From d6ef68881e263812d4901f632786015005c4d050 Mon Sep 17 00:00:00 2001
From: iridiumine <42236072+iridiumine@users.noreply.github.com>
Date: Tue, 21 Jul 2026 09:16:48 +0800
Subject: [PATCH] [NPU] Adapt MiMo-V2.5-W8A8 (#29131)

---
 .../quantization/modelslim/modelslim.py       |  4 ++
 .../sglang/test/ascend/test_ascend_utils.py   |  1 +
 .../llm_models/test_npu_mimo_v2_5_w8a8.py     | 49 +++++++++++++++++++
 3 files changed, 54 insertions(+)
 create mode 100644 test/manual/ascend/llm_models/test_npu_mimo_v2_5_w8a8.py

diff --git a/python/sglang/srt/layers/quantization/modelslim/modelslim.py b/python/sglang/srt/layers/quantization/modelslim/modelslim.py
index 9975e6ae7..e02cdb667 100644
--- a/python/sglang/srt/layers/quantization/modelslim/modelslim.py
+++ b/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):
diff --git a/python/sglang/test/ascend/test_ascend_utils.py b/python/sglang/test/ascend/test_ascend_utils.py
index 8a70c8776..3c63bbc50 100644
--- a/python/sglang/test/ascend/test_ascend_utils.py
+++ b/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"
diff --git a/test/manual/ascend/llm_models/test_npu_mimo_v2_5_w8a8.py b/test/manual/ascend/llm_models/test_npu_mimo_v2_5_w8a8.py
new file mode 100644
index 000000000..e71088831
--- /dev/null
+++ b/test/manual/ascend/llm_models/test_npu_mimo_v2_5_w8a8.py
@@ -0,0 +1,49 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import MIMO_V2_5_W8A8_WEIGHTS_PATH
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestMiMoV25W8A8GraphWithMTP(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify the inference accuracy of MiMo-V2.5-W8A8 on GSM8K with cuda graph and MTP (speculative decoding).
+
+    [Test Category] Model
+    [Test Target] XiaomiMiMo/MiMo-V2.5-W8A8
+    [Test Config] Prefill+Decode, cuda graph enabled, EAGLE speculative decoding, modelslim quantization
+    """
+
+    model = MIMO_V2_5_W8A8_WEIGHTS_PATH
+    accuracy = 0.9
+    other_args = [
+        "--trust-remote-code",
+        "--mem-fraction-static",
+        "0.75",
+        "--attention-backend",
+        "ascend",
+        "--tp-size",
+        "8",
+        "--reasoning-parser",
+        "mimo",
+        "--speculative-algorithm",
+        "EAGLE",
+        "--speculative-num-steps",
+        "3",
+        "--speculative-eagle-topk",
+        "1",
+        "--speculative-num-draft-tokens",
+        "4",
+        "--enable-multi-layer-eagle",
+        "--quantization",
+        "modelslim",
+        "--speculative-draft-model-quantization",
+        "unquant",
+        "--dp-size",
+        "2",
+        "--enable-dp-attention",
+        "--enable-dp-lm-head",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()