ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Move GLM4 f32 attention fix to the correct function (#13750)

Browse source
This commit is contained in:
0cc4m 2025-05-24 16:49:12 +02:00 • committed by GitHub
parent 4c32832c59
commit 259469c4b5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/llama-graph.cpp
View file

@ -1287,6 +1287,10 @@ ggml_tensor * llm_graph_context::build_attn(
if (wo) { if (wo) {
cur = build_lora_mm(wo, cur); cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4) {
// GLM4 seems to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
} }
if (wo_b) { if (wo_b) {
@ -1367,10 +1371,6 @@ ggml_tensor * llm_graph_context::build_attn(
if (wo) { if (wo) {
cur = build_lora_mm(wo, cur); cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4) {
// GLM4 seems to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
} }
if (wo_b) { if (wo_b) {