[Model] Liger support for SmolLM3 (#798)

## Summary
<!--- This is a required section; please describe the main purpose of
this proposed code change. --->
This PR adds Liger support for the SmolLM3 model.
<!---
## Details
This is an optional section; is there anything specific that reviewers
should be aware of?
--->
The model architecture is similar to that of llama, so I was able to use
most of their ops.

## Testing Done
<!--- This is a required section; please describe how this change was
tested. --->

<!-- 
Replace BLANK with your device type. For example, A100-80G-PCIe

Complete the following tasks before sending your PR, and replace `[ ]`
with
`[x]` to indicate you have done them. 
-->

- Hardware Type: A100-80G-PCIe
- [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: Steven Shimizu <[email protected]>

Author: edbeeching

src/liger_kernel/transformers/__init__.py

@@ -43,6 +43,7 @@
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen2_vl  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_qwen3_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_smollm3  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -100,6 +101,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2_vl",
         "apply_liger_kernel_to_qwen3",
         "apply_liger_kernel_to_qwen3_moe",
+        "apply_liger_kernel_to_smollm3",
     }
 
     if name in monkey_patch_symbols:
@@ -155,5 +157,6 @@ def __getattr__(name: str):
             "apply_liger_kernel_to_qwen2_vl",
             "apply_liger_kernel_to_qwen3",
             "apply_liger_kernel_to_qwen3_moe",
+            "apply_liger_kernel_to_smollm3",
         ]
     )

src/liger_kernel/transformers/model/smollm3.py

@@ -0,0 +1,189 @@
+from typing import TYPE_CHECKING
+from typing import List
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from torch.distributed.fsdp import FullyShardedDataParallel
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.fsdp import _FSDPForwardRedirection
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.utils import PEFT_AVAILABLE
+
+if TYPE_CHECKING:
+    from transformers.cache_utils import Cache
+
+if PEFT_AVAILABLE:
+    from peft.utils.other import ModulesToSaveWrapper
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+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[Union["Cache", 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,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    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.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, Smollm3ForCausalLM
+
+    >>> model = Smollm3ForCausalLM.from_pretrained("HuggingFaceTB/SmolLM3-3B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM3-3B")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> 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]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    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,
+        **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, :]
+
+    shift_labels = kwargs.pop("shift_labels", None)
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if skip_logits and labels is None and shift_labels is None:
+        raise ValueError("skip_logits is True, but labels and shift_labels are None")
+
+    if skip_logits is None:
+        # By default, if in training mode, don't materialize logits
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        loss = lce_maybe_trainable_lm_head(
+            self,
+            hidden_states=kept_hidden_states,
+            hidden_size=self.config.hidden_size,
+            labels=labels,
+            shift_labels=shift_labels,
+            **kwargs,
+        )
+
+    else:
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **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,
+    )
+
+
+def lce_maybe_trainable_lm_head(self, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    lm_head = self.lm_head
+
+    # Unwrap the module if lm_head has been added as trainable module in PEFT LoRA configuration,
+    # i.e. listed in the modules_to_save field of LoraConfig, so the lm_head weights are read
+    # from the unwrapped module.
+    # See https://huggingface.co/docs/peft/package_reference/lora for reference.
+    if PEFT_AVAILABLE and isinstance(lm_head, ModulesToSaveWrapper):
+        lm_head = lm_head.modules_to_save.default
+
+    # If FSDP is used and lm_head is trainable, e.g., during full fine-tuning or with LoRA,
+    # reading the lm_head module weights and calling the kernel must be done within FSDP forward pass
+    # so the module entire parameters are summoned and kept in memory during the kernel execution.
+    if isinstance(lm_head, FullyShardedDataParallel):
+        return _FSDPForwardRedirection()(
+            lm_head,
+            _liger_for_causal_lm_loss,
+            lm_head.module,
+            hidden_states,
+            hidden_size,
+            labels,
+            shift_labels,
+            **loss_kwargs,
+        )
+
+    # FSDP is not used so we can read the lm_head weights and call the kernel directly
+    return _liger_for_causal_lm_loss(
+        lm_head=self.lm_head,
+        hidden_states=hidden_states,
+        hidden_size=hidden_size,
+        labels=labels,
+        shift_labels=shift_labels,
+        **loss_kwargs,
+    )
+
+
+def _liger_for_causal_lm_loss(lm_head, hidden_states, hidden_size, labels, shift_labels, **loss_kwargs):
+    return LigerForCausalLMLoss(
+        hidden_states=hidden_states,
+        lm_head_weight=lm_head.weight,
+        labels=labels,
+        hidden_size=hidden_size,
+        shift_labels=shift_labels,
+        **loss_kwargs,
+    )

src/liger_kernel/transformers/monkey_patch.py

@@ -29,6 +29,7 @@
 from liger_kernel.transformers.model.phi3 import lce_forward_deprecated as phi3_lce_forward_deprecated
 from liger_kernel.transformers.model.qwen2 import lce_forward as qwen2_lce_forward
 from liger_kernel.transformers.model.qwen2 import lce_forward_deprecated as qwen2_lce_forward_deprecated
+from liger_kernel.transformers.model.smollm3 import lce_forward as smollm3_lce_forward
 from liger_kernel.transformers.qwen2vl_mrope import liger_multimodal_rotary_pos_emb
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
@@ -290,6 +291,77 @@ def apply_liger_kernel_to_llama(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_smollm3(
+    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 SmolLM3 model
+
+    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.
+        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.smollm3 import modeling_smollm3
+    from transformers.models.smollm3.modeling_smollm3 import SmolLM3Model
+
+    if rope:
+        modeling_smollm3.apply_rotary_pos_emb = liger_rotary_pos_emb
+    if rms_norm:
+        modeling_smollm3.SmolLM3RMSNorm = LigerRMSNorm
+    if swiglu:
+        modeling_smollm3.SmolLM3MLP = LigerSwiGLUMLP
+
+    if cross_entropy:
+        if transformer_version >= version.parse(SUPPORTED_TRANSFORMER_VERSION):
+            from transformers.loss.loss_utils import nn
+
+            nn.functional.cross_entropy = liger_cross_entropy
+        else:
+            logger.warning(TRANSFORMER_DEPRECATION_WARNING)
+            modeling_smollm3.CrossEntropyLoss = LigerCrossEntropyLoss
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(smollm3_lce_forward, model)
+        else:
+            modeling_smollm3.SmolLM3ForCausalLM.forward = smollm3_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 (e.g. SmolLM3RMSNorm or SmolLM3MLP)
+
+        # get the base model from the model instance
+        base_model: SmolLM3Model = getattr(model, model.base_model_prefix, model)
+
+        if rms_norm:
+            _patch_rms_norm_module(base_model.norm)
+
+        for decoder_layer in base_model.layers:
+            if swiglu:
+                _patch_swiglu_module(decoder_layer.mlp, LigerSwiGLUMLP)
+            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_llava(
     cross_entropy: bool = False,
     fused_linear_cross_entropy: bool = True,
@@ -1801,6 +1873,7 @@ def apply_liger_kernel_to_glm4(
     "qwen2_vl_text": apply_liger_kernel_to_qwen2_vl,
     "qwen2_5_vl": apply_liger_kernel_to_qwen2_5_vl,
     "qwen2_5_vl_text": apply_liger_kernel_to_qwen2_5_vl,
+    "smollm3": apply_liger_kernel_to_smollm3,
     "phi3": apply_liger_kernel_to_phi3,
     "paligemma": apply_liger_kernel_to_paligemma,
 }