liger-kernel-nightly 0.6.4.dev20251202094519__py3-none-any.whl → 0.6.4.dev20251206103502__py3-none-any.whl

liger_kernel/ops/rms_norm.py

@@ -351,7 +351,7 @@ def _block_rms_norm_backward_kernel(
 
         # calculate the gradient of W
         if casting_mode == _CASTING_MODE_LLAMA:
-			# TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
+            # TODO(tcc): use tl.sum(..., dtype=tl.float32) once we upgrade to triton>=3.3.0
             dW_row += tl.sum((dY_row * (X_row * rstd_row[:, None]).to(X_dtype)).to(tl.float32), 0)
         else:
             # here X_row is already in fp32 (see previous if block)

liger_kernel/transformers/__init__.py

@@ -41,6 +41,7 @@ if TYPE_CHECKING:
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_glm4v_moe  # noqa: F401
+    from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_gpt_oss  # 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_hunyuan_v1_dense  # noqa: F401
     from liger_kernel.transformers.monkey_patch import apply_liger_kernel_to_hunyuan_v1_moe  # noqa: F401
@@ -110,6 +111,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_glm4",
         "apply_liger_kernel_to_glm4v",
         "apply_liger_kernel_to_glm4v_moe",
+        "apply_liger_kernel_to_gpt_oss",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_internvl",
         "apply_liger_kernel_to_llama",
@@ -187,6 +189,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_glm4",
             "apply_liger_kernel_to_glm4v",
             "apply_liger_kernel_to_glm4v_moe",
+            "apply_liger_kernel_to_gpt_oss",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_internvl",
             "apply_liger_kernel_to_llama",

