fix: dbo fit on 0.9.1

Signed-off-by: 赵江江 <[email protected]>

Author: 赵江江

vllm_ascend/platform.py

@@ -27,6 +27,7 @@
 from vllm.logger import logger
 from vllm.platforms import Platform, PlatformEnum
 
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import check_ascend_config, init_ascend_config
 from vllm_ascend.utils import (ASCEND_QUATIZATION_METHOD,
                                check_torchair_cache_exist,
@@ -147,14 +148,19 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
 
         check_ascend_config(vllm_config, enforce_eager)
 
+        if vllm_config.speculative_config and envs_ascend.VLLM_ASCEND_ENABLE_DBO:
+            raise ValueError(
+                "DBO and mtp can't work at the same time. Please `export VLLM_ASCEND_ENABLE_DBO=0`"
+            )
+
         if enforce_eager or compilation_config.level == CompilationLevel.NO_COMPILATION:
             logger.info("Compilation disabled, using eager mode by default")
             compilation_config.level = CompilationLevel.NO_COMPILATION
         elif compilation_config.level != CompilationLevel.PIECEWISE:
             logger.warning(
                 "NPU does not support %s compilation level. Setting level to NO_COMPILATION",
                 compilation_config.level)
-            compilation_config.level = CompilationLevel.NO_COMPILATION
+            compilation_config.level = CompilationLevel.NfvO_COMPILATION
         elif ascend_config.torchair_graph_config.enabled:
             logger.info(
                 "Torchair compilation enabled on NPU. Setting level to NO_COMPILATION"

vllm_ascend/worker/model_runner_v1.py

@@ -219,9 +219,6 @@ def __init__(self, vllm_config: VllmConfig, device: torch.device):
         self.spec_token_num = 0
         self.decode_token_per_req = 1
         if self.speculative_config:
-            if envs_ascend.VLLM_ASCEND_ENABLE_DBO:
-                raise NotImplementedError(
-                    "DBO and mtp can't work at the same currently")
             self.use_spec_decode = True
             self.spec_token_num = self.speculative_config.num_speculative_tokens
             assert self.spec_token_num > 0