Use pypi cutlass

gptqmodel/models/loader.py

@@ -580,14 +580,14 @@ def build_layerwise_device_map(
             device_ids = list(range(num_gpus))
             device_map: Dict[str, str] = {}
             mod2name = {m: n for n, m in model.named_modules()}
-            
+
             if torch.cuda.is_available():
                 device_strs = [f"cuda:{i}" for i in range(num_gpus)]
             elif hasattr(torch, "xpu") and torch.xpu.is_available():
                 device_strs = [f"xpu:{i}" for i in range(num_gpus)]
             else:
                 device_strs = ["cpu"] * num_gpus
-            
+
             def assign(mod, device_id):
                 if mod is None:
                     return
@@ -726,7 +726,7 @@ def assign(mod, device_id):
                 )
             if not _validate_machete_device_support():
                 raise ValueError(
-                    f"Kernel: Machete kernel requires compute capability >= 9.0. Detected capability: {torch.cuda.get_device_capability()}"
+                    f"Kernel: Machete kernel currently supports Hopper GPUs (SM 90). Detected capability: {torch.cuda.get_device_capability()}."
                 )
 
         if backend in [BACKEND.MARLIN, BACKEND.MARLIN_FP16] and (