liger_kernel/transformers/model/gpt_oss.py

@@ -0,0 +1,211 @@
+from typing import List
+from typing import Optional
+from typing import Union
+
+import torch
+
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import load_balancing_loss_func
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from liger_kernel.transformers.model.loss_utils import unpack_cross_entropy_result
+from liger_kernel.transformers.model.output_classes import LigerMoeCausalLMOutputWithPast
+
+
+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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> LigerMoeCausalLMOutputWithPast:
+    r"""
+        Forward pass for causal language modeling with Mixture of Experts (MoE) architecture using Liger Kernel optimizations.
+
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary. Indices can be obtained using tokenizers.
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings.
+        past_key_values (`List[torch.FloatTensor]` or `Cache`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks) that can be used to speed up
+            sequential decoding. See `past_key_values` input for more details.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+        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]`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the router logits of all MoE layers. See `router_logits` under returned tensors
+            for more detail.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence.
+        logits_to_keep (`int` or `torch.Tensor`, *optional*, defaults to 0):
+            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).
+        skip_logits (`bool`, *optional*):
+            Whether to skip logit computation and directly compute loss. If `None`, defaults to `True` during training
+            when labels are provided (to save memory), and `False` during inference.
+
+    Returns:
+        `LigerMoeCausalLMOutputWithPast`: An output object containing:
+            - loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+                Language modeling loss (for next-token prediction), including the auxiliary load balancing loss.
+            - aux_loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Auxiliary load balancing loss for the sparse MoE modules.
+            - logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*):
+                Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+                Note: logits are `None` during training when `skip_logits=True` to save memory.
+            - past_key_values (`Cache`, *optional*, returned when `use_cache=True` is passed):
+                Cached key and value projection states for faster sequential decoding.
+            - hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`):
+                Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for each layer) of shape
+                `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer.
+            - attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+                sequence_length)`. Attentions weights after the attention softmax.
+            - router_logits (`tuple(torch.FloatTensor)`, *optional*, returned when `output_router_logits=True`):
+                Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, sequence_length, num_experts)`.
+                Router logits of the MoE layers, useful to compute the auxiliary loss and z_loss.
+            - token_accuracy (`torch.FloatTensor`, *optional*, returned when `labels` is provided):
+                Token-level prediction accuracy.
+
+    Example:
+
+    ```python
+        >>> from transformers import AutoTokenizer, GptOssForCausalLM
+        >>> from liger_kernel.transformers import apply_liger_kernel_to_gpt_oss
+
+        >>> # Apply Liger Kernel patches for optimized performance
+        >>> apply_liger_kernel_to_gpt_oss()
+
+        >>> model = GptOssForCausalLM.from_pretrained("openai/gpt-oss-20b")
+        >>> tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Inference: Forward pass returns logits
+        >>> outputs = model(**inputs)
+        >>> outputs.logits.shape
+        torch.Size([1, 12, 201088])
+
+        >>> # Get next token prediction
+        >>> next_token_logits = outputs.logits[:, -1, :]
+        >>> predicted_token_id = next_token_logits.argmax(dim=-1)
+
+        >>> # Training: Forward pass with labels returns loss
+        >>> labels = inputs.input_ids.clone()
+        >>> outputs = model(**inputs, labels=labels)
+        >>> outputs.loss
+        tensor(2.6454)
+    ```"""
+
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_router_logits = (
+        output_router_logits if output_router_logits is not None else self.config.output_router_logits
+    )
+
+    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: MoeModelOutputWithPast = 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,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    # 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
+    token_accuracy = None
+
+    if skip_logits is None:
+        skip_logits = self.training and (labels is not None or shift_labels is not None)
+
+    if skip_logits:
+        result = 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,
+        )
+        loss, _, token_accuracy = unpack_cross_entropy_result(result)
+    else:  # if in inference model materialize logits
+        logits = self.lm_head(kept_hidden_states)
+        if labels is not None or shift_labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                shift_labels=shift_labels,
+                vocab_size=self.vocab_size,
+                **kwargs,
+            )
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(loss.device)  # make sure to reside in the same device
+
+    return LigerMoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+        token_accuracy=token_accuracy,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -20,6 +20,7 @@ from liger_kernel.transformers.model.gemma import lce_forward as gemma_lce_forwa
 from liger_kernel.transformers.model.gemma import lce_forward_deprecated as gemma_lce_forward_deprecated
 from liger_kernel.transformers.model.gemma2 import lce_forward as gemma2_lce_forward
 from liger_kernel.transformers.model.gemma2 import lce_forward_deprecated as gemma2_lce_forward_deprected
+from liger_kernel.transformers.model.gpt_oss import lce_forward as gpt_oss_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward as llama_lce_forward
 from liger_kernel.transformers.model.llama import lce_forward_deprecated as llama_lce_forward_deprecated
 from liger_kernel.transformers.model.llava import lce_forward as llava_lce_forward
@@ -1459,6 +1460,79 @@ def apply_liger_kernel_to_qwen3_moe(
                 _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
 
 
+def apply_liger_kernel_to_gpt_oss(
+    rope: bool = True,
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = True,
+    rms_norm: bool = True,
+    swiglu: bool = False,  # Set to False by default since GPT-OSS has custom expert implementation
+    model: PreTrainedModel = None,
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace GPT-OSS models.
+    NOTE: GPT-OSS is supported in transformers >= 4.55.0
+    NOTE: SwiGLU patching is disabled by default for GPT-OSS as it uses a custom expert
+          implementation with clamping and MXFP4 quantization.
+
+    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 False.
+            Note: GPT-OSS uses a custom expert implementation, so SwiGLU patching is disabled by default.
+        model (PreTrainedModel): The model instance to apply Liger kernels to, if the model has already been
+            loaded. Default is None.
+    """
+    if version.parse(transformers.__version__) < version.parse("4.55.0"):
+        logger.warning("GPT-OSS support requires transformers >= 4.55.0")
+        return
+
+    assert not (cross_entropy and fused_linear_cross_entropy), (
+        "cross_entropy and fused_linear_cross_entropy cannot both be True."
+    )
+
+    from transformers.models.gpt_oss import modeling_gpt_oss
+    from transformers.models.gpt_oss.modeling_gpt_oss import GptOssModel
+
+    if rope:
+        modeling_gpt_oss.apply_rotary_pos_emb = liger_rotary_pos_emb
+
+    if rms_norm:
+        modeling_gpt_oss.GptOssRMSNorm = LigerRMSNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        if model is not None:
+            model.forward = MethodType(gpt_oss_lce_forward, model)
+        else:
+            modeling_gpt_oss.GptOssForCausalLM.forward = gpt_oss_lce_forward
+
+    # Note: SwiGLU patching is not implemented for GPT-OSS due to custom expert implementation
+    # with clamping (swiglu_limit=7.0) and MXFP4 quantization
+
+    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: GptOssModel = 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 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,
@@ -2752,6 +2826,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "glm4": apply_liger_kernel_to_glm4,
     "glm4v": apply_liger_kernel_to_glm4v,
     "glm4v_moe": apply_liger_kernel_to_glm4v_moe,
+    "gpt_oss": apply_liger_kernel_to_gpt_oss,
     "internvl": apply_liger_kernel_to_internvl,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,

