ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-06 14:30:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

WIP

Browse Source
This commit is contained in:
Kawrakow
2025-11-26 12:11:27 +00:00
parent 5d68e4eb35
commit 4303587f1c
2 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
src/llama-build-context.cpp
View File

@@ -634,7 +634,7 @@ ggml_tensor * llm_build_context::llm_build_ffn(
ggml_tensor * act_scales,
llm_ffn_op_type type_op,
llm_ffn_gate_type type_gate,
const llm_build_cb & cb, int il) {
const llm_build_cb & cb, int il, ggml_cgraph * graph) {
if (!up_b && !up_s && !gate_b && !gate_s && !down_b && !down_s &&
up->extra && gate->extra && down->extra && type_gate == LLM_FFN_PAR &&
@@ -662,6 +662,9 @@ ggml_tensor * llm_build_context::llm_build_ffn(
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
if (graph) {
ggml_build_forward_expand(graph, cur);
}
ffn.push_back(cur);
}
if (ffn.size() == 1) return ffn.front();
@@ -1526,7 +1529,7 @@ ggml_cgraph * llm_build_context::build_llama() {
model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
LLM_FFN_SILU, LLM_FFN_PAR, cb, il, gf);
cb(cur, "ffn_out", il);
} else if (model.arch == LLM_ARCH_LLAMA4) {
// llama4 MoE
@@ -9152,6 +9155,7 @@ ggml_tensor * llm_build_context::build_std_attention(ggml_cgraph * gf, ggml_tens
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
cb(cur, "kqv_wo", il_cb);
ggml_build_forward_expand(gf, cur);
// TODO: wo_b
attn.push_back(cur);
}

2
src/llama-build-context.h
View File

@@ -331,7 +331,7 @@ struct llm_build_context {
ggml_tensor * act_scales,
llm_ffn_op_type type_op,
llm_ffn_gate_type type_gate,
const llm_build_cb & cb, int il);
const llm_build_cb & cb, int il, ggml_cgraph * graph = nullptr);
static ggml_tensor * llm_build_moe_ffn(ggml_context * ctx, llama_context & lctx,
ggml_tensor * cur,