Merge pull request huggingface#36 from huggingface/sparse-llama4-moe

Add support for sparse `Llama4TextMoe` layer from the kernel hub

Author: ArthurZucker

src/transformers/integrations/hub_kernels.py

@@ -26,6 +26,13 @@
     _hub_kernels_available = True
 
     _KERNEL_MAPPING: Dict[str, Dict[Union[Device, str], LayerRepository]] = {
+        "Llama4TextMoe": {
+            "cuda": LayerRepository(
+                # Move to kernels-community/moe once we release.
+                repo_id="kernels-community/moe-new-models",
+                layer_name="Llama4TextMoe",
+            )
+        },
         "MultiScaleDeformableAttention": {
             "cuda": LayerRepository(
                 repo_id="kernels-community/deformable-detr",

src/transformers/models/llama4/modeling_llama4.py

@@ -33,6 +33,7 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
+from ...integrations.hub_kernels import use_kernel_forward_from_hub
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_outputs import (
@@ -150,6 +151,7 @@ def extra_repr(self):
         return f"{tuple(self.weight.shape)}, eps={self.eps}"
 
 
+@use_kernel_forward_from_hub("Llama4TextMoe")
 class Llama4TextMoe(nn.Module):
     def __init__(self, config):
         super().__init__()