liger-kernel-nightly 0.5.8.dev20250502223102__tar.gz → 0.5.8.dev20250503025537__tar.gz

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.8.dev20250502223102
+Version: 0.5.8.dev20250503025537
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -317,6 +317,7 @@ loss.backward()
 | 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        |
+| Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, 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-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/README.md

@@ -269,6 +269,7 @@ loss.backward()
 | 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        |
+| Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, 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-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "liger_kernel_nightly"
-version = "0.5.8.dev20250502223102"
+version = "0.5.8.dev20250503025537"
 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-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/__init__.py

@@ -39,6 +39,7 @@ if TYPE_CHECKING:
     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.monkey_patch import apply_liger_kernel_to_qwen3  # noqa: F401
 
 
 # Check if 'transformers' is installed
@@ -93,6 +94,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_qwen2",
         "apply_liger_kernel_to_qwen2_5_vl",
         "apply_liger_kernel_to_qwen2_vl",
+        "apply_liger_kernel_to_qwen3",
     }
 
     if name in monkey_patch_symbols:
@@ -144,5 +146,6 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_qwen2",
             "apply_liger_kernel_to_qwen2_5_vl",
             "apply_liger_kernel_to_qwen2_vl",
+            "apply_liger_kernel_to_qwen3",
         ]
     )

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/gemma.py

@@ -200,6 +200,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -207,7 +210,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -215,8 +218,7 @@ def lce_forward(
             **loss_kwargs,
         )
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/gemma2.py

@@ -212,6 +212,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -219,7 +222,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -229,8 +232,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        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)

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/glm4.py

@@ -88,6 +88,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -95,7 +98,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -104,8 +107,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/llama.py

@@ -209,6 +209,9 @@ def lce_forward(
     )
 
     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, :]
 
     if self.config.pretraining_tp > 1:
         raise Exception("Liger Kernel does not support pretraining_tp!!")
@@ -219,7 +222,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -228,8 +231,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/mistral.py

@@ -91,6 +91,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     loss = None
@@ -98,7 +101,7 @@ def lce_forward(
 
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -107,8 +110,7 @@ def lce_forward(
         )
 
     else:
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
 
         loss = None
         if labels is not None:

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/mixtral.py

@@ -225,6 +225,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -232,7 +235,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -241,8 +244,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
 
         loss = None
         if labels is not None:

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/mllama.py

@@ -215,6 +215,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -222,7 +225,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -231,8 +234,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/olmo2.py

@@ -88,6 +88,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -95,7 +98,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -104,8 +107,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/phi3.py

@@ -213,6 +213,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -220,7 +223,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -229,8 +232,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/model/qwen2.py

@@ -199,6 +199,9 @@ def lce_forward(
     )
 
     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 = loss_kwargs.pop("shift_labels", None)
     logits = None
@@ -206,7 +209,7 @@ def lce_forward(
     # if in training mode, don't materialize logits
     if self.training and (labels is not None or shift_labels is not None):
         loss = LigerForCausalLMLoss(
-            hidden_states=hidden_states,
+            hidden_states=kept_hidden_states,
             lm_head_weight=self.lm_head.weight,
             labels=labels,
             shift_labels=shift_labels,
@@ -215,8 +218,7 @@ def lce_forward(
         )
 
     else:  # if in inference mode materialize logits
-        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
-        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        logits = self.lm_head(kept_hidden_states)
         if labels is not None:
             loss = self.loss_function(
                 logits=logits,

liger_kernel_nightly-0.5.8.dev20250503025537/src/liger_kernel/transformers/model/qwen3.py

@@ -0,0 +1,118 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.models.qwen3.modeling_qwen3 import _CONFIG_FOR_DOC
+from transformers.models.qwen3.modeling_qwen3 import QWEN3_INPUTS_DOCSTRING
+from transformers.utils import add_start_docstrings_to_model_forward
+from transformers.utils import replace_return_docstrings
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@add_start_docstrings_to_model_forward(QWEN3_INPUTS_DOCSTRING)
+@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+def lce_forward(
+    self,
+    input_ids: Optional[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,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs,
+) -> 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, Qwen3ForCausalLM
+
+    >>> model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-8B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
+
+    >>> 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
+    )
+
+    # 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,
+        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 self.training and (labels is not None or shift_labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=kept_hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            shift_labels=shift_labels,
+            hidden_size=self.config.hidden_size,
+            **kwargs,
+        )
+
+    else:  # if in inference mode materialize logits
+        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,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel/transformers/monkey_patch.py

@@ -1048,6 +1048,60 @@ def apply_liger_kernel_to_qwen2(
     print("Applied Liger kernels to Qwen2")
 
