EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Revert "[Perf] Add FIA interface in FA case" (#3553)

Browse Source 
Reverts vllm-project/vllm-ascend#3321
The output dimension mismatch and accuracy issue
- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

Signed-off-by: ZYang6263 <zy626375@gmail.com>
This commit is contained in:
ZYang6263
2025-10-20 19:56:10 +08:00
committed by GitHub
parent 34c2996ab8
commit b9e2896eb1
2 changed files with 12 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
vllm_ascend/attention/attention_v1.py
View File

@@ -348,50 +348,15 @@ class AscendAttentionBackendImpl(AttentionImpl):
mask = torch_npu.npu_format_cast(mask.contiguous(), mask = torch_npu.npu_format_cast(mask.contiguous(),
ACL_FORMAT_FRACTAL_NZ) ACL_FORMAT_FRACTAL_NZ)
num_tokens = query.shape[0] torch_npu._npu_flash_attention(query=query,
if torch.version.cann.startswith("8.3") and self.head_size != 256: key=key,
query_start_loc = attn_metadata.actual_seq_lengths_q value=value,
num_tokens = query_start_loc[-1] mask=mask,
softmax_lse = torch.empty(num_tokens, seq_len=attn_metadata.seq_lens,
dtype=query.dtype, scale_value=self.scale,
device=query.device) num_heads=self.num_heads,
workspace = torch_npu._npu_fused_infer_attention_score_get_max_workspace( num_kv_heads=self.num_kv_heads,
query=query, out=output)
key=key,
value=value,
atten_mask=attn_metadata.attn_mask,
input_layout="TND",
actual_seq_lengths=attn_metadata.actual_seq_lengths_q,
actual_seq_lengths_kv=attn_metadata.actual_seq_lengths_q,
num_key_value_heads=self.num_kv_heads,
num_heads=self.num_heads,
scale=self.scale,
sparse_mode=3)
torch_npu.npu_fused_infer_attention_score.out(
query=query,
key=key,
value=value,
atten_mask=attn_metadata.attn_mask,
input_layout="TND",
actual_seq_lengths=attn_metadata.actual_seq_lengths_q,
actual_seq_lengths_kv=attn_metadata.actual_seq_lengths_q,
num_key_value_heads=self.num_kv_heads,
num_heads=self.num_heads,
scale=self.scale,
sparse_mode=3,
workspace=workspace,
out=[output, softmax_lse])
else:
torch_npu._npu_flash_attention(query=query,
key=key,
value=value,
mask=mask,
seq_len=attn_metadata.seq_lens,
scale_value=self.scale,
num_heads=self.num_heads,
num_kv_heads=self.num_kv_heads,
out=output)
assert output is not None assert output is not None
return output[:num_tokens, :, :] return output[:num_tokens, :, :]

9
vllm_ascend/worker/model_runner_v1.py
View File

@@ -898,12 +898,9 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# Prefill without cache situation. # Prefill without cache situation.
elif attn_state == AscendAttentionState.PrefillNoCache: elif attn_state == AscendAttentionState.PrefillNoCache:
if torch.version.cann.startswith("8.3"): max_seq_len = max(seq_lens.max().item(), 0)
return self.attn_mask_builder.get_splitfuse_attn_mask() return self.attn_mask_builder.get_attn_mask(
else: max_seq_len, self.dtype, self.device)
max_seq_len = max(seq_lens, default=0)
return self.attn_mask_builder.get_attn_mask(
max_seq_len, self.dtype, self.device)
# Prefill with cache hit. # Prefill with cache hit.
elif attn_state == AscendAttentionState.PrefillCacheHit: elif attn_state == AscendAttentionState.PrefillCacheHit:
return self.attn_mask_builder.get_attn_mask( return self.attn_mask_builder.get_attn_mask(