ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-26 01:19:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

Split mode graph for Qwen3 (#1106)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2026-01-05 14:31:36 +02:00
committed by GitHub
parent cac2b046f0
commit 359cf817a9
3 changed files with 51 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/llama-load-tensors.cpp
View File

@@ -1116,8 +1116,8 @@ bool create_tensors_helper::create_qwen3_tensors(const LLM_TN & tn) {
// output
{
model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
model.output_norm = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (model.output == NULL) {
model.output = create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
@@ -1125,21 +1125,20 @@ bool create_tensors_helper::create_qwen3_tensors(const LLM_TN & tn) {
}
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(i);
auto & layer = model.layers[i];
layer.attn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.attn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
use_mmap_buffer &= !merge_qkv(tn, i, 0);
layer.wo = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd});
layer.attn_k_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k});
layer.attn_q_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k});
layer.attn_k_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k});
layer.attn_q_norm = create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k});
layer.ffn_norm = create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_norm = create_tensor(ctx_split, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
create_std_ffn(i, tn, layer, n_ff, n_embd, ctx_split);
}
return use_mmap_buffer;