feat: Add support for dense FFN in GraniteMoeHybrid

This was already partially supported via reusing the granite ffn builder,
and there may be models that leverage this architecture going forward. The
naming is a bit odd, but in the transformers version, it reuses the same
model class and simply has zero regular experts and a single shared expert
(which is the same as a single dense FFN).

Branch: GraniteFour

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Author: gabe-l-hart

convert_hf_to_gguf.py

@@ -6538,13 +6538,25 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        has_experts = bool(self.hparams.get('num_local_experts'))
+
         if name.endswith("shared_mlp.input_linear.weight"):
             ffn_dim = self.hparams["shared_intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
             gate, up = data_torch.split(ffn_dim, dim=-2)
+            if has_experts:
+                return [
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                    (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                ]
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), up),
+            ]
+
+        if not has_experts and name.endswith("shared_mlp.output_linear.weight"):
             return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN, bid), data_torch)
             ]
 
         return super().modify_tensors(data_torch, name, bid)
@@ -6569,7 +6581,7 @@ def modify_tensors(
     ) -> Iterable[tuple[str, Tensor]]:
         if (
             name.endswith("block_sparse_moe.input_linear.weight")
-            or name.endswith("shared_mlp.input_linear.weight")
+            or "shared_mlp" in name
         ):
             return GraniteMoeModel.modify_tensors(self, data_torch, name, bid)
         return super().modify_tensors(data_torch, name, bid)

gguf-py/gguf/constants.py

@@ -2190,13 +2190,18 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.FFN_NORM,
+        # MoE
         MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_GATE_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.FFN_GATE_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_DOWN_SHEXP,
+        # Dense
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
     ],
     MODEL_ARCH.CHAMELEON: [
         MODEL_TENSOR.TOKEN_EMBD,