ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 09:09:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix split mode graph for GPT-OSS with partial offload (#1128)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2026-01-10 07:57:43 +02:00
committed by GitHub
parent 08a0da389c
commit dd3c3f72f2
1 changed files with 14 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
src/llama-load-tensors.cpp
View File

@@ -2583,6 +2583,7 @@ bool create_tensors_helper::create_openai_moe_tensors(const LLM_TN & tn) {
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_split = ctx_for_layer_split(i);
ggml_context * ctx_layer = ctx_for_layer(i);
auto & layer = model.layers[i];
layer.attn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
@@ -2602,11 +2603,14 @@ bool create_tensors_helper::create_openai_moe_tensors(const LLM_TN & tn) {
layer.ffn_up_exps = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0, &ctx_ffn_up);
// bias
ggml_context *ctx_ffn_gate_b, *ctx_ffn_up_b, *ctx_ffn_down_b;
layer.ffn_gate_inp_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_INP, "bias", i), {n_expert}, 0);
layer.ffn_gate_exps_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0, &ctx_ffn_gate_b);
layer.ffn_down_exps_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "bias", i), { n_embd, n_expert}, 0, &ctx_ffn_down_b);
layer.ffn_up_exps_b = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS, "bias", i), {n_ff_exp, n_expert}, 0, &ctx_ffn_up_b);
ggml_context *ctx_ffn_gate_b, *ctx_ffn_up_b, *ctx_ffn_down_b;
auto ctx_gate_b = ctx_ffn_gate == ctx_split ? ctx_split : ctx_layer;
auto ctx_down_b = ctx_ffn_down == ctx_split ? ctx_split : ctx_layer;
auto ctx_up_b = ctx_ffn_up == ctx_split ? ctx_split : ctx_layer;
layer.ffn_gate_exps_b = create_tensor(ctx_gate_b, tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert}, 0, &ctx_ffn_gate_b);
layer.ffn_down_exps_b = create_tensor(ctx_down_b, tn(LLM_TENSOR_FFN_DOWN_EXPS, "bias", i), { n_embd, n_expert}, 0, &ctx_ffn_down_b);
layer.ffn_up_exps_b = create_tensor(ctx_up_b, tn(LLM_TENSOR_FFN_UP_EXPS, "bias", i), {n_ff_exp, n_expert}, 0, &ctx_ffn_up_b);
if (ctx_ffn_gate_b != ctx_ffn_gate) {
layer.ffn_gate_exps_b_dup = create_tensor(ctx_ffn_gate, tn(LLM_TENSOR_FFN_GATE_EXPS, "bias", i), {n_ff_exp, n_expert},
@@ -3169,9 +3173,6 @@ bool create_tensors_helper::create_tensors() {
if (layer.ffn_gate_exps_b) {
prepare_split_tensors( 0, ctx_split, layer.ffn_gate_exps_b, layer.split_ffn_gate_exps_b, split, mem_used);
}
if (layer.ffn_gate_inp_b) {
prepare_split_tensors(-1, ctx_split, layer.ffn_gate_inp_b, layer.split_ffn_gate_inp_b, split, mem_used);
}
}
}
@@ -3181,6 +3182,12 @@ bool create_tensors_helper::create_tensors() {
prepare_split_tensors(-1, ctx_split, layer.ffn_gate_inp, layer.split_ffn_gate_inp, shared_split, mem_used);
}
}
if (layer.ffn_gate_inp_b) {
if (auto it = split_tensors.find(layer.ffn_gate_inp_b); it != split_tensors.end()) {
auto shared_split = create_split(ggml_nrows(layer.ffn_gate_inp_b), -1, cur_splits, mem_used);
prepare_split_tensors(-1, ctx_split, layer.ffn_gate_inp_b, layer.split_ffn_gate_inp_b, shared_split, mem_used);
}
}
if (layer.ffn_exp_probs_b) {
if (auto it = split_tensors.find(layer.ffn_exp_probs_b); it != split_tensors.end()) {
auto shared_split = create_split(ggml_nrows(layer.ffn_exp_probs_b), -1, cur_splits, mem_used);