splitfuse access optimize

Author: qyqc731

vllm_ascend/attention/attention_mask.py

@@ -44,6 +44,7 @@ def __init__(
 
         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)
 
     @staticmethod
     def get_mask_scale_factor(dtype: torch.dtype = torch.float16):
@@ -66,24 +67,9 @@ def get_attn_mask(self, max_seq_len: int, dtype: torch.dtype,
 
     def get_splitfuse_attn_mask(
         self,
-        seq_lens: torch.Tensor,
-        position: torch.Tensor,
-        dtype: torch.dtype,
         device: torch.device,
     ) -> torch.Tensor:
-        if dtype not in [torch.float16, torch.bfloat16]:
-            raise ValueError(
-                "splitfuse_attn_mask now only supports bf16 and fp16")
-        max_seq_len = max(seq_lens, default=0)
-        self._update_attn_cache(max_seq_len, dtype)
-        # FIXME: Currently the mask value of chunked-prefill situation and Prefill-Only situation
-        # is not the same. Fix this in the future when kernel is ready.
-        mask_scale_factor = AttentionMaskBuilder.get_mask_scale_factor(dtype)
-        attn_mask = torch.index_select(self.attn_mask_cache,
-                                       dim=0,
-                                       index=position)[:, :max_seq_len]
-        attn_mask *= mask_scale_factor
-        return attn_mask.contiguous().to(device, non_blocking=True)
+        return self.chunked_prefill_attn_mask.to(device)
 
     def _update_attn_cache(self, seqlen: int, dtype: torch.dtype):
         if seqlen > self._seq_len_cached:

vllm_ascend/attention/attention_v1.py

@@ -463,7 +463,7 @@ def _forward_v1_style(
             query=query,
             key=key,
             value=value,
-            atten_mask=attn_metadata.attn_mask.to(device=query.device),
+            atten_mask=attn_metadata.attn_mask,
             block_table=attn_metadata.block_tables,
             input_layout="TND",
             block_size=block_size,

vllm_ascend/worker/model_runner_v1.py

@@ -797,7 +797,8 @@ 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 torch.triu(torch.ones(2048, 2048), diagonal=1).to(torch.int8)
+            return self.attn_mask_builder.get_splitfuse_attn_mask(
+                self.device)
         # Prefill without cache situation.
         elif attn_state == AscendAttentionState.PrefillNoCache:
             max_seq_len = max(seq_lens, default=0)