add support_offload_to_disk check

gptqmodel/models/base.py

@@ -205,6 +205,8 @@ class BaseQModel(nn.Module):
 
     support_batch_quantize = True
 
+    support_offload_to_disk = True
+
     ATTENTION_MASKS_DTYPE = torch.bool # default to bool
 
     ATTENTION_MASKS_REQUIRED_FOR_INPUT: bool = False

gptqmodel/models/definitions/gpt_oss.py

@@ -126,6 +126,8 @@ def forward(self, hidden_states):
         return router_scores, router_indices
 
 class GPTOSSGPTQ(BaseQModel):
+    support_offload_to_disk = False
+
     dynamic_expert_index = "num_local_experts"
 
     pre_lm_head_norm_module = "model.norm"

gptqmodel/models/definitions/llama4.py

@@ -12,6 +12,7 @@ class Llama4QModel(BaseQModel):
     # some bug in the attention_mask of transformers.modeling_llama4,
     # so batch quantization for Llama4 is temporarily not supported.
     support_batch_quantize = False
+    support_offload_to_disk = False
     loader = AutoModelForImageTextToText
 
     pre_lm_head_norm_module = "language_model.model.norm"

gptqmodel/models/loader.py

@@ -205,6 +205,12 @@ def skip(*args, **kwargs):
         if hasattr(config, "hidden_act") and config.hidden_act == "xielu":
             quantize_config.offload_to_disk = False
 
+        # some models need convert moe-experts after model loaded, like GPTOSS and Llama4
+        # so offload_to_disk is not supported for them.
+        if not cls.support_offload_to_disk:
+            quantize_config.offload_to_disk = False
+            log.warn(f"{cls} doesn't support offload_to_disk, set quantize_config.offload_to_disk to False.")
+
         if quantize_config.offload_to_disk:
             model = build_shell_model(cls.loader, config=config, **model_init_kwargs)
             model._model_init_kwargs = model_init_kwargs