ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-07 04:20:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make sure we pick the reduced tensor from the right GPU

Browse Source
This commit is contained in:
Kawrakow
2026-02-23 09:37:34 +00:00
parent 68bd30d99c
commit fd2a70913c
4 changed files with 45 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
src/llama-build-context.cpp
View File

@@ -1986,9 +1986,12 @@ static ggml_tensor * build_output(llama_context & lctx, ggml_context * ctx, ggml
}
}
} else {
if (cur->op == GGML_OP_REDUCE && cur->src[lctx.model.main_gpu]) {
int idx = lctx.model.default_layer_device[lctx.model.hparams.n_layer];
int idx_out = ggml_backend_sched_get_backend_idx(lctx.sched, lctx.model.output->buffer);
if (idx_out >= 0) idx = idx_out;
if (cur->op == GGML_OP_REDUCE && cur->src[idx]) {
// avoid copy to main GPU
cur->view_src = cur->src[lctx.model.main_gpu];
cur->view_src = cur->src[idx];
}
if (output_norm) {
cur = llm_build_context::llm_build_norm(ctx, cur, lctx.model.hparams, output_norm, NULL, LLM_NORM_RMS, cb, -1);
@@ -4458,12 +4461,33 @@ ggml_cgraph * llm_build_context::build_qwen3next() {
if (hparams.is_recurrent(il)) {
if (inpL->op == GGML_OP_REDUCE && inpL->src[model.default_layer_device[il]]) {
inpL->view_src = inpL->src[model.default_layer_device[il]];
//printf("Using reduce result on device %d\n", model.default_layer_device[il]);
//inpL = inpL->src[model.default_layer_device[il]];
int idx = model.default_layer_device[il];
if (inpL->op == GGML_OP_REDUCE) {
if (model.layers[il].wqkv) {
int idx_wqkv = ggml_backend_sched_get_backend_idx(lctx.sched, model.layers[il].wqkv->buffer);
//int idx_s_l = -1;
//if (kv_self.s_l[il]) idx_s_l = ggml_backend_sched_get_backend_idx(lctx.sched, kv_self.s_l[il]->buffer);
//printf("Layer %2d: %d, %d, %d\n", il, idx, idx_wqkv, idx_s_l);
if (idx_wqkv >= 0) idx = idx_wqkv;
}
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[model.default_layer_device[il]] : model.layers[il].attn_norm;
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;
//if (inpL->op == GGML_OP_REDUCE && inpL->src[model.default_layer_device[il]]) {
// if (model.layers[il].wqkv) {
// int idx = ggml_backend_sched_get_backend_idx(lctx.sched, model.layers[il].wqkv->buffer);
// printf("Layer %2d wqkv backend is %d\n", il, idx);
// //auto backend = ggml_backend_sched_get_tensor_backend(lctx.sched, model.layers[il].wqkv);
// //if (backend) printf("Layer %2d wqkv backend is %s\n", il, ggml_backend_name(backend));
// //else printf("Backend for wqkv in layer %2d is not known\n", il);
// }
// inpL->view_src = inpL->src[model.default_layer_device[il]];
// //printf("Using reduce result on device %d\n", model.default_layer_device[il]);
// //inpL = inpL->src[model.default_layer_device[il]];
//}
//auto norm = model.layers[il].attn_norm->extra ? ((ggml_split_tensor_t *)model.layers[il].attn_norm->extra)->splits[model.default_layer_device[il]] : 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);