Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.dockerignore
#	.github/workflows/build.yml
#	.github/workflows/docker.yml
#	Makefile
#	README.md
#	examples/infill/infill.cpp
#	examples/perplexity/perplexity.cpp
#	examples/server/README.md
#	examples/speculative/speculative.cpp
#	flake.lock
#	ggml/src/CMakeLists.txt
#	scripts/sync-ggml.last
#	tests/test-backend-ops.cpp
#	tests/test-sampling.cpp
This commit is contained in:
Concedo
2024-09-27 11:21:28 +08:00
parent 6342b414ea 95bc82fbc0
commit ea55f69dc1
39 changed files with 2587 additions and 1564 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/arg.cpp
+28 -28
View File
@@ -692,7 +692,7 @@ 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}));
).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),
@@ -1103,7 +1103,7 @@ 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",
@@ -1122,77 +1122,77 @@ 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",
@@ -1206,7 +1206,7 @@ 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()),
@@ -1214,7 +1214,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
// 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()),
@@ -1222,7 +1222,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
// 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"),
@@ -1356,7 +1356,7 @@ 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"},
@@ -1364,14 +1364,14 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
[](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"
@@ -1386,7 +1386,7 @@ 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",
@@ -1434,7 +1434,7 @@ 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",
@@ -1461,7 +1461,7 @@ 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),
@@ -1471,7 +1471,7 @@ 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"),
@@ -1534,7 +1534,7 @@ 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
@@ -1742,7 +1742,7 @@ 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"),
@@ -1780,14 +1780,14 @@ 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),
@@ -1795,7 +1795,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
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),
common/log.cpp
+1 -1
View File
@@ -82,7 +82,7 @@ struct gpt_log_entry {
}
}
if (level != GGML_LOG_LEVEL_NONE && prefix) {
if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
if (timestamp) {
// [M.s.ms.us]
fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
common/log.h
+2
View File
@@ -83,8 +83,10 @@ void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps); // w
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, 0, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT, 0, __VA_ARGS__)
#define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO, verbosity, __VA_ARGS__)
#define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN, verbosity, __VA_ARGS__)
#define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
#define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
#define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT, verbosity, __VA_ARGS__