vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-05 21:29:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix compute buffer estimate: don't reverse CUDA VRAM for output layer

Browse source
This commit is contained in:
Zonghang Li 2025-06-27 12:42:16 +00:00
parent 3a03549fed
commit 11ce0d58f7
1 changed files with 14 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
src/llama.cpp
View file

@ -21359,6 +21359,16 @@ void * llama_context_setup_backend(
bool ok = true;
GGML_ASSERT(ctx->sched.size() == gf.size());
for (size_t i = 0; i < gf.size(); ++i) {
#if defined(GGML_USE_CUDA)
if ((cparams.rank == 0 && (i == 0 || i == gf.size() - 1))
|| model->n_gpu_layers == 0) {
continue;
}
#elif defined(GGML_USE_METAL)
#endif
ok = ok & ggml_backend_sched_reserve(ctx->sched[i], gf[i]);
}
if (!ok) {
@ -21963,7 +21973,8 @@ void llama_model_compute_buf_size(
(n_inp_pos + n_kq_mask) * type_size_f32 + n_v * type_size_f16 + nb_attn_norm_w,
});
}
*gpu_buf += (n_out_embd + n_result) * type_size_f32 + nb_output_w;
// we run the output layer on CPU by default
// *gpu_buf += (n_out_embd + n_result) * type_size_f32 + nb_output_w;
gpu_host_buf = (n_inp_toks + n_inp_embd + n_bak_embd + n_inp_pos + n_kq_mask + n_inp_out_ids + n_out_embd) * type_size_f32;
} else {
if (has_gpu_layers) {
@ -22036,7 +22047,7 @@ void llama_model_compute_buf_size(
LLAMA_LOG_INFO("\n");
LLAMA_LOG_INFO("%s: here the compute buffer size is a predicted upper bound, not an exact value\n", __func__);
LLAMA_LOG_INFO("%s: (rank %d) compute buffer size = %7.2f MiB (GPU) + %7.2f MiB (CPU / GPU-host buffer)\n", __func__,
LLAMA_LOG_INFO("%s: (rank %d) compute buffer size = %7.2f MiB (GPU) + %7.2f MiB (CPU & GPU-host buffer)\n", __func__,
my_rank, *gpu_buf / (1024.0 * 1024.0), *cpu_buf / (1024.0 * 1024.0));
if (backend == BACKEND_CUDA) {