optimize again

Author: qyqc731

vllm_ascend/attention/attention_mask.py

@@ -39,12 +39,13 @@ def __init__(
         self,
         max_seq_len: int,
         dtype: torch.dtype,
+        device: torch.device,
     ):
         attn_mask = _generate_attn_mask(max_seq_len, dtype)
 
         self._seq_len_cached = attn_mask.shape[0]
         self.attn_mask_cache = attn_mask
-        self.chunked_prefill_attn_mask = torch.triu(torch.ones(2048, 2048), diagonal=1).to(torch.int8)
+        self.chunked_prefill_attn_mask = torch.triu(torch.ones(2048, 2048), diagonal=1).to(torch.int8).to(device)
 
     @staticmethod
     def get_mask_scale_factor(dtype: torch.dtype = torch.float16):
@@ -67,9 +68,8 @@ def get_attn_mask(self, max_seq_len: int, dtype: torch.dtype,
 
     def get_splitfuse_attn_mask(
         self,
-        device: torch.device,
     ) -> torch.Tensor:
-        return self.chunked_prefill_attn_mask.to(device)
+        return self.chunked_prefill_attn_mask
 
     def _update_attn_cache(self, seqlen: int, dtype: torch.dtype):
         if seqlen > self._seq_len_cached:

vllm_ascend/worker/model_runner_v1.py

@@ -293,7 +293,7 @@ def __init__(self, vllm_config: VllmConfig, device: torch.device):
         )
 
         self.attn_mask_builder = AttentionMaskBuilder(
-            self.model_config.max_model_len, self.dtype)
+            self.model_config.max_model_len, self.dtype, self.device)
 
         # Set up speculative decoding.
         self.spec_attn_mask = None
@@ -797,8 +797,7 @@ def _make_attention_mask(self, seq_lens, position,
                              attn_state) -> torch.Tensor:
         # Chunk Prefill situation.
         if attn_state == AscendAttentionState.ChunkedPrefill and not self.vllm_config.model_config.use_mla:
-            return self.attn_mask_builder.get_splitfuse_attn_mask(
-                self.device)
+            return self.attn_mask_builder.get_splitfuse_attn_mask()
         # Prefill without cache situation.
         elif attn_state == AscendAttentionState.PrefillNoCache:
             max_seq_len = max(seq_lens, default=0)