liger-kernel-nightly 0.5.6.dev20250402215200__tar.gz → 0.5.6.dev20250403001329__tar.gz

{liger_kernel_nightly-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.6.dev20250402215200
+Version: 0.5.6.dev20250403001329
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -311,6 +311,8 @@ 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         |
+| Gemma3 (Text)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3_text`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Gemma3 (Multimodal)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/README.md

@@ -263,6 +263,8 @@ 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         |
+| Gemma3 (Text)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3_text`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Gemma3 (Multimodal)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.6.dev20250402215200"
+version = "0.5.6.dev20250403001329"
 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-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/src/liger_kernel/transformers/__init__.py

@@ -10,6 +10,8 @@ from liger_kernel.transformers.monkey_patch import _apply_liger_kernel  # noqa:
 from liger_kernel.transformers.monkey_patch import _apply_liger_kernel_to_instance  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma2  # noqa: F401
+from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3  # noqa: F401
+from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gemma3_text  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_granite  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llama  # noqa: F401
 from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_llava  # noqa: F401

liger_kernel_nightly-0.5.6.dev20250403001329/src/liger_kernel/transformers/gema3_rms.py

@@ -0,0 +1,8 @@
+from .rms_norm import LigerRMSNorm
+
+
+class LigerRMSNormForGemma3(LigerRMSNorm):
+    """Gemma3RMSNorm has a dim argument not hidden_size used in q_norm and k_norm."""
+
+    def __init__(self, dim, eps=0.000001, offset=1.0, casting_mode="gemma", init_fn="zeros", in_place=False):
+        super().__init__(dim, eps, offset, casting_mode, init_fn, in_place)

liger_kernel_nightly-0.5.6.dev20250403001329/src/liger_kernel/transformers/model/gemma3.py

