Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

fix(sdnq): sync benchmark prep mirror with kernel helpers

Browse Source 
upstream replaced math.log2/triton.next_power_of_2 with sdnq utils in
the attention prep; mirror the same hadamard sizing in make_prep_fn.
This commit is contained in:
CalamitousFelicitousness
2026-07-09 17:42:49 +01:00
parent 75a245cbeb
commit d2937166d3
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
cli/sdnq-attention-benchmark.py
+2 -2
View File
@@ -478,15 +478,15 @@ def run_correctness():
def make_prep_fn(q, k, v, attn_mask, kwargs):
# mirror sdnq_triton_atten's prep call so the prep column measures the same code path
import triton
from modules.sdnq.kernels import triton_atten as atten_module
from modules.sdnq.quant_utils import get_hadamard, get_hadamard_group_size
from modules.sdnq.utils import next_power_of_2
matmul_dtype = kwargs.get("matmul_dtype", "int8")
do_quantize = kwargs.get("do_quantize", True)
hadamard_group_size = kwargs.get("hadamard_group_size", 256)
hadamard = None
if kwargs.get("use_hadamard", False) and do_quantize and matmul_dtype not in {None, "none", "no"}:
channel_size = min(triton.next_power_of_2(q.shape[-1]), triton.next_power_of_2(k.shape[-1]))
channel_size = next_power_of_2(min(q.shape[-1], k.shape[-1]))
hadamard_group_size = min(hadamard_group_size, channel_size)
enabled, hadamard_group_size = get_hadamard_group_size(channel_size, hadamard_group_size)
if enabled: