xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-09 17:45:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

hotfix attn alibi wo head mapping (#496)

Browse Source 
Co-authored-by: oliveryuan <oliveryuan@basemind.com>
This commit is contained in:
Song 2023-07-19 02:31:48 +08:00 committed by GitHub
parent 453bafb96f
commit bda41c70dd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 3 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tests/kernels/test_attention.py
View File

@ -199,6 +199,7 @@ def run_single_query_cached_kv_attention(
]
block_tables.append(block_table)
block_tables = torch.tensor(block_tables, dtype=torch.int, device='cuda')
head_mapping = torch.arange(num_heads, dtype=torch.int32, device="cuda")
scale = float(1.0 / (head_size**0.5))
output = torch.empty(num_tokens,
@ -211,6 +212,7 @@ def run_single_query_cached_kv_attention(
query,
key_cache,
value_cache,
head_mapping,
scale,
block_tables,
context_lens,

1
vllm/model_executor/layers/attention.py
View File

@ -408,6 +408,7 @@ class PagedAttentionWithALiBi(PagedAttention):
query,
key_cache,
value_cache,
self.head_mapping,
self.scale,
input_metadata.block_tables,
input_metadata.context_lens,