liger-kernel-nightly 0.6.2.dev20250903164435__py3-none-any.whl → 0.6.2.dev20250911161200__py3-none-any.whl

liger_kernel/transformers/__init__.py

@@ -36,6 +36,7 @@ if TYPE_CHECKING:
     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_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_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_llama4  # noqa: F401
@@ -95,6 +96,7 @@ def __getattr__(name: str):
         "apply_liger_kernel_to_gemma3_text",
         "apply_liger_kernel_to_glm4",
         "apply_liger_kernel_to_glm4v",
+        "apply_liger_kernel_to_glm4v_moe",
         "apply_liger_kernel_to_granite",
         "apply_liger_kernel_to_llama",
         "apply_liger_kernel_to_llava",
@@ -159,6 +161,7 @@ if _TRANSFORMERS_AVAILABLE:
             "apply_liger_kernel_to_gemma3_text",
             "apply_liger_kernel_to_glm4",
             "apply_liger_kernel_to_glm4v",
+            "apply_liger_kernel_to_glm4v_moe",
             "apply_liger_kernel_to_granite",
             "apply_liger_kernel_to_llama",
             "apply_liger_kernel_to_llava",

liger_kernel/transformers/fused_linear_cross_entropy.py

@@ -25,7 +25,8 @@ class LigerFusedLinearCrossEntropyLoss(torch.nn.Module):
         assert reduction in {
             "mean",
             "sum",
-        }, f"reduction must be 'mean' or 'sum'. Got: {reduction}"
+            "none", 
+        }, f"reduction must be 'mean' or 'sum' or 'none'. Got: {reduction}"
         assert softcap is None or softcap > 0, f"softcap must greater than 0.0 or None. Got: {softcap}"
         self.ce_weight = ce_weight
         self.ignore_index = ignore_index

