[Bugfix] Allow grouped weight-only int8 MoE under moe_wna16 (alternative a of 2)

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe/compressed_tensors_moe_wna16.py

@@ -82,7 +82,6 @@ def __init__(
             else:
                 scale = kInt4StaticGroupScale
         elif self.num_bits == 8:
-            assert self.group_size == -1
             scale = kInt8StaticGroupScale
         else:
             raise ValueError(

vllm/model_executor/layers/quantization/moe_wna16.py

@@ -222,7 +222,6 @@ def __init__(self, quant_config: MoeWNA16Config, moe: "FusedMoEConfig") -> None:
             else:
                 scale = kInt4StaticGroupScale
         elif num_bits == 8:
-            assert group_size == -1
             quant_type = INT8_DTYPE
             scale = kInt8StaticGroupScale
         else:
@@ -254,6 +253,14 @@ def create_weights(
         group_size = self.quant_config.group_size
         group_size_div_factor = 1
 
+        # group_size == -1 means per-channel: one scale per output row, i.e. a
+        # single group spanning the whole reduction axis. The loop below cannot
+        # normalise it (x % -1 == 0 for every x, so it exits immediately and
+        # leaves a negative divisor for the scale shapes), so map it to the
+        # axis length up front.
+        if group_size == -1:
+            group_size = min(intermediate_size_per_partition, hidden_size)
+
         # make intermediate_size and hidden_size divisible by group_size
         # we reduce the group size to ensure that
         # and we would repeat the loaded_weight later