From b2ab0960178cc737d8537208f1e1de79404ca1ac Mon Sep 17 00:00:00 2001 From: Harry Mellor <19981378+hmellor@users.noreply.github.com> Date: Tue, 11 Aug 2026 20:22:56 +0100 Subject: [PATCH] [Docs] Fix broken autorefs cross-reference in TurboQuant v2 docstring (#51857) Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Andreas Karatzas --- .../attention/ops/turboquant_soa/triton_turboquant_decode_v2.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/v1/attention/ops/turboquant_soa/triton_turboquant_decode_v2.py b/vllm/v1/attention/ops/turboquant_soa/triton_turboquant_decode_v2.py index 99198f7a23..321f42bcb1 100644 --- a/vllm/v1/attention/ops/turboquant_soa/triton_turboquant_decode_v2.py +++ b/vllm/v1/attention/ops/turboquant_soa/triton_turboquant_decode_v2.py @@ -6,7 +6,7 @@ FLUTE-paper optimizations applied: 1. Grouped Q heads: grid over (B, Hk, splits) instead of (B, Hq, splits). Each program loads BLOCK_M Q heads sharing a KV head into a 2D tile, enabling tl.dot on tensor cores (MFMA/WMMA) for both Q·K and P·V. - 2. Vectorized pair LUT: precompute pair_table[i][j] = (T[i], T[j]) + 2. Vectorized pair LUT: precompute `pair_table[i][j] = (T[i], T[j])` offline. At runtime, extract adjacent index pairs and fetch two dequantized centroids with a single gather, halving LUT lookups. 3. exp2 instead of exp: scores pre-scaled by log2(e) so the hardware-