liger_kernel/transformers/model/glm4v_moe.py

@@ -0,0 +1,152 @@
+from typing import Optional
+from typing import Tuple
+from typing import Union
+
+import torch
+
+from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeCausalLMOutputWithPast
+from transformers.utils.deprecation import deprecate_kwarg
+
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+
+
+@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[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    skip_logits: Optional[bool] = None,
+    **kwargs,
+) -> Union[Tuple, Glm4vMoeCausalLMOutputWithPast]:
+    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]`.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+            The temporal, height and width of feature shape of each video in LLM.
+        rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+            The rope index difference between sequence length and multimodal rope.
+
+
+        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).
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
+    >>> import torch
+
+    >>> MODEL_PATH = "zai-org/GLM-4.5V"
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image",
+                    "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+                },
+                {
+                    "type": "text",
+                    "text": "describe this image"
+                }
+            ],
+        }
+    ]
+    >>> processor = AutoProcessor.from_pretrained(MODEL_PATH)
+    >>> model = Glm4vMoeForConditionalGeneration.from_pretrained(
+        pretrained_model_name_or_path=MODEL_PATH,
+        torch_dtype="auto",
+        device_map="auto",
+    )
+    >>> inputs = processor.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_dict=True,
+        return_tensors="pt"
+    ).to(model.device)
+    >>> inputs.pop("token_type_ids", None)
+    >>> generated_ids = model.generate(**inputs, max_new_tokens=8192)
+    >>> output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
+    ```
+    """
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        pixel_values=pixel_values,
+        pixel_values_videos=pixel_values_videos,
+        image_grid_thw=image_grid_thw,
+        video_grid_thw=video_grid_thw,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        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 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 = 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:
+        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 Glm4vMoeCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=outputs.rope_deltas,
+    )

liger_kernel/transformers/monkey_patch.py

@@ -1928,6 +1928,107 @@ def apply_liger_kernel_to_glm4v(
                     _patch_rms_norm_module(decoder_layer.post_mlp_layernorm)
 
 
+def apply_liger_kernel_to_glm4v_moe(
+    rope: bool = False,
+    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 GLM4v_moe models.
+
+    Args:
+        rope (bool): Whether to apply Liger's rotary position embedding. Default is False.
+        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 SwiGLUMLP. 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.glm4v_moe import modeling_glm4v_moe
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeForConditionalGeneration
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeModel
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextModel
+    from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeVisionModel
+
+    from liger_kernel.transformers.model.glm4v_moe import lce_forward as glm4v_moe_lce_forward
+    from liger_kernel.transformers.rms_norm import LigerRMSNormForGlm4
+
+    if rope:
+        raise NotImplementedError("liger_rotary_pos_emb is not available for Glm4 models.")
+    if rms_norm:
+        modeling_glm4v_moe.Glm4vRMSNorm = LigerRMSNormForGlm4
+    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(glm4v_moe_lce_forward, model)
+        else:
+            modeling_glm4v_moe.Glm4vMoeForConditionalGeneration.forward = glm4v_moe_lce_forward
+
+    if model is not None:
+        # The model instance already exists, so we need to additionally patch the
+        # instance variables that reference already-instantiated modules
+        if isinstance(model, (Glm4vMoeForConditionalGeneration, Glm4vMoeModel)):
+            # Note: language_model and visual properties can be accessed throught conditional class for BC.
+            # Not sure if it is subject to changes in the future.
+            # Reference: https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm4v_moe/modeling_glm4v_moe.py#L337
+            text_model: Glm4vMoeTextModel = model.language_model
+            vision_model: Glm4vMoeVisionModel = model.visual
+            Glm4vMoeTextMoE = modeling_glm4v_moe.Glm4vMoeTextMoE
+        elif isinstance(model, Glm4vMoeTextModel):
+            text_model: Glm4vMoeTextModel = model
+            vision_model = None
+        else:
+            # Note: Currently there's no support for patching vision model only. Feel free to raise an issue if needed.
+            raise TypeError(
+                f"Unsupported glm4v_moe model type. `model` must be `Glm4vMoeForConditionalGeneration`, `Glm4vMoeVisionModel` or `Glm4vMoeTextModel`. Got: {type(model)}"
+            )
+
+        if vision_model is not None:
+            _patch_rms_norm_module(vision_model.post_conv_layernorm)
+            _patch_rms_norm_module(vision_model.post_layernorm)
+            for vision_block in vision_model.blocks:
+                if rms_norm:
+                    _patch_rms_norm_module(vision_block.norm1)
+                    _patch_rms_norm_module(vision_block.norm2)
+                if swiglu:
+                    _patch_swiglu_module(vision_block.mlp, LigerSwiGLUMLP)
+
+        if text_model is not None:
+            if rms_norm:
+                _patch_rms_norm_module(text_model.norm)
+            for decoder_layer in text_model.layers:
