[main] remove dbo code

Signed-off-by: zzzzwwjj <[email protected]>

Author: zzzzwwjj

examples/offline_dualbatch_overlap_npu.py

@@ -213,9 +213,6 @@ class AscendMetadata:
     # (num_tokens,)
     slot_mapping: torch.Tensor = None
 
-    # *************************** Other Properties *************************** #
-    enable_dbo_across_dp: bool = False
-
     prefill: Optional[AscendMetadataForPrefill] = None
 
     decode_meta: Optional[AscendMetadataForDecode] = None
@@ -374,7 +371,6 @@ def build(
             slot_mapping=slot_mapping,
             attn_mask=attn_mask,
             attn_state=attn_state,
-            enable_dbo_across_dp=common_attn_metadata.enable_dbo_across_dp,
             num_prefills=num_prefills,
             num_decodes=num_decodes,
             prefill=prefill_metadata,

vllm_ascend/attention/attention_v1.py

@@ -184,7 +184,6 @@ class AscendMLAMetadata:
 
     decode: Optional[AscendMLADecodeMetadata] = None
     prefill: Optional[AscendMLAPrefillMetadata] = None
-    enable_dbo_across_dp: bool = False
 
     def __post_init__(self):
         pass
@@ -538,7 +537,6 @@ def build(
             query_start_loc=query_start_loc,
             block_tables=block_table,
             seq_lens=seq_lens,
-            enable_dbo_across_dp=common_attn_metadata.enable_dbo_across_dp,
         )
 
     def build_for_graph_capture(

vllm_ascend/attention/mla_v1.py

@@ -138,7 +138,6 @@ class AscendSFAMetadata:
 
     decode: Optional[AscendSFADecodeMetadata] = None
     prefill: Optional[AscendSFAPrefillMetadata] = None
-    enable_dbo_across_dp: bool = False
 
     def __post_init__(self):
         pass
@@ -434,7 +433,6 @@ def build(
             query_start_loc=query_start_loc,
             block_tables=block_table,
             seq_lens=seq_lens,
-            enable_dbo_across_dp=common_attn_metadata.enable_dbo_across_dp,
         )
 
 

vllm_ascend/attention/sfa_v1.py

@@ -91,8 +91,6 @@ class AscendCommonAttentionMetadata:
 
     attn_state: Any = None
 
-    enable_dbo_across_dp: bool = False
-
     is_only_prefill: bool = False
 
     graph_pad_size: int = -1

vllm_ascend/attention/utils.py

@@ -82,9 +82,6 @@
     "VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP":
     lambda: bool(int(os.getenv("VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP", '0'))
                  ),
-    # Whether to enable DBO feature for deepseek model.
-    "VLLM_ASCEND_ENABLE_DBO":
-    lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_DBO", '0'))),
     # Whether to enable the model execute time observe profile. Disable it when
     # running vllm ascend in production environment.
     "VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE":