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

[Fixbugs]: fix refactor cause to 310p chunkprefill error (#6340)

Browse Source 
Adapt modelrunner refactor change to make 310p work

- vLLM version: v0.14.1
- vLLM main:
dc917cceb8

Signed-off-by: Tflowers-0129 <2906339855@qq.com>
This commit is contained in:
Shaoxu Cheng
2026-01-28 16:41:32 +08:00
committed by GitHub
parent 325cb16e3f
commit 9fadc8df4f
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
vllm_ascend/_310p/attention/attention_v1.py
View File

@@ -101,8 +101,7 @@ class AscendAttentionBackendImpl310(_BaseImpl):
out=output, out=output,
) )
out_real = output[:real_tokens, :, :] return output[:aligned_tokens, :, :]
return out_real
def _forward_chunked_prefill_310p(self, query, attn_metadata, output): def _forward_chunked_prefill_310p(self, query, attn_metadata, output):
assert attn_metadata is not None assert attn_metadata is not None
@@ -110,6 +109,10 @@ class AscendAttentionBackendImpl310(_BaseImpl):
if query.dtype == torch.float32: if query.dtype == torch.float32:
query = query.to(torch.float16) query = query.to(torch.float16)
num_actual_tokens = int(attn_metadata.num_actual_tokens)
query = query[:num_actual_tokens]
output = output[:num_actual_tokens]
qsl_cpu = attn_metadata.query_start_loc.detach().to("cpu", dtype=torch.int32) qsl_cpu = attn_metadata.query_start_loc.detach().to("cpu", dtype=torch.int32)
qlens = (qsl_cpu[1:] - qsl_cpu[:-1]).to(torch.int32) qlens = (qsl_cpu[1:] - qsl_cpu[:-1]).to(torch.int32)
@@ -163,8 +166,7 @@ class AscendAttentionBackendImpl310(_BaseImpl):
k = key[:num_tokens] k = key[:num_tokens]
v = value[:num_tokens] v = value[:num_tokens]
out = self._forward_prefill_310p_fallback(q, k, v, attn_metadata, output) out = self._forward_prefill_310p_fallback(q, k, v, attn_metadata, output)
output[:num_tokens] = out return out
return output
if state == AscendAttentionState.ChunkedPrefill: if state == AscendAttentionState.ChunkedPrefill:
self._forward_chunked_prefill_310p(query, attn_metadata, output) self._forward_chunked_prefill_310p(query, attn_metadata, output)