+                if swiglu:
+                    decoder_layer.mlp = _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)
+        if isinstance(Glm4vMoeTextMoE, type) and isinstance(decoder_layer.mlp, Glm4vMoeTextMoE):
+            experts = getattr(decoder_layer.mlp, "experts", None)
+            if experts is not None:
+                for expert in experts:
+                    _patch_swiglu_module(expert, LigerSwiGLUMLP)
+            if decoder_layer.mlp.shared_experts is not None:
+                _patch_swiglu_module(decoder_layer.mlp.shared_experts, LigerSwiGLUMLP)
+            for decoder_layer in text_model.layers:
+                if rms_norm:
+                    _patch_rms_norm_module(decoder_layer.input_layernorm)
+                    _patch_rms_norm_module(decoder_layer.post_attention_layernorm)
+
+
 # Model type corresponds to the keys defined in transformers/models/auto/modeling_auto.py
 MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma": apply_liger_kernel_to_gemma,
@@ -1936,6 +2037,7 @@ MODEL_TYPE_TO_APPLY_LIGER_FN = {
     "gemma3": apply_liger_kernel_to_gemma3,
     "glm4": apply_liger_kernel_to_glm4,
     "glm4v": apply_liger_kernel_to_glm4v,
+    "glm4v_moe": apply_liger_kernel_to_glm4v_moe,
     "llama": apply_liger_kernel_to_llama,
     "llama4_text": apply_liger_kernel_to_llama4,
     "llama4": apply_liger_kernel_to_llama4,

{liger_kernel_nightly-0.6.2.dev20250903164435.dist-info → liger_kernel_nightly-0.6.2.dev20250911161200.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: liger_kernel_nightly
-Version: 0.6.2.dev20250903164435
+Version: 0.6.2.dev20250911161200
 Summary: Efficient Triton kernels for LLM Training
 License: BSD 2-CLAUSE LICENSE
         Copyright 2024 LinkedIn Corporation 

{liger_kernel_nightly-0.6.2.dev20250903164435.dist-info → liger_kernel_nightly-0.6.2.dev20250911161200.dist-info}/RECORD

@@ -41,14 +41,14 @@ liger_kernel/ops/tvd.py,sha256=FHJtLQI95ijqgg9UtaHpMAjSCiPxB6CduPwPMcGxelc,6405
 liger_kernel/ops/utils.py,sha256=uoFKQqo-34N2TWQNvXMFywqGiOMMXNEVBxVojzlUAa0,3836
 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=jkokP69dbCzUDTz-H6QowB5xNEflmgQ7Zv-_4MVuxpY,8440
+liger_kernel/transformers/__init__.py,sha256=GA4GdoP0VKKGzXoLN51DacuTtNuwciS96ZAblkDYl3w,8631
 liger_kernel/transformers/auto_model.py,sha256=0qCTRZt280Bj_LcFdzo9hlaR-BWNazawXOGgoCZjgEg,1545
 liger_kernel/transformers/cross_entropy.py,sha256=z3KTWQnFxr_IZaVjtYt0ZNEWQdDdYThN35xWkHlDGH0,1683
 liger_kernel/transformers/dyt.py,sha256=i-4GPaMrl-jab9TVI5qN0-H9qycn_mCbV82ozU4nbmU,723
 liger_kernel/transformers/fsdp.py,sha256=CUiyjTmjkjY7pLXQv8ly9rnzgXw6529csd9pvtJNMYc,3096
 liger_kernel/transformers/functional.py,sha256=-vpz95wbv5wLpInjSG06KNHETsEgKnRIiV-lMYHVs68,7841
 liger_kernel/transformers/fused_add_rms_norm.py,sha256=7_Bzg-x6lLe6W1qG2DtjDALhEpNZlC6N5GppEs9cTYY,1199
-liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=ZMxkiJzGz1KtqgAdsqPODq3bugHBx_80kPYcd5z-xmM,1990
+liger_kernel/transformers/fused_linear_cross_entropy.py,sha256=3O_e3AcFmEftT4GawvgBStD9DJGcL8FqIqXGJQSxfQ0,2021
 liger_kernel/transformers/fused_linear_jsd.py,sha256=bZ4otCvWBuOnA5XdQL-FzZVItJlDt-ht9e_pG7PG93E,3999
 liger_kernel/transformers/fused_neighborhood_attention.py,sha256=TxYDUAt9B6WSP14aJP66C_2Mbds2sSIPGnamhUSTrC8,7957
 liger_kernel/transformers/geglu.py,sha256=mrgqzIUVd6lN7fkDKLkw5YaESDxDtFgbot430WwPVOQ,1107
@@ -58,7 +58,7 @@ 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=pG3Yf0fMg4_0pAncc2wLtpdfXvmC5CROpNJ43-MmElM,93075
+liger_kernel/transformers/monkey_patch.py,sha256=DpYF1mVjmErtkFgnxi20rnyeXr7bvVSg3BR166cMF5I,98472
 liger_kernel/transformers/multi_token_attention.py,sha256=K3NIY9_5TPgZ4_Rahn0xnkMXxD_fmlJHK4CWGYvGQp0,1752
 liger_kernel/transformers/qwen2vl_mrope.py,sha256=5EwSqrMdsL9MYspeBMXBsNJKvH0MOmRrtJXAJlnnlOI,1047
 liger_kernel/transformers/rms_norm.py,sha256=vkekcvTeWY8vL4H6hg3t0XeY0Ew_3OFMPHuzqlxPPVw,2719
@@ -76,6 +76,7 @@ liger_kernel/transformers/model/gemma2.py,sha256=R_JFPyWTk7RyA7D05ZiIaNO5pX8gWcv
 liger_kernel/transformers/model/gemma3.py,sha256=FKO4j3t4W_5uECRA1lhVnXC-It2GhirHm4tpCf9ApAc,12785
 liger_kernel/transformers/model/glm4.py,sha256=GlnEhdGJuDIqp2R9qC54biY3HwV1tWmfpJm6ijoAsrM,5257
 liger_kernel/transformers/model/glm4v.py,sha256=zbV3agptEYpGAD0eeCRwIpJAhJUviTT5xQbbLlgpVnc,5957
+liger_kernel/transformers/model/glm4v_moe.py,sha256=xTPjRqI3E_--0gbqM8MYlBYvVecGkx9iU_2AnFRhfmQ,6172
 liger_kernel/transformers/model/llama.py,sha256=i8jJgyZsMKWQ-zKloETLugtwFpUOdaWxLDceciFXKd4,12832
 liger_kernel/transformers/model/llama4.py,sha256=IgbB8sTh3dlETQnaNNy1bZLuXy-Nt7qmeAjF27ydGpg,4210
 liger_kernel/transformers/model/llava.py,sha256=bLCioday_SOm69ogMDBhy_4UsVkH2-BSl93-EXY6-7I,15076
@@ -96,9 +97,9 @@ 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-0.6.2.dev20250903164435.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
-liger_kernel_nightly-0.6.2.dev20250903164435.dist-info/METADATA,sha256=BgiSTSMznb0cvZyFqU68T0sEIAOBcf9hvuO6jIPCcC8,24504
-liger_kernel_nightly-0.6.2.dev20250903164435.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
-liger_kernel_nightly-0.6.2.dev20250903164435.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
-liger_kernel_nightly-0.6.2.dev20250903164435.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
-liger_kernel_nightly-0.6.2.dev20250903164435.dist-info/RECORD,,
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/LICENSE,sha256=OhzLDHJ0to4a8sodVLELZiCFylZ1NAAYLs-HrjPy0ag,1312
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/METADATA,sha256=K99hPx0Vx-l-9kUhuhAfUuqgyqiA0l5SS08o5Z1PxzU,24504
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/NOTICE,sha256=njwnoPZLh9AN8SJQzxvCGLHi-8X__AvWRze6joNXIY8,2066
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/WHEEL,sha256=iAkIy5fosb7FzIOwONchHf19Qu7_1wCWyFNR5gu9nU0,91
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/top_level.txt,sha256=2eghu4hA3LnkM7ElW92tQ8zegWKgSbeo-k-aGe1YnvY,13
+liger_kernel_nightly-0.6.2.dev20250911161200.dist-info/RECORD,,