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 15, 06:00 UTC
[HOT-FIX] Add `kv_sharing_target_layer_name` argument to cutlass_mla backend (#19374)
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Pavani Majety committed
Jun 9, 2025 at 16:00 UTC
8058c91108a3611c48ef0b54448ce6b48c017f5d
1 file changed
+2
-1
vllm/v1/attention/backends/mla/cutlass_mla.py
+2
-1
@@ -40,12 +40,13 @@ class CutlassMLAImpl(MLACommonImpl[MLACommonMetadata]):
40
blocksparse_params: Optional[dict[str, Any]],
41
logits_soft_cap: Optional[float],
42
attn_type: str,
43
+ kv_sharing_target_layer_name: Optional[str],
44
# MLA Specific Arguments
45
**mla_args) -> None:
46
super().__init__(num_heads, head_size, scale, num_kv_heads,
47
alibi_slopes, sliding_window, kv_cache_dtype,
48
blocksparse_params, logits_soft_cap, attn_type,
48
- **mla_args)
49
+ kv_sharing_target_layer_name, **mla_args)
50
51
unsupported_features = [
52
alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap