graph : add precision handling for MELLUM architecture (#30584)
This commit is contained in:
1 parent
b797b9ca81
commit
5e03a1ff45
1 file changed
+1
-1
+1
-1
@@ -2452,7 +2452,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
|
||||
}
|
||||
|
||||
experts = build_lora_mm_id(down_exps, cur, selected_experts, down_exps_s, slots); // [n_embd, n_expert_used, n_tokens]
|
||||
if (arch == LLM_ARCH_MISTRAL4) {
|
||||
if (arch == LLM_ARCH_MISTRAL4 || arch == LLM_ARCH_MELLUM) {
|
||||
// src1 can exceed F16 range
|
||||
ggml_prec_set_src(experts, GGML_PREC_F32, 1);
|
||||
}
|
||||
|
||||
Reference in new issue
Block a user