[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 <akaratza@amd.com>
Harry Mellor committed
Aug 11, 2026 at 20:22 UTC
b2ab0960178cc737d8537208f1e1de79404ca1ac
1 file changed
+1
-1
vllm/v1/attention/ops/turboquant_soa/triton_turboquant_decode_v2.py
+1
-1
@@ -6,7 +6,7 @@ FLUTE-paper optimizations applied:
6
1. Grouped Q heads: grid over (B, Hk, splits) instead of (B, Hq, splits).
7
Each program loads BLOCK_M Q heads sharing a KV head into a 2D tile,
8
enabling tl.dot on tensor cores (MFMA/WMMA) for both Q·K and P·V.
9
- 2. Vectorized pair LUT: precompute pair_table[i][j] = (T[i], T[j])
9
+ 2. Vectorized pair LUT: precompute `pair_table[i][j] = (T[i], T[j])`
10
offline. At runtime, extract adjacent index pairs and fetch two
11
dequantized centroids with a single gather, halving LUT lookups.
12
3. exp2 instead of exp: scores pre-scaled by log2(e) so the hardware-