[main][bugfix] bugfix for qwen3 moe quantization (#4599)

### What this PR does / why we need it?
Fix the issue where the qwen3 moe service cannot be started due to
upgrading the vllm version

Error info:
AttributeError: 'AscendFusedMoE' object has no attribute 'use dp
chunking'

### Does this PR introduce _any_ user-facing change?
no


- vLLM version: v0.11.2

---------

Signed-off-by: Wang Kunpeng <1289706727@qq.com>

examples/offline_data_parallel.py

@@ -111,6 +111,10 @@ def parse_args():
     parser.add_argument("--enable-expert-parallel",
                         action="store_true",
                         help="Enable expert parallel, used in MOE models.")
+    parser.add_argument("--quantization",
+                        type=str,
+                        default="",
+                        help="Use quantization models")
     return parser.parse_args()
 
 
@@ -134,6 +138,7 @@ def main(
     enable_expert_parallel,
     enforce_eager,
     trust_remote_code,
+    quantization,
 ):
     # DP only support on V1 engine
     os.environ["VLLM_DP_RANK"] = str(global_dp_rank)
@@ -185,6 +190,7 @@ def main(
         enforce_eager=enforce_eager,
         enable_expert_parallel=enable_expert_parallel,
         trust_remote_code=trust_remote_code,
+        quantization=quantization,
     )
     outputs = llm.generate(prompts, sampling_params)
     # Print the outputs.
@@ -220,6 +226,8 @@ if __name__ == "__main__":
     assert dp_size % node_size == 0, "dp_size should be divisible by node_size"
     dp_per_node = dp_size // node_size
 
+    quantization = args.quantization if args.quantization else None
+
     from multiprocessing import Process
 
     procs = []
@@ -238,6 +246,7 @@ if __name__ == "__main__":
                 args.enable_expert_parallel,
                 args.enforce_eager,
                 args.trust_remote_code,
+                quantization,
             ),
         )
         proc.start()

tests/e2e/multicard/test_data_parallel.py

@@ -27,13 +27,17 @@ from unittest.mock import patch
 
 import pytest
 
-MODELS = ["Qwen/Qwen3-0.6B", "Qwen/Qwen3-30B-A3B"]
+MODELS = [
+    "Qwen/Qwen3-0.6B", "Qwen/Qwen3-30B-A3B", "vllm-ascend/Qwen3-30B-A3B-W8A8"
+]
 
 
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("max_tokens", [32])
 @patch.dict(os.environ, {"ASCEND_RT_VISIBLE_DEVICES": "0,1"})
 def test_data_parallel_inference(model, max_tokens):
+    moe_models = ["Qwen/Qwen3-30B-A3B", "vllm-ascend/Qwen3-30B-A3B-W8A8"]
+    quantization_models = ["vllm-ascend/Qwen3-30B-A3B-W8A8"]
     script = "examples/offline_data_parallel.py"
 
     env = os.environ.copy()
@@ -54,8 +58,11 @@ def test_data_parallel_inference(model, max_tokens):
         "--trust-remote-code",
     ]
 
-    if model == "Qwen/Qwen3-30B-A3B":
+    if model in moe_models:
         cmd.append("--enable-expert-parallel")
+    if model in quantization_models:
+        cmd.append("--quantization")
+        cmd.append("ascend")
 
     print(f"Running subprocess: {' '.join(cmd)}")
     proc = subprocess.run(cmd,

vllm_ascend/quantization/quant_config.py

@@ -408,11 +408,10 @@ class AscendFusedMoEMethod(FusedMoEMethodBase):
         quant_config: The Ascend quantization config.
     """
 
-    def __init__(self,
+    def __init__(self, quant_config: AscendQuantConfig, prefix: str,
-                 quant_config: AscendQuantConfig,
+                 packed_modules_mapping: Dict[str,
-                 prefix: str,
+                                              Any], layer: torch.nn.Module):
-                 packed_modules_mapping: Dict[str, Any],
+        super().__init__(layer.moe_config)
-                 layer: torch.nn.Module = None):
         self.quant_method = get_quant_method(quant_config.quant_description,
                                              prefix,
                                              "moe",

vllm_ascend/torchair/ops/torchair_fused_moe.py

@@ -16,7 +16,7 @@
 # Adapted from vllm/tests/kernels/test_moe.py
 
 import os
-from typing import Any, Callable, Optional, Tuple, Union
+from typing import Any, Callable, Dict, Optional, Tuple, Union
 
 import torch
 import torch.distributed as dist
@@ -45,7 +45,9 @@ from vllm_ascend.ascend_forward_context import FusedMoEState
 from vllm_ascend.distributed.parallel_state import get_mc2_group
 from vllm_ascend.eplb.core.eplb_utils import determine_default_log2phy_map
 from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
-from vllm_ascend.quantization.quant_config import AscendFusedMoEMethod
+from vllm_ascend.quantization.quant_config import (AscendFusedMoEMethod,
+                                                   AscendQuantConfig)
+from vllm_ascend.quantization.utils import get_quant_method
 from vllm_ascend.torchair.ops.sequence_parallel import MetadataForPadding
 from vllm_ascend.torchair.utils import (get_all_reduce_merge_state,
                                         get_rm_router_logits_state,
@@ -936,6 +938,15 @@ class TorchairAscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
                 ep_group=get_ep_group())
 
 
+class TorchairAscendFusedMoEMethod(AscendFusedMoEMethod):
+
+    def __init__(self, quant_config: AscendQuantConfig, prefix: str,
+                 packed_modules_mapping: Dict[str, Any]):
+        self.quant_method = get_quant_method(quant_config.quant_description,
+                                             prefix, "moe",
+                                             packed_modules_mapping)
+
+
 class TorchairAscendFusedMoE(FusedMoE):
 
     # The moe_counter parameter is required during the initialization of EPLB
@@ -1115,7 +1126,7 @@ class TorchairAscendFusedMoE(FusedMoE):
                 self.quant_method = TorchairAscendUnquantizedFusedMoEMethod(
                     self.moe)
             else:
-                self.quant_method = AscendFusedMoEMethod(
+                self.quant_method = TorchairAscendFusedMoEMethod(
                     quant_config, prefix, quant_config.packed_modules_mapping)
 
         assert self.quant_method is not None