marlin: support tp>1 when group_size==-1

server/text_generation_server/utils/weights.py

@@ -513,6 +513,12 @@ class Weights:
                     "Cannot load `marlin` weight, make sure the model is already quantized, or quantize it with `text-generation-server quantize ORIGINAL_MODEL_ID NEW_MODEL_ID`"
                 )
 
+            num_groups = self._get_slice(f"{prefix}.s").get_shape()[0]
+            if num_groups == 1:
+                # The number of groups is 1 when group_size == -1. share
+                # scales between all shards in this case.
+                s = self.get_tensor(f"{prefix}.s")
+            else:
                 s = self.get_sharded(f"{prefix}.s", dim=0)
             weight = MarlinWeight(B=B, s=s)