[NPU] Adapt MiMo-V2.5-W8A8 (#29131)

python/sglang/srt/layers/quantization/modelslim/modelslim.py

@@ -175,6 +175,10 @@ class ModelSlimConfig(QuantizationConfig):
                 prefix_in_quant_config = prefix.replace(
                     proj_name, packed_modules_mapping_subset[proj_name][0]
                 )
+                # Verify the remapped prefix exists in quant_description.
+                # If not (e.g. json uses fused name as-is), fall back to original.
+                if prefix_in_quant_config + ".weight" not in self.quant_description:
+                    prefix_in_quant_config = prefix
             if self.is_layer_skipped(
                 prefix, packed_modules_mapping_subset
             ) or self.is_layer_skipped(prefix, self.packed_modules_mapping):

python/sglang/test/ascend/test_ascend_utils.py

@@ -149,6 +149,7 @@ META_LLAMA_3_1_8B_INSTRUCT = os.path.join(
 )
 MIMO_7B_RL_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "XiaomiMiMo/MiMo-7B-RL")
 MIMO_V2_FLASH_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "XiaomiMiMo/MiMo-V2-Flash")
+MIMO_V2_5_W8A8_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "solinliu/MiMo-V2.5-W8A8")
 MINICPM3_4B_WEIGHTS_PATH = os.path.join(MODEL_WEIGHTS_DIR, "OpenBMB/MiniCPM3-4B")
 MISTRAL_7B_INSTRUCT_V0_2_WEIGHTS_PATH = os.path.join(
     MODEL_WEIGHTS_DIR, "mistralai/Mistral-7B-Instruct-v0.2"