ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-23 14:44:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Disable some fusion and make rope cahe off by default

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-11-04 07:42:55 +02:00
parent 1cfd19862f
commit 45601bc1bf
4 changed files with 6 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
common/common.cpp
View File

@@ -1106,8 +1106,8 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
params.fused_mmad = false;
return true;
}
if (arg == "-no-rcache" || arg == "--no-rope-cache") {
params.rope_cache = false;
if (arg == "-rcache" || arg == "--rope-cache") {
params.rope_cache = true;
return true;
}
if (arg == "-ser" || arg == "--smart-expert-reduction") {

2
common/common.h
View File

@@ -249,7 +249,7 @@ struct gpt_params {
bool fused_up_gate = true; // fused up*unary(gate) op
bool fused_mmad = true; // fused mul+multi_add op
bool grouped_expert_routing = false; // if to use grouped expert routing (BailingMoeV2 arch)
bool rope_cache = true; // if to use RoPE cache (for supported models)
bool rope_cache = false; // if to use RoPE cache (for supported models)
int min_experts = -1;
float thresh_experts = 0;

8
ggml/src/ggml-cuda.cu
View File

@@ -3244,11 +3244,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
ggml_cuda_op_rms_norm(ctx, dst);
break;
case GGML_OP_FUSED_RMS_NORM:
//if (i + 6 < cgraph->n_nodes) {
// printf("=== Fused rms_norm(%s)\n", dst->name);
// for (int j = 1; j <= 6; ++j) printf(" %s(%s)\n", ggml_op_name(cgraph->nodes[i+j]->op), cgraph->nodes[i+j]->name);
//}
if (ENABLE_FUSION && i + 4 < cgraph->n_nodes &&
if (false && ENABLE_FUSION && i + 4 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_VIEW &&
cgraph->nodes[i+2]->op == GGML_OP_FUSED_RMS_NORM &&
cgraph->nodes[i+3]->op == GGML_OP_ROPE_FAST &&
@@ -3256,7 +3252,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
ggml_cuda_op_fused_rms_rope_fast(ctx, cgraph->nodes[i+3], cgraph->nodes[i+4])) {
i += 4;
}
else if (ENABLE_FUSION && i + 4 < cgraph->n_nodes &&
else if (false && ENABLE_FUSION && i + 4 < cgraph->n_nodes &&
cgraph->nodes[i+1]->op == GGML_OP_ROPE_FAST &&
cgraph->nodes[i+2]->op == GGML_OP_RESHAPE &&
cgraph->nodes[i+3]->op == GGML_OP_FUSED_RMS_NORM &&

2
src/llama.cpp
View File

@@ -3833,7 +3833,7 @@ struct llama_context_params llama_context_default_params() {
/*.grouped_expert_routing =*/ false,
/*.fused_up_gate =*/ true,
/*.fused_mmad =*/ true,
/*.rope_cache =*/ true,
/*.rope_cache =*/ false,
/*.min_experts =*/ -1,
/*.thtesh_experts =*/ 0.0f,
/*.only_active_experts =*/ false,