CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

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

Browse Source 
* llama: use max. GPU layers by default, auto -fa

* ggml-backend: abort instead of segfault
This commit is contained in:
Johannes Gäßler
2025-08-30 16:32:10 +02:00
committed by GitHub
parent 38ad381f9f
commit e81b8e4b7f
19 changed files with 235 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
common/arg.cpp
View File

@@ -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(
@@ -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;
@@ -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;
@@ -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;
@@ -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;
@@ -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;
@@ -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;