tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-26 19:55:04 +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

7
common/arg.cpp
View File

@ -2437,6 +2437,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
}
));
add_opt(common_arg(
{"--no-op-offload"},
string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
[](common_params & params) {
params.no_op_offload = true;
}
));
add_opt(common_arg(
{"--lora"}, "FNAME",
"path to LoRA adapter (can be repeated to use multiple adapters)",

1
common/common.cpp
View File

@ -1113,6 +1113,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
cparams.offload_kqv = !params.no_kv_offload;
cparams.flash_attn = params.flash_attn;
cparams.no_perf = params.no_perf;
cparams.op_offload = !params.no_op_offload;
if (params.reranking) {
cparams.embeddings = true;

1
common/common.h
View File

@ -332,6 +332,7 @@ struct common_params {
bool no_kv_offload = false; // disable KV offloading
bool warmup = true; // warmup run
bool check_tensors = false; // validate tensor data
bool no_op_offload = false; // globally disable offload host tensor operations to device
bool single_turn = false; // single turn chat conversation