A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Misaligned params in TreeAttentionImpl (#22226)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Aug 5, 2025 at 13:40 UTC
cdfd6871a5c4f125c9b3707ec5c1260db54f4b03
1 file changed
+1
-5
vllm/v1/attention/backends/tree_attn.py
+1
-5
@@ -4,7 +4,7 @@
4
5
import ast
6
from dataclasses import dataclass
7
-from typing import TYPE_CHECKING, Any, Optional
7
+from typing import TYPE_CHECKING, Optional
8
9
import torch
10
@@ -313,15 +313,11 @@ class TreeAttentionImpl(AttentionImpl):
313
alibi_slopes: Optional[list[float]],
314
sliding_window: Optional[int],
315
kv_cache_dtype: str,
316
- blocksparse_params: Optional[dict[str, Any]] = None,
316
logits_soft_cap: Optional[float] = None,
317
attn_type: AttentionType = AttentionType.DECODER,
318
kv_sharing_target_layer_name: Optional[str] = None,
319
use_irope: bool = False,
320
) -> None:
322
- if blocksparse_params is not None:
323
- raise ValueError(
324
- "TreeAttention does not support block-sparse attention.")
321
self.num_heads = num_heads
322
self.head_size = head_size
323
self.scale = float(scale)