ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-21 13:44:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

qwen3next: disable flash-attn for cpu-only contexts

Browse Source
This commit is contained in:
yurko
2026-02-08 01:04:38 -08:00
parent a822db6f18
commit 627d46912c
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/llama.cpp
View File

@@ -4508,6 +4508,13 @@ struct llama_context * llama_new_context_with_model(
params.flash_attn = false;
}
// Qwen3Next currently has a CPU-only flash-attn assertion path in iqk FA kernels.
// Keep CPU runs safe by disabling FA when no layers are offloaded.
if (params.flash_attn && model->arch == LLM_ARCH_QWEN3NEXT && model->n_gpu_layers == 0) {
LLAMA_LOG_WARN("%s: flash_attn is unstable for CPU-only Qwen3Next - forcing off\n", __func__);
params.flash_attn = false;
}
//if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
// LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
// params.flash_attn = false;