ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-06 12:00:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

Graph parallel for dense Qwen-3.5 models (#1331)

Browse Source 
* Graph parallel for idense Qwen-3.5 models

* Cleanup
This commit is contained in:
Kawrakow
2026-02-27 07:03:25 +01:00
committed by GitHub
parent facc8fdc44
commit 1e6d36b1b4
3 changed files with 28 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llama-build-context.cpp
View File

@@ -4649,7 +4649,19 @@ ggml_cgraph * llm_build_context::build_qwen35() {
if (hparams.is_recurrent(il)) {
ggml_tensor * inpSA = inpL;
cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, cb, il);
int idx = model.default_layer_device[il];
if (inpL->op == GGML_OP_REDUCE) {
if (kv_self.s_l[il]) {
// This shouldn't be necessary, but just in case.
int idx_s_l = ggml_backend_sched_get_backend_idx(lctx.sched, kv_self.s_l[il]->buffer);
if (idx_s_l >= 0) idx = idx_s_l;
}
if (inpL->src[idx]) {
inpL->view_src = inpL->src[idx];
}
}
auto norm = model.layers[il].attn_norm->extra ? ((ggml_split_tensor_t *)model.layers[il].attn_norm->extra)->splits[idx] : model.layers[il].attn_norm;
cur = llm_build_norm(ctx0, inpL, hparams, norm, nullptr, LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
cur = delta.build_layer_attn_linear(ctx0, gf, cur, causal_mask, identity, diag_mask, il, cb);
if (il == n_layer - 1 && inp_out_ids) {