[Perf] Tune Triton GDN decode tiles on SM100-family GPUs (#61355)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Thien Tran
2026-10-13 17:01:12 +08:00
committed by GitHub
co-authored by OpenAI Codex
parent 183625d4f9
commit 656d0d4506
2 changed files with 52 additions and 14 deletions
@@ -10,6 +10,7 @@
import torch
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from .op import exp, log
@@ -439,6 +440,17 @@ def fused_recurrent_gated_delta_rule_packed_decode(
f"Packed decode kernel only supports NK=1 (got K={K}, BK={BK})."
)
BV = min(triton.next_power_of_2(V), 32)
if (
K == V == 128
and current_platform.is_cuda()
and current_platform.is_device_capability_family(100, device_id=dev.index)
):
# Tuned on GB300; validated on GB200 and with 148 SMs.
# MRv2 decode warmup/graph capture compiles the reachable BV variants.
if B * HV <= 64:
BV = 4
elif initial_state.dtype == torch.float32:
BV = 8
num_stages = 3
num_warps = 1