@@ -0,0 +1,335 @@
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+import torch.nn as nn
+
+from transformers.cache_utils import Cache
+from transformers.cache_utils import HybridCache
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.gemma3.modeling_gemma3 import _CONFIG_FOR_DOC
+from transformers.models.gemma3.modeling_gemma3 import GEMMA3_INPUTS_DOCSTRING
+from transformers.models.gemma3.modeling_gemma3 import Gemma3CausalLMOutputWithPast
+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
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+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(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+def causal_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[HybridCache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Gemma3ForCausalLM
+
+    >>> model = Gemma3ForCausalLM.from_pretrained("google/gemma-2-9b")
+    >>> tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-9b")
+
+    >>> prompt = "What is your favorite condiment?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # 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]
+    "What is your favorite condiment?"
+    ```"""
+
+    if self.training and self.config._attn_implementation != "eager":
+        logger.warning_once(
+            "It is strongly recommended to train Gemma3 models with the `eager` attention implementation "
+            f"instead of `{self.config._attn_implementation}`. Use `eager` with `AutoModelForCausalLM.from_pretrained('<path-to-checkpoint>', attn_implementation='eager')`."
+        )
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **loss_kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+    kept_hidden_states = hidden_states[:, slice_indices, :]
+    loss = None
+    logits = None
+    if self.training and (labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            softcap=self.config.final_logit_softcapping,
+            **loss_kwargs,
+        )
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if self.config.final_logit_softcapping is not None:
+            logits = logits / self.config.final_logit_softcapping
+            logits = torch.tanh(logits)
+            logits = logits * self.config.final_logit_softcapping
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+def multimodal_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, Gemma3CausalLMOutputWithPast]:
+    r"""
+        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, Gemma3ForConditionalGeneration
+
+    >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/Gemma3-test-224px-hf")
+    >>> processor = AutoProcessor.from_pretrained("google/Gemma3-test-224px-hf")
+
+    >>> prompt = "answer en Where is the cow standing?"
+    >>> url = "https://huggingface.co/gv-hf/Gemma3-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")
+
+    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
+
+    # Replace image id woth PAD if the image token if OOV, to avoid index-errors
+    if input_ids is not None and self.config.image_token_index >= self.vocab_size:
+        special_image_mask = input_ids == self.config.image_token_index
+        llm_input_ids = input_ids.clone()
+        llm_input_ids[special_image_mask] = 0
+    else:
+        llm_input_ids = input_ids
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(llm_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  # Gemma3 positions are 1-indexed
+
+    # Merge text and images
+    if pixel_values is not None:
+        image_features = self.get_image_features(pixel_values)
+
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(self.config.image_token_index, dtype=torch.long, device=inputs_embeds.device)
+            )
+        else:
+            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 = (special_image_mask).sum(dim=1).sum(dim=0)[0]
+            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.text_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 = nn.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 Gemma3CausalLMOutputWithPast(
+        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-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/src/liger_kernel/transformers/monkey_patch.py

@@ -694,6 +694,177 @@ def apply_liger_kernel_to_gemma2(
                 _patch_rms_norm_module_for_gemma2(decoder_layer.post_feedforward_layernorm)
 
 
+def apply_liger_kernel_to_gemma3_text(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    geglu: bool = True,
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Gemma3
+
+    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.
+        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."
+    )
+
+    from transformers.models.gemma3 import modeling_gemma3
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3DecoderLayer
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForCausalLM
+
+    from liger_kernel.transformers.gema3_rms import LigerRMSNormForGemma3
+    from liger_kernel.transformers.model.gemma3 import causal_forward
+
+    _patch_rms_norm_module_for_gemma3 = partial(
+        _patch_rms_norm_module, offset=1.0, casting_mode="gemma", in_place=False
+    )
+
+    if rope:
+        modeling_gemma3.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_gemma3.Gemma3RMSNorm = LigerRMSNormForGemma3
+
+    if geglu:
+        modeling_gemma3.Gemma3MLP = LigerGEGLUMLP
+
+    # Handle loss function
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_gemma3.Gemma3ForCausalLM.forward = causal_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        if isinstance(model, Gemma3ForCausalLM):
+            # get the base model from the model instance
+            base_model = model.model
+
+            if rms_norm:
+                _patch_rms_norm_module_for_gemma3(base_model.norm)
+
+            for decoder_layer in base_model.layers:
+                decoder_layer: Gemma3DecoderLayer
+                if geglu:
+                    _bind_method_to_module(decoder_layer.mlp, "forward", LigerGEGLUMLP.forward)
+                if rms_norm:
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.post_attention_layernorm)
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.pre_feedforward_layernorm)
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.post_feedforward_layernorm)
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.self_attn.q_norm)
+                    _patch_rms_norm_module_for_gemma3(decoder_layer.self_attn.k_norm)
+
+        else:
+            raise TypeError("The model must be Gemma3ForCausalLM.")
+
+
+def apply_liger_kernel_to_gemma3(
+    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 Gemma3
+
+    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."
+    )
+
+    from transformers.models.gemma3 import modeling_gemma3
+    from transformers.models.gemma3.modeling_gemma3 import Gemma3ForConditionalGeneration
+    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.gemma3 import multimodal_forward
+
+    _patch_rms_norm_module_for_gemma3 = partial(
+        _patch_rms_norm_module, offset=1.0, casting_mode="gemma", in_place=False
+    )
+
+    if layer_norm:
+        modeling_siglip.nn.LayerNorm = LigerLayerNorm
+
+    apply_liger_kernel_to_gemma3_text(
+        rope=rope, cross_entropy=False, fused_linear_cross_entropy=False, rms_norm=rms_norm, geglu=geglu
+    )
+
+    if cross_entropy:
+        modeling_gemma3.nn.CrossEntropyLoss = LigerCrossEntropyLoss
+
+    if fused_linear_cross_entropy:
+        modeling_gemma3.Gemma3ForConditionalGeneration.forward = multimodal_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+
+        if isinstance(model, Gemma3ForConditionalGeneration):
+            if isinstance(model.vision_tower, SiglipVisionModel):
+                vision_tower = 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)
+            else:
+                raise TypeError("The vision tower must be SiglipVisionModel")
+
+            if rms_norm:
+                _patch_rms_norm_module_for_gemma3(model.multi_modal_projector.mm_soft_emb_norm)
+
+            apply_liger_kernel_to_gemma3_text(
+                rope=rope,
+                cross_entropy=False,
+                fused_linear_cross_entropy=False,
+                rms_norm=rms_norm,
+                geglu=geglu,
+                model=model.language_model,
+            )
+
+        else:
+            raise TypeError("The model must be Gemma3ForConditionalGeneration.")
+
+
 def apply_liger_kernel_to_paligemma(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1152,6 +1323,8 @@ def apply_liger_kernel_to_olmo2(
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
     "gemma2": apply_liger_kernel_to_gemma2,
+    "gemma3_text": apply_liger_kernel_to_gemma3_text,
+    "gemma3": apply_liger_kernel_to_gemma3,
     "llama": apply_liger_kernel_to_llama,
     "llava": apply_liger_kernel_to_llava,
     "granite": apply_liger_kernel_to_granite,

{liger_kernel_nightly-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.6.dev20250402215200
+Version: 0.5.6.dev20250403001329
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -311,6 +311,8 @@ 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         |
+| Gemma3 (Text)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3_text`   | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy         |
+| Gemma3 (Multimodal)      | `liger_kernel.transformers.apply_liger_kernel_to_gemma3`   | LayerNorm, 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        |

{liger_kernel_nightly-0.5.6.dev20250402215200 → liger_kernel_nightly-0.5.6.dev20250403001329}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -146,6 +146,7 @@ src/liger_kernel/transformers/functional.py
 src/liger_kernel/transformers/fused_linear_cross_entropy.py
 src/liger_kernel/transformers/fused_linear_jsd.py
 src/liger_kernel/transformers/geglu.py
+src/liger_kernel/transformers/gema3_rms.py
 src/liger_kernel/transformers/group_norm.py
 src/liger_kernel/transformers/jsd.py
 src/liger_kernel/transformers/kl_div.py
@@ -161,6 +162,7 @@ src/liger_kernel/transformers/experimental/embedding.py
 src/liger_kernel/transformers/model/__init__.py
 src/liger_kernel/transformers/model/gemma.py
 src/liger_kernel/transformers/model/gemma2.py
+src/liger_kernel/transformers/model/gemma3.py
 src/liger_kernel/transformers/model/llama.py
 src/liger_kernel/transformers/model/llava.py
 src/liger_kernel/transformers/model/loss_utils.py
@@ -203,6 +205,7 @@ 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/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/Llava/llava-1.5-7b-hf/preprocessor_config.json
 test/resources/fake_configs/Llava/llava-1.5-7b-hf/processor_config.json