Skip to content

Commit

Permalink
server : add more env vars, improve gen-docs (ggerganov#9635)
Browse files Browse the repository at this point in the history 
* server : add more env vars, improve gen-docs

* update server docs

* LLAMA_ARG_NO_CONTEXT_SHIFT
  • Loading branch information
@ngxson @dsx1986
ngxson authored and dsx1986 committed Oct 29, 2024
1 parent ff56563 commit 5a49d62
Show file tree
Hide file tree
Showing 4 changed files with 157 additions and 107 deletions.
56 changes: 28 additions & 28 deletions common/arg.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -691,7 +691,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params) {
params.ctx_shift = false;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONTEXT_SHIFT"));
add_opt(llama_arg(
{"--chunks"}, "N",
format("max number of chunks to process (default: %d, -1 = all)", params.n_chunks),
Expand Down Expand Up @@ -1102,7 +1102,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
else if (value == "last") { params.pooling_type = LLAMA_POOLING_TYPE_LAST; }
else { throw std::invalid_argument("invalid value"); }
}
).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
add_opt(llama_arg(
{"--attention"}, "{causal,non,causal}",
"attention type for embeddings, use model default if unspecified",
Expand All @@ -1121,77 +1121,77 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
else { throw std::invalid_argument("invalid value"); }
}
));
).set_env("LLAMA_ARG_ROPE_SCALING_TYPE"));
add_opt(llama_arg(
{"--rope-scale"}, "N",
"RoPE context scaling factor, expands context by a factor of N",
[](gpt_params & params, const std::string & value) {
params.rope_freq_scale = 1.0f / std::stof(value);
}
));
).set_env("LLAMA_ARG_ROPE_SCALE"));
add_opt(llama_arg(
{"--rope-freq-base"}, "N",
"RoPE base frequency, used by NTK-aware scaling (default: loaded from model)",
[](gpt_params & params, const std::string & value) {
params.rope_freq_base = std::stof(value);
}
));
).set_env("LLAMA_ARG_ROPE_FREQ_BASE"));
add_opt(llama_arg(
{"--rope-freq-scale"}, "N",
"RoPE frequency scaling factor, expands context by a factor of 1/N",
[](gpt_params & params, const std::string & value) {
params.rope_freq_scale = std::stof(value);
}
));
).set_env("LLAMA_ARG_ROPE_FREQ_SCALE"));
add_opt(llama_arg(
{"--yarn-orig-ctx"}, "N",
format("YaRN: original context size of model (default: %d = model training context size)", params.yarn_orig_ctx),
[](gpt_params & params, int value) {
params.yarn_orig_ctx = value;
}
));
).set_env("LLAMA_ARG_YARN_ORIG_CTX"));
add_opt(llama_arg(
{"--yarn-ext-factor"}, "N",
format("YaRN: extrapolation mix factor (default: %.1f, 0.0 = full interpolation)", (double)params.yarn_ext_factor),
[](gpt_params & params, const std::string & value) {
params.yarn_ext_factor = std::stof(value);
}
));
).set_env("LLAMA_ARG_YARN_EXT_FACTOR"));
add_opt(llama_arg(
{"--yarn-attn-factor"}, "N",
format("YaRN: scale sqrt(t) or attention magnitude (default: %.1f)", (double)params.yarn_attn_factor),
[](gpt_params & params, const std::string & value) {
params.yarn_attn_factor = std::stof(value);
}
));
).set_env("LLAMA_ARG_YARN_ATTN_FACTOR"));
add_opt(llama_arg(
{"--yarn-beta-slow"}, "N",
format("YaRN: high correction dim or alpha (default: %.1f)", (double)params.yarn_beta_slow),
[](gpt_params & params, const std::string & value) {
params.yarn_beta_slow = std::stof(value);
}
));
).set_env("LLAMA_ARG_YARN_BETA_SLOW"));
add_opt(llama_arg(
{"--yarn-beta-fast"}, "N",
format("YaRN: low correction dim or beta (default: %.1f)", (double)params.yarn_beta_fast),
[](gpt_params & params, const std::string & value) {
params.yarn_beta_fast = std::stof(value);
}
));
).set_env("LLAMA_ARG_YARN_BETA_FAST"));
add_opt(llama_arg(
{"-gan", "--grp-attn-n"}, "N",
format("group-attention factor (default: %d)", params.grp_attn_n),
[](gpt_params & params, int value) {
params.grp_attn_n = value;
}
));
).set_env("LLAMA_ARG_GRP_ATTN_N"));
add_opt(llama_arg(
{"-gaw", "--grp-attn-w"}, "N",
format("group-attention width (default: %.1f)", (double)params.grp_attn_w),
[](gpt_params & params, int value) {
params.grp_attn_w = value;
}
));
).set_env("LLAMA_ARG_GRP_ATTN_W"));
add_opt(llama_arg(
{"-dkvc", "--dump-kv-cache"},
"verbose print of the KV cache",
Expand All @@ -1205,23 +1205,23 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params) {
params.no_kv_offload = true;
}
));
).set_env("LLAMA_ARG_NO_KV_OFFLOAD"));
add_opt(llama_arg(
{"-ctk", "--cache-type-k"}, "TYPE",
format("KV cache data type for K (default: %s)", params.cache_type_k.c_str()),
[](gpt_params & params, const std::string & value) {
// TODO: get the type right here
params.cache_type_k = value;
}
));
).set_env("LLAMA_ARG_CACHE_TYPE_K"));
add_opt(llama_arg(
{"-ctv", "--cache-type-v"}, "TYPE",
format("KV cache data type for V (default: %s)", params.cache_type_v.c_str()),
[](gpt_params & params, const std::string & value) {
// TODO: get the type right here
params.cache_type_v = value;
}
));
).set_env("LLAMA_ARG_CACHE_TYPE_V"));
add_opt(llama_arg(
{"--perplexity", "--all-logits"},
format("return logits for all tokens in the batch (default: %s)", params.logits_all ? "true" : "false"),
Expand Down Expand Up @@ -1355,22 +1355,22 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params, const std::string & value) {
params.rpc_servers = value;
}
));
).set_env("LLAMA_ARG_RPC"));
#endif
add_opt(llama_arg(
{"--mlock"},
"force system to keep model in RAM rather than swapping or compressing",
[](gpt_params & params) {
params.use_mlock = true;
}
));
).set_env("LLAMA_ARG_MLOCK"));
add_opt(llama_arg(
{"--no-mmap"},
"do not memory-map model (slower load but may reduce pageouts if not using mlock)",
[](gpt_params & params) {
params.use_mmap = false;
}
));
).set_env("LLAMA_ARG_NO_MMAP"));
add_opt(llama_arg(
{"--numa"}, "TYPE",
"attempt optimizations that help on some NUMA systems\n"
Expand All @@ -1385,7 +1385,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
else { throw std::invalid_argument("invalid value"); }
}
));
).set_env("LLAMA_ARG_NUMA"));
add_opt(llama_arg(
{"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
"number of layers to store in VRAM",
Expand Down Expand Up @@ -1433,7 +1433,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the split mode has no effect.\n");
}
}
));
).set_env("LLAMA_ARG_SPLIT_MODE"));
add_opt(llama_arg(
{"-ts", "--tensor-split"}, "N0,N1,N2,...",
"fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1",
Expand All @@ -1460,7 +1460,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting a tensor split has no effect.\n");
}
}
));
).set_env("LLAMA_ARG_TENSOR_SPLIT"));
add_opt(llama_arg(
{"-mg", "--main-gpu"}, "INDEX",
format("the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: %d)", params.main_gpu),
Expand All @@ -1470,7 +1470,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
fprintf(stderr, "warning: llama.cpp was compiled without support for GPU offload. Setting the main GPU has no effect.\n");
}
}
));
).set_env("LLAMA_ARG_MAIN_GPU"));
add_opt(llama_arg(
{"--check-tensors"},
format("check model tensor data for invalid values (default: %s)", params.check_tensors ? "true" : "false"),
Expand Down Expand Up @@ -1533,7 +1533,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params, const std::string & value) {
params.model_alias = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_ALIAS"));
add_opt(llama_arg(
{"-m", "--model"}, "FNAME",
ex == LLAMA_EXAMPLE_EXPORT_LORA
Expand Down Expand Up @@ -1741,7 +1741,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params, const std::string & value) {
params.public_path = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_STATIC_PATH"));
add_opt(llama_arg(
{"--embedding", "--embeddings"},
format("restrict to only support embedding use case; use only with dedicated embedding models (default: %s)", params.embedding ? "enabled" : "disabled"),
Expand Down Expand Up @@ -1779,22 +1779,22 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](gpt_params & params, const std::string & value) {
params.ssl_file_key = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_KEY_FILE"));
add_opt(llama_arg(
{"--ssl-cert-file"}, "FNAME",
"path to file a PEM-encoded SSL certificate",
[](gpt_params & params, const std::string & value) {
params.ssl_file_cert = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
add_opt(llama_arg(
{"-to", "--timeout"}, "N",
format("server read/write timeout in seconds (default: %d)", params.timeout_read),
[](gpt_params & params, int value) {
params.timeout_read = value;
params.timeout_write = value;
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_TIMEOUT"));
add_opt(llama_arg(
{"--threads-http"}, "N",
format("number of threads used to process HTTP requests (default: %d)", params.n_threads_http),
Expand Down
85 changes: 58 additions & 27 deletions examples/gen-docs/gen-docs.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,42 +6,73 @@

