 'language_model']
+
+    from transformers.models.gemma2.modeling_gemma2 import Gemma2ForCausalLM
+    from transformers.models.paligemma import modeling_paligemma
+    from transformers.models.paligemma.modeling_paligemma import PaliGemmaForConditionalGeneration
+    from transformers.models.siglip import modeling_siglip
+    from transformers.models.siglip.modeling_siglip import SiglipEncoderLayer
+    from transformers.models.siglip.modeling_siglip import SiglipVisionModel
+
+    from liger_kernel.transformers.model.paligemma import lce_forward
+
+    # The vision_tower is a SiglipVisionModel
+    if layer_norm:
+        modeling_siglip.nn.LayerNorm = LigerLayerNorm
+
+    # SiglipMLP is standard FFN so LigerGEGLUMLP is not compatible
+    # The multi_modal_projector is Linear, nothing to do
+
+    # The language_model is Gemma2ForCausalLM
+    apply_liger_kernel_to_gemma2(rope=rope, cross_entropy=False, fused_linear_cross_entropy=False, geglu=geglu)
+    # Handle loss function
+    if cross_entropy:
+        modeling_paligemma.nn.CrossEntropyLoss = LigerCrossEntropyLoss
+    if fused_linear_cross_entropy:
+        modeling_paligemma.PaliGemmaForConditionalGeneration.forward = 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 not isinstance(model, PaliGemmaForConditionalGeneration):
+            raise TypeError("model have to be of type PaliGemmaForConditionalGeneration")
+
+        vision_tower: SiglipVisionModel = model.vision_tower
+
+        _patch_layer_norm_module(vision_tower.vision_model.post_layernorm)
+
+        for layer in vision_tower.vision_model.encoder.layers:
+            layer: SiglipEncoderLayer
+            if layer_norm:
+                _patch_layer_norm_module(layer.layer_norm1)
+                _patch_layer_norm_module(layer.layer_norm2)
+
+        language_model: Gemma2ForCausalLM = model.language_model
+
+        apply_liger_kernel_to_gemma2(
+            rope=rope,
+            cross_entropy=False,
+            fused_linear_cross_entropy=False,
+            rms_norm=rms_norm,
+            geglu=geglu,
+            model=language_model,
+        )
+
+
 def apply_liger_kernel_to_qwen2(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -959,6 +1043,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "qwen2_vl": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "phi3": apply_liger_kernel_to_phi3,
+    "paligemma": apply_liger_kernel_to_paligemma,
 }
 
 

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.5.dev20250314002525
+Version: 0.5.5.dev20250315175408
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -310,6 +310,7 @@ loss.backward()
 | Mixtral     | `liger_kernel.transformers.apply_liger_kernel_to_mixtral`  | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy        |
 | 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, FusedLinearCrossEntropy         |
+| Paligemma, Paligemma2, & Paligemma2 Mix      | `liger_kernel.transformers.apply_liger_kernel_to_paligemma`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -163,6 +163,7 @@ src/liger_kernel/transformers/model/mistral.py
 src/liger_kernel/transformers/model/mixtral.py
 src/liger_kernel/transformers/model/mllama.py
 src/liger_kernel/transformers/model/olmo2.py
+src/liger_kernel/transformers/model/paligemma.py
 src/liger_kernel/transformers/model/phi3.py
 src/liger_kernel/transformers/model/qwen2.py
 src/liger_kernel/transformers/model/qwen2_5_vl.py
@@ -197,6 +198,7 @@ test/convergence/fp32/test_mini_models.py
 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/Google/Paligemma/paligemma-3b-pt-224/tokenizer_config.json
 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

{liger_kernel_nightly-0.5.5.dev20250314002525 → liger_kernel_nightly-0.5.5.dev20250315175408}/test/chunked_loss/test_jsd_loss.py

@@ -27,7 +27,6 @@ class HFJSDLoss(HFDistillationLoss):
         ignore_index: int = -100,
         weight_hard_loss: float = 0.5,
         weight_soft_loss: float = 0.5,
-        beta: float = 0.5,
     ):
         super().__init__(
             ignore_index=ignore_index,
@@ -35,7 +34,6 @@ class HFJSDLoss(HFDistillationLoss):
             weight_soft_loss=weight_soft_loss,
             temperature=temperature,
         )
-        self.beta = (beta,)
 
     def distillation_loss(self, student_logits, teacher_logits, beta=0.5):
         """
@@ -50,15 +48,20 @@ class HFJSDLoss(HFDistillationLoss):
         student_log_probs = F.log_softmax(student_logits, dim=-1)
         teacher_log_probs = F.log_softmax(teacher_logits, dim=-1)
 
-        # Compute probabilities (only required for mean calculation)
-        mean_probs = beta * student_log_probs.exp() + (1 - beta) * teacher_log_probs.exp()
-        log_mean_probs = mean_probs.log()
+        if beta == 0:
+            jsd_loss = F.kl_div(student_log_probs, teacher_log_probs, reduction="none", log_target=True)
+        elif beta == 1:
+            jsd_loss = F.kl_div(teacher_log_probs, student_log_probs, reduction="none", log_target=True)
+        else:
+            # Compute probabilities (only required for mean calculation)
+            mean_probs = (1 - beta) * student_log_probs.exp() + beta * teacher_log_probs.exp()
+            log_mean_probs = mean_probs.log()
 
-        student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="batchmean", log_target=True)
-        teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="batchmean", log_target=True)
+            student_kl = F.kl_div(log_mean_probs, student_log_probs, reduction="batchmean", log_target=True)
+            teacher_kl = F.kl_div(log_mean_probs, teacher_log_probs, reduction="batchmean", log_target=True)
 
-        # JSD is the weighted average of the KL divergences
-        jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
+            # JSD is the weighted average of the KL divergences
+            jsd_loss = beta * teacher_kl + (1 - beta) * student_kl
         return jsd_loss
 
 
@@ -88,12 +91,12 @@ class TorchLMHeadJSD(torch.nn.Module):
         # smaller student model weights
         self.student_lin = torch.nn.Linear(in_features=H // 2, out_features=V, bias=bias, dtype=dtype, device=device)
         self.teacher_lin = torch.nn.Linear(in_features=H, out_features=V, bias=bias, dtype=dtype, device=device)
+        self.beta = beta
         self.jsd = HFJSDLoss(
             ignore_index=ignore_index,
             weight_hard_loss=weight_hard_loss,
             weight_soft_loss=weight_soft_loss,
             temperature=temperature,
-            beta=beta,
         ).get_batch_loss_metrics
 
     def forward(self, student_input, teacher_input, target):
@@ -105,6 +108,7 @@ class TorchLMHeadJSD(torch.nn.Module):
             target,
             self.student_lin.bias,
             self.teacher_lin.bias,
+            beta=self.beta,
         )
         return jsd_loss
 
@@ -132,6 +136,7 @@ class LigerLMHeadJSD(torch.nn.Module):
             weight_soft_loss=weight_soft_loss,
             ignore_index=ignore_index,
             temperature=temperature,
+            beta=beta,
         )
 
     def forward(self, student_input, teacher_input, target):