[Fixbug] Fix soc_version for 310p

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

Author: zhangxinyuehfad

tests/ut/worker/test_model_runner_v1.py

@@ -24,25 +24,25 @@
     "soc_version, enable_expert_parallel, world_size, num_tokens, mc2_tokens_capacity, quant_type, expected_method",
     [
         # Case 1: Expert parallel is disabled, should always be 'allgather'
-        ("A2", False, 8, 100, 256, None, "allgather"),
-        ("A3", False, 16, 500, 256, None, "allgather"),
+        ("A2", False, 8, 100, 256, None, MoECommType.ALLGATHER),
+        ("A3", False, 16, 500, 256, None, MoECommType.ALLGATHER),
 
         # Case 2: A2 SOC with w4a8_dynamic -> use alltoall when not mc2
-        ("A2", True, 8, 100, 256, "w4a8_dynamic", "alltoall"),
-        ("A2", True, 16, 257, 256, "w4a8_dynamic", "alltoall"),
-        ("A2", True, 16, 100, 256, "w4a8_dynamic", "mc2"),  # meets mc2 condition
+        ("A2", True, 8, 100, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
+        ("A2", True, 16, 257, 256, "w4a8_dynamic", MoECommType.ALLTOALL),
+        ("A2", True, 16, 100, 256, "w4a8_dynamic", MoECommType.MC2),  # meets mc2 condition
 
         # Case 3: A2 SOC without w4a8_dynamic -> fallback to allgather
-        ("A2", True, 8, 100, 256, None, "allgather"),
-        ("A2", True, 16, 257, 256, None, "allgather"),
+        ("A2", True, 8, 100, 256, None, MoECommType.ALLGATHER),
+        ("A2", True, 16, 257, 256, None, MoECommType.ALLGATHER),
 
         # Case 4: A3 SOC
-        ("A3", True, 8, 100, 256, None, "mc2"),
-        ("A3", True, 8, 257, 256, None, "alltoall"),
+        ("A3", True, 8, 100, 256, None, MoECommType.MC2),
+        ("A3", True, 8, 257, 256, None, MoECommType.ALLTOALL),
 
         # Case 5: P3 SOC
-        ("310P", True, 8, 100, 256, None, "allgather"),
-        ("310P", True, 8, 257, 256, None, "allgather"),
+        ("310P", True, 8, 100, 256, None, MoECommType.ALLGATHER),
+        ("310P", True, 8, 257, 256, None, MoECommType.ALLGATHER),
     ])
 # yapf: enable
 def test_select_moe_comm_method(soc_version, enable_expert_parallel,

vllm_ascend/__init__.py

@@ -15,6 +15,10 @@
 # This file is a part of the vllm-ascend project.
 #
 
+import logging
+
+logger = logging.getLogger(__name__)
+
 # Supported SOC_VERSION codes
 ASCEND_A2_SOC_VERSION = range(220, 226)
 ASCEND_A3_SOC_VERSION = range(250, 256)
@@ -28,20 +32,33 @@ def register():
 
 
 def register_model():
-    import torch_npu  # type: ignore
-
-    from vllm_ascend import _build_info  # type: ignore
-    raw = torch_npu.npu.get_soc_version()
-    soc_version = (
-        "ASCEND910B1" if raw in ASCEND_A2_SOC_VERSION else
-        "ASCEND910_9392" if raw in ASCEND_A3_SOC_VERSION else
-        "ASCEND310P3" if raw in ASCEND_310P_SOC_VERSION else "UNDEFINED")
-    if soc_version == "UNDEFINED":
-        raise RuntimeError("Unsupported or undefined Ascend SOC version.")
-    elif soc_version != _build_info.__soc_version__:
-        raise RuntimeError(
-            f"Built for SOC version {_build_info.__soc_version__}, but need running on default {soc_version}. Please reinstall vllm-ascend with the default SOC_VERSION."
-        )
+    try:
+        import torch_npu  # type: ignore
+    except ImportError:
+        logger.warning("torch_npu not found. Proceeding with CPU testing.")
     else:
-        from .models import register_model
-        register_model()
+        from vllm_ascend import _build_info  # type: ignore
+        ascend_soc_version = getattr(_build_info, "__ascend_soc_version__",
+                                     None)
+        raw = torch_npu.npu.get_soc_version()
+        soc_version = (
+            "ASCEND910B1" if raw in ASCEND_A2_SOC_VERSION else
+            "ASCEND910_9392" if raw in ASCEND_A3_SOC_VERSION else
+            "ASCEND310P3" if raw in ASCEND_310P_SOC_VERSION else "UNDEFINED")
+
+        if soc_version == "UNDEFINED":
+            raise RuntimeError("Unsupported or undefined Ascend SOC version.")
+        elif soc_version != _build_info.__soc_version__:
+            raise RuntimeError(
+                f"Built for SOC version {_build_info.__soc_version__}, "
+                f"but running on SOC version {soc_version}. "
+                f"Please reinstall vllm-ascend with 'export SOC_VERSION={soc_version}' ."
+            )
+        elif ascend_soc_version is None:
+            raise RuntimeError(
+                "vllm_ascend._build_info is missing '__ascend_soc_version__'. "
+                f"Please reinstall vllm-ascend with 'export SOC_VERSION={soc_version}' ."
+            )
+
+    from .models import register_model
+    register_model()

vllm_ascend/torchair/ops/torchair_fused_moe.py

@@ -48,7 +48,8 @@
 from vllm_ascend.quantization.quant_config import AscendFusedMoEMethod
 from vllm_ascend.torchair.utils import npu_stream_switch, npu_wait_tensor
 from vllm_ascend.utils import (dispose_tensor, get_all_reduce_merge_state,
-                               get_rm_router_logits_state, is_310p)
+                               get_rm_router_logits_state, is_310p,
+                               vllm_version_is)
 
 
 def torchair_fused_experts_with_mc2(

vllm_ascend/utils.py

@@ -23,7 +23,6 @@
 import os
 from contextlib import contextmanager, nullcontext
 from dataclasses import dataclass
-from enum import Enum
 from threading import Lock
 from typing import TYPE_CHECKING, Any, List, Optional, Tuple, Union
 

vllm_ascend/worker/model_runner_v1.py

@@ -116,8 +116,7 @@
 from vllm_ascend.spec_decode.interface import SpecDcodeType
 from vllm_ascend.spec_decode.mtp_proposer import MtpProposer
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, ACL_FORMAT_FRACTAL_NZ,
-                               AscendSocVersion, ProfileExecuteDuration,
-                               get_ascend_soc_version, get_graph_params,
+                               ProfileExecuteDuration, get_graph_params,
                                is_310p, lmhead_tp_enable, set_graph_params,
                                vllm_version_is)
 from vllm_ascend.worker.npu_input_batch import CachedRequestState, InputBatch