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 10, 16:30 UTC
[PERF] Speed-up of GDN attention decode part (Qwen3-Next) (#31722)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Vadim Gimpelson committed
Jan 6, 2026 at 21:32 UTC
22dffca9822987f0e912bfd9635e94bbdd05def3
1 file changed
+1
-1
vllm/model_executor/layers/fla/ops/fused_recurrent.py
+1
-1
@@ -189,7 +189,7 @@ def fused_recurrent_gated_delta_rule_fwd(
189
B, T, H, K, V = *k.shape, v.shape[-1]
190
HV = v.shape[2]
191
N = B if cu_seqlens is None else len(cu_seqlens) - 1
192
- BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
192
+ BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 32)
193
NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
194
assert NK == 1, "NK > 1 is not supported yet"
195
num_stages = 3