tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-30 04:45:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add --no-op-offload to improve -ot pp perf in MoE models like llama4 400B (#13386)

Browse Source
This commit is contained in:
David Huang
2025-05-11 20:18:39 +08:00
committed by GitHub
parent 3eac209319
commit 7f323a589f
11 changed files with 57 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/llama-context.cpp
View File

@ -93,6 +93,7 @@ llama_context::llama_context(
}
cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
cparams.op_offload = params.op_offload;
const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
@ -243,7 +244,7 @@ llama_context::llama_context(
}
}
sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));
sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel, cparams.op_offload));
if (pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
@ -1871,6 +1872,7 @@ llama_context_params llama_context_default_params() {
/*.offload_kqv =*/ true,
/*.flash_attn =*/ false,
/*.no_perf =*/ true,
/*.op_offload =*/ true,
};
return result;