llama.cpp
TP: fix split state and granularity for fused QKV gemma4, qwen35
#28965
Merged
Merged
TP: fix split state and granularity for fused QKV gemma4, qwen35 #28965
ggerganov
merged
3 commits into
ggml-org:master
from
dfriehs:tp-fused-qkv
model: calculate split states for attn_qkv from n_head * n_embd_head_k
3da324e3
model: handle fused full attention layers for qwen35/qwen35moe
b1a9fc96
CISC
approved these changes
on 2026-09-16
model: add TODO: [TAG_SPLIT_QGATE_QWEN]
c2651dc8
Login to write a write a comment.
Login via GitHub