{liger_kernel_nightly-0.6.4.dev20251202094519.dist-info → liger_kernel_nightly-0.6.4.dev20251206103502.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.4.dev20251202094519
+Version: 0.6.4.dev20251206103502
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 
@@ -312,6 +312,7 @@ loss.backward()
 | OLMo2   | `liger_kernel.transformers.apply_liger_kernel_to_olmo2`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | Olmo3   | `liger_kernel.transformers.apply_liger_kernel_to_olmo3`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | GLM-4   | `liger_kernel.transformers.apply_liger_kernel_to_glm4`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
+| GPT-OSS   | `liger_kernel.transformers.apply_liger_kernel_to_gpt_oss`     | RoPE, RMSNorm, CrossEntropyLoss, FusedLinearCrossEntropy |
 | InternVL3   | `liger_kernel.transformers.apply_liger_kernel_to_internvl`     | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
 | HunyuanV1   | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_dense`    |  RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |
 | HunyuanV1 MoE | `liger_kernel.transformers.apply_liger_kernel_to_hunyuan_v1_moe` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy       |

{liger_kernel_nightly-0.6.4.dev20251202094519.dist-info → liger_kernel_nightly-0.6.4.dev20251206103502.dist-info}/RECORD

@@ -33,7 +33,7 @@ liger_kernel/ops/llama4_rope.py,sha256=-aqdZzllklTN8b9--e-TsWY_ntGCN8-tyseT4x0bd
 liger_kernel/ops/multi_token_attention.py,sha256=Oz_RXDp-OSS_R_HuGmaETHdAJ7Toda_70OfE7TXMUlY,7645
 liger_kernel/ops/poly_norm.py,sha256=5IdJEZnbbhblkL_X8UhSD4A2CooQbOAZJw8nAekWNs4,11372
 liger_kernel/ops/qwen2vl_mrope.py,sha256=3GExhYpLgB4VUtyZyjRk8XjEur3W4EWF6HQ67ML5vBU,8481
-liger_kernel/ops/rms_norm.py,sha256=2V8qheEvidBm0VxwfOoAnV837F6blmtTgP91VYdGs7c,19211
+liger_kernel/ops/rms_norm.py,sha256=owWgM1jE5aP4clshCNWiulnemHPzR72D9QN2kc3eoe0,19220
 liger_kernel/ops/rope.py,sha256=v-7JHRrv-5ImoROkpKfl30WwWI4qTa2tAl7zQeB4ml4,8956
 liger_kernel/ops/softmax.py,sha256=tgORx6MK1IDDtZKqGarj0IPIVjqAIEUXXYPiinhRdtI,5864
 liger_kernel/ops/sparsemax.py,sha256=AeWe1xgkHJFEKWTj2vu_0hj7LztGvjqXAps-QTpCY0U,5087
@@ -43,7 +43,7 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=kYp84AOA7D9PYrvBUSrNsfQIt8elr_uA9OxCkbfiUFA,3980
 liger_kernel/ops/experimental/embedding.py,sha256=tolj3tItkzpSb30zWqDN2_yX4ectflaQ8HMyKyFIQc8,4172
 liger_kernel/ops/experimental/mm_int8int2.py,sha256=TrS9lpwekrik_w5qE7AhMJD1bcq-OidjtbsW80oZ6IM,13314
-liger_kernel/transformers/__init__.py,sha256=CgwhrY5cdx6OcRgR2ZZJbOIkLswQWPTr-BAaoxDNNOY,10687
+liger_kernel/transformers/__init__.py,sha256=4sqcDbOZ_JtS9Ag-7oyuhq5jN298GyzjJFu9J-DyyZQ,10872
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=DMtHkKrVJDSsels7KgGQJqrXkEAd6Zopcdr-5oRmQgE,2010
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
@@ -60,7 +60,7 @@ liger_kernel/transformers/jsd.py,sha256=DGqRnxIZxsvxo0_tbbxX3b-sDbDjC_yKufyRIHCc
 liger_kernel/transformers/kl_div.py,sha256=WLffFbh1EExD2Eb1F7lN11fo9JJC-0751WJjZAF1Fj8,409
 liger_kernel/transformers/layer_norm.py,sha256=c9pk3PEasOKYR0rhe5e5nNrnYKVCEW4VC8S6LpCq9EQ,906
 liger_kernel/transformers/llama4_rope.py,sha256=kS6PSHEwf3dS7hD7C7p8S0geugx2EMCiP0h0F7LsUoY,3639
-liger_kernel/transformers/monkey_patch.py,sha256=4LV6LSz_AAop6HWk1spZm1QigPN9nUDPJu9tK21-jIo,132446
+liger_kernel/transformers/monkey_patch.py,sha256=0ER5BjQXcIKwgL2e7ji3_DIm1DaJzevzo53aXSe2YJU,135862
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/poly_norm.py,sha256=g5tC75i3qy1_N26ZUP-jfpct7ivQAEdJfIfx8IXzeyE,1377
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
@@ -82,6 +82,7 @@ liger_kernel/transformers/model/gemma3.py,sha256=ZUrFCc-pfF8jYHV0HsptBr98hx6p2q9
 liger_kernel/transformers/model/glm4.py,sha256=bSp22iPIjsli4-c_usUOsyh1Bs2gIK8X6ynS0azseUs,5900
 liger_kernel/transformers/model/glm4v.py,sha256=dd-BQpccDCp1SbIxcJ5rG8xcwYQK3KOv1Tgm9TGnZc4,6594
 liger_kernel/transformers/model/glm4v_moe.py,sha256=zKhMdOOrRhlrvCSFaeVYfddL1ubpY8edEO91TN81n98,7135
+liger_kernel/transformers/model/gpt_oss.py,sha256=8jEAQQNEXgVA-yuvEjKkBQvCvZy0E9ns-O9BPlajXXU,11197
 liger_kernel/transformers/model/hunyuan_v1.py,sha256=MJvP9xkUFePIV0HLETJM4YPbVCEPkAE1ZI5Jxyiebh0,5731
 liger_kernel/transformers/model/internvl.py,sha256=OOutracs9qrPHSU7FVYar08yinvGrHQVPvo39JEws6w,6473
 liger_kernel/transformers/model/llama.py,sha256=kqZeONzwTBzudoChlKMzq1w23BtYGbxWZC1l1V__JTw,13410
@@ -110,9 +111,9 @@ liger_kernel/transformers/trainer/__init__.py,sha256=p7yQfklV8-467qSz_ZMimkbDF7H
 liger_kernel/transformers/trainer/orpo_trainer.py,sha256=tX0h63aOFe3rNqTmk6JpMf75UPo981yzEa6TghnjS0Q,5370
 liger_kernel/triton/__init__.py,sha256=qCiCamzCRv6lpV8IqpAc9YMdNKC7GKurClWceQPnlis,92
 liger_kernel/triton/monkey_patch.py,sha256=Rd0hUHAzDkFfHvnX7-PBaNK5EKnZhtfM_h-fgQH9HPY,1568
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/METADATA,sha256=ZgwmbyNXCX9-W_MkkjoTyoMWce_cteqdiqlMCeQ_Tt8,25238
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.4.dev20251202094519.dist-info/RECORD,,
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/METADATA,sha256=9fKlKAtH1HWCDW5wysDkZA-z5cqMjTz3YDhacJ_dhn8,25375
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.4.dev20251206103502.dist-info/RECORD,,