[Misc] Add assertion and helpful message for marlin24 compressed mode…

…ls (vllm-project#11388)
ayylemao · Dec 24, 2024 · d4eacfb · d4eacfb
1 parent b5e10a9
commit d4eacfb
Showing 1 changed file with 4 additions and 0 deletions.
diff --git a/...el_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py b/...el_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py
@@ -61,6 +61,10 @@ def create_weights(self, layer: torch.nn.Module, input_size: int,
                        params_dtype: torch.dtype, weight_loader: Callable,
                        **kwargs):
 
+        assert params_dtype == torch.float16, (
+            "float16 is required for marlin24 compressd models. Set dtype=torch.float16"  # noqa: E501
+        )
+
         pack_factor = 32 // self.quant_type.size_bits
         output_size_per_partition = sum(output_partition_sizes)