 
+def apply_liger_kernel_to_qwen3(
+    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 Qwen3 models.
+    """
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.qwen3 import modeling_qwen3
+    from transformers.models.qwen3.modeling_qwen3 import Qwen3Model
+
+    from liger_kernel.transformers.model.qwen3 import lce_forward as qwen3_lce_forward
+
+    if rope:
+        modeling_qwen3.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_qwen3.Qwen3RMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_qwen3.Qwen3ForCausalLM.forward = qwen3_lce_forward
+
+    if swiglu:
+        modeling_qwen3.Qwen3MLP = 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: Qwen3Model = 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_qwen2_vl(
     rope: bool = True,
     cross_entropy: bool = False,
@@ -1400,6 +1454,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "mixtral": apply_liger_kernel_to_mixtral,
     "olmo2": apply_liger_kernel_to_olmo2,
     "qwen2": apply_liger_kernel_to_qwen2,
+    "qwen3": apply_liger_kernel_to_qwen3,
     "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-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel_nightly.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.5.8.dev20250502223102
+Version: 0.5.8.dev20250503025537
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -317,6 +317,7 @@ loss.backward()
 | 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        |
+| Qwen3   | `liger_kernel.transformers.apply_liger_kernel_to_qwen3`    |  RoPE, 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-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/src/liger_kernel_nightly.egg-info/SOURCES.txt

@@ -176,6 +176,7 @@ 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/model/qwen3.py
 src/liger_kernel/transformers/trainer/__init__.py
 src/liger_kernel/transformers/trainer/orpo_trainer.py
 src/liger_kernel/triton/__init__.py

{liger_kernel_nightly-0.5.8.dev20250502223102 → liger_kernel_nightly-0.5.8.dev20250503025537}/test/convergence/bf16/test_mini_models.py

@@ -33,6 +33,7 @@ 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 liger_kernel.transformers import apply_liger_kernel_to_qwen3
 from test.utils import DEFAULT_DATASET_PATH
 from test.utils import MiniModelConfig
 from test.utils import assert_verbose_allclose
@@ -51,6 +52,7 @@ 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 revert_liger_kernel_to_qwen3
 from test.utils import set_seed
 from test.utils import simple_collate_fn
 from test.utils import supports_bfloat16
@@ -82,6 +84,14 @@ try:
 except ImportError:
     QWEN2_5_VL_AVAILABLE = False
 
+try:
+    from transformers.models.qwen3.configuration_qwen3 import Qwen3Config
+    from transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM
+
+    QWEN3_AVAILABLE = True
+except ImportError:
+    QWEN3_AVAILABLE = False
+
 try:
     from transformers.models.granite import GraniteConfig
     from transformers.models.granite import GraniteForCausalLM
@@ -358,6 +368,33 @@ MINI_MODEL_SETUPS = {
     ),
 }
 
+if QWEN3_AVAILABLE:
+    MINI_MODEL_SETUPS["mini_qwen3"] = MiniModelConfig(
+        liger_kernel_patch_func=apply_liger_kernel_to_qwen3,
+        liger_kernel_patch_revert_func=revert_liger_kernel_to_qwen3,
+        model_class=Qwen3ForCausalLM,
+        mini_model_config=Qwen3Config(
+            attention_dropout=0.0,
+            bos_token_id=1,
+            eos_token_id=2,
+            hidden_act="silu",
+            hidden_size=896,
+            initializer_range=0.02,
+            intermediate_size=4864,
+            max_position_embeddings=32768,
+            num_attention_heads=8,
+            num_hidden_layers=4,
+            num_key_value_heads=2,
+            rms_norm_eps=1e-6,
+            rope_theta=1000000.0,
+            sliding_window=131072,
+            tie_word_embeddings=True,
+            use_cache=True,
+            vocab_size=32000,
+            attn_implementation="sdpa",
+        ),
+    )
+
 if GEMMA3_AVAILABLE:
     MINI_MODEL_SETUPS["mini_gemma3_text"] = MiniModelConfig(
         liger_kernel_patch_func=apply_liger_kernel_to_gemma3_text,
@@ -851,6 +888,25 @@ def run_mini_model(
             1e-2,
             marks=pytest.mark.skipif(not supports_bfloat16(), reason="bfloat16 not supported on this GPU"),
         ),
+        pytest.param(
+            "mini_qwen3",
+            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 QWEN3_AVAILABLE,
+                    reason="Qwen3 not available in this version of transformers",
+                ),
+            ],
+        ),
         pytest.param(
             "mini_qwen2_vl",
             32,