feat(convert_hf_to_gguf): Add GraniteMoeModel

GraniteMoe has the same configuration deltas as Granite

Branch: GraniteMoE

Signed-off-by: Gabe Goodhart <[email protected]>

Author: gabe-l-hart

convert_hf_to_gguf.py

@@ -4102,14 +4102,23 @@ def set_gguf_parameters(self):
         #   consistency
         if attention_scale := self.hparams.get("attention_multiplier"):
             self.gguf_writer.add_attention_scale(attention_scale)
+            logger.info("gguf: (granite) attention_scale = %s", attention_scale)
         if embedding_scale := self.hparams.get("embedding_multiplier"):
             self.gguf_writer.add_embedding_scale(embedding_scale)
+            logger.info("gguf: (granite) embedding_scale = %s", embedding_scale)
         if residual_scale := self.hparams.get("residual_multiplier"):
             self.gguf_writer.add_residual_scale(residual_scale)
-        if logits_scaling := self.hparams.get("logits_scaling"):
-            self.gguf_writer.add_logit_scale(logits_scaling)
+            logger.info("gguf: (granite) residual_scale = %s", residual_scale)
+        if logits_scale := self.hparams.get("logits_scaling"):
+            self.gguf_writer.add_logit_scale(logits_scale)
+            logger.info("gguf: (granite) logits_scale = %s", logits_scale)
 
 
+@Model.register("GraniteMoeForCausalLM")
+class GraniteMoeModel(GraniteModel):
+    """Conversion for IBM's GraniteMoeForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_MOE
+
 ###### CONVERSION LOGIC ######
 
 # tree of lazy tensors

gguf-py/gguf/constants.py

@@ -1253,7 +1253,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.FFN_NORM,
-        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
     ],

gguf-py/gguf/tensor_mapping.py

@@ -251,11 +251,12 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
-            "layers.{bid}.feed_forward.gate",             # mixtral
-            "model.layers.{bid}.block_sparse_moe.gate",   # mixtral
-            "model.layers.{bid}.mlp.gate",                # qwen2moe olmoe
-            "transformer.decoder_layer.{bid}.router",     # Grok
-            "transformer.blocks.{bid}.ffn.router.layer",  # dbrx
+            "layers.{bid}.feed_forward.gate",                   # mixtral
+            "model.layers.{bid}.block_sparse_moe.gate",         # mixtral
+            "model.layers.{bid}.mlp.gate",                      # qwen2moe olmoe
+            "transformer.decoder_layer.{bid}.router",           # Grok
+            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
+            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -329,7 +330,6 @@ class TensorNameMap:
             "transformer.decoder_layer.{bid}.moe.linear",       # Grok (merged)
             "transformer.blocks.{bid}.ffn.experts.mlp.w1",      # dbrx
             "model.layers.{bid}.mlp.experts.gate_proj",         # qwen2moe olmoe (merged)
-            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
         ),
 
         MODEL_TENSOR.FFN_GATE_SHEXP: (