Skip to content

llama: use FA + max. GPU layers by default #15434

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 8 commits into from
Aug 30, 2025
Merged

llama: use FA + max. GPU layers by default #15434

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
86f0cea
llama: use max. GPU layers by default, auto -fa
JohannesGaessler Aug 19, 2025
97ce75a
disable -fa for server test
JohannesGaessler Aug 27, 2025
9be3435
remove redundant defaults
JohannesGaessler Aug 29, 2025
8d77368
ggml-backend: abort instead of segfault
JohannesGaessler Aug 29, 2025
634c522
address review comments
JohannesGaessler Aug 30, 2025
a3aa838
add comment [no ci]
JohannesGaessler Aug 30, 2025
1de69b8
fix unittest, remove metal ifdef
JohannesGaessler Aug 30, 2025
3762f43
add comment [no ci]
JohannesGaessler Aug 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
30 changes: 12 additions & 18 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1545,10 +1545,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL}));
add_opt(common_arg(
{"-fa", "--flash-attn"},
string_format("enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled"),
[](common_params & params) {
params.flash_attn = true;
{"-fa", "--flash-attn"}, "FA",
string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')", llama_flash_attn_type_name(params.flash_attn_type)),
[](common_params & params, const std::string & value) {
if (value == "on" || value == "enabled") {
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
} else if (value == "off" || value == "disabled") {
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
} else if (value == "auto") {
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
} else {
throw std::runtime_error(string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
}
}
).set_env("LLAMA_ARG_FLASH_ATTN"));
add_opt(common_arg(
Expand Down Expand Up @@ -3459,8 +3467,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand All @@ -3475,8 +3481,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand All @@ -3491,8 +3495,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF";
params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand All @@ -3508,10 +3510,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf";
params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.speculative.n_gpu_layers = 99;
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand All @@ -3527,10 +3526,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf";
params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF";
params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf";
params.speculative.n_gpu_layers = 99;
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand All @@ -3545,8 +3541,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.model.hf_repo = "ggml-org/Qwen3-Coder-30B-A3B-Instruct-Q8_0-GGUF";
params.model.hf_file = "qwen3-coder-30b-a3b-instruct-q8_0.gguf";
params.port = 8012;
params.n_gpu_layers = 99;
params.flash_attn = true;
params.n_ubatch = 1024;
params.n_batch = 1024;
params.n_ctx = 0;
Expand Down
8 changes: 5 additions & 3 deletions common/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -901,7 +901,8 @@ struct common_init_result common_init_from_params(common_params & params) {

llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
if (model == NULL) {
LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str());
LOG_ERR("%s: failed to load model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
__func__, params.model.path.c_str());
return iparams;
}

Expand All @@ -911,7 +912,8 @@ struct common_init_result common_init_from_params(common_params & params) {

llama_context * lctx = llama_init_from_model(model, cparams);
if (lctx == NULL) {
LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str());
LOG_ERR("%s: failed to create context with model '%s', try reducing --n-gpu-layers if you're running out of VRAM\n",
__func__, params.model.path.c_str());
llama_model_free(model);
return iparams;
}
Expand Down Expand Up @@ -1152,10 +1154,10 @@ struct llama_context_params common_context_params_to_llama(const common_params &
cparams.yarn_orig_ctx = params.yarn_orig_ctx;
cparams.pooling_type = params.pooling_type;
cparams.attention_type = params.attention_type;
cparams.flash_attn_type = params.flash_attn_type;
cparams.cb_eval = params.cb_eval;
cparams.cb_eval_user_data = params.cb_eval_user_data;
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;
cparams.swa_full = params.swa_full;
Expand Down
2 changes: 1 addition & 1 deletion common/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -309,6 +309,7 @@ struct common_params {
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
enum llama_flash_attn_type flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO; // whether to use Flash Attention

struct common_params_sampling sampling;
struct common_params_speculative speculative;
Expand Down Expand Up @@ -372,7 +373,6 @@ struct common_params {
bool multiline_input = false; // reverse the usage of `\`
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = false; // flash attention
bool no_perf = false; // disable performance metrics
bool ctx_shift = false; // context shift on infinite text generation
bool swa_full = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
Expand Down
2 changes: 1 addition & 1 deletion examples/diffusion/diffusion-cli.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -564,7 +564,7 @@ int main(int argc, char ** argv) {
ctx_params.n_ctx = params.n_ctx;
ctx_params.n_batch = params.n_batch;
ctx_params.n_ubatch = params.n_ubatch;
ctx_params.flash_attn = params.flash_attn;
ctx_params.flash_attn_type = params.flash_attn_type;
ctx_params.no_perf = params.no_perf;
ctx_params.type_k = params.cache_type_k;
ctx_params.type_v = params.cache_type_v;
Expand Down
Loading