// Export usage message (-h) to markdown format

static void write_table_header(std::ofstream & file) {
file << "| Argument | Explanation |\n";
file << "| -------- | ----------- |\n";
}

static void write_table_entry(std::ofstream & file, const llama_arg & opt) {
file << "| `";
// args
for (const auto & arg : opt.args) {
if (arg == opt.args.front()) {
file << arg;
if (opt.args.size() > 1) file << ", ";
} else {
file << arg << (arg != opt.args.back() ? ", " : "");
}
}
// value hint
if (opt.value_hint) {
std::string md_value_hint(opt.value_hint);
string_replace_all(md_value_hint, "|", "\\|");
file << " " << md_value_hint;
}
if (opt.value_hint_2) {
std::string md_value_hint_2(opt.value_hint_2);
string_replace_all(md_value_hint_2, "|", "\\|");
file << " " << md_value_hint_2;
}
// help text
std::string md_help(opt.help);
string_replace_all(md_help, "\n", "<br/>");
string_replace_all(md_help, "|", "\\|");
file << "` | " << md_help << " |\n";
}

static void write_table(std::ofstream & file, std::vector<llama_arg *> & opts) {
write_table_header(file);
for (const auto & opt : opts) {
write_table_entry(file, *opt);
}
}

static void export_md(std::string fname, llama_example ex) {
std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);

gpt_params params;
auto ctx_arg = gpt_params_parser_init(params, ex);

file << "| Argument | Explanation |\n";
file << "| -------- | ----------- |\n";
std::vector<llama_arg *> common_options;
std::vector<llama_arg *> sparam_options;
std::vector<llama_arg *> specific_options;
for (auto & opt : ctx_arg.options) {
file << "| `";
// args
for (const auto & arg : opt.args) {
if (arg == opt.args.front()) {
file << arg;
if (opt.args.size() > 1) file << ", ";
} else {
file << arg << (arg != opt.args.back() ? ", " : "");
}
}
// value hint
if (opt.value_hint) {
std::string md_value_hint(opt.value_hint);
string_replace_all(md_value_hint, "|", "\\|");
file << " " << md_value_hint;
// in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
if (opt.is_sparam) {
sparam_options.push_back(&opt);
} else if (opt.in_example(ctx_arg.ex)) {
specific_options.push_back(&opt);
} else {
common_options.push_back(&opt);
}
if (opt.value_hint_2) {
std::string md_value_hint_2(opt.value_hint_2);
string_replace_all(md_value_hint_2, "|", "\\|");
file << " " << md_value_hint_2;
}
// help text
std::string md_help(opt.help);
string_replace_all(md_help, "\n", "<br/>");
string_replace_all(md_help, "|", "\\|");
file << "` | " << md_help << " |\n";
}

file << "**Common params**\n\n";
write_table(file, common_options);
file << "\n\n**Sampling params**\n\n";
write_table(file, sparam_options);
file << "\n\n**Example-specific params**\n\n";
write_table(file, specific_options);
}

int main(int, char **) {
Expand Down
Loading

0 comments on commit 5a49d62

Please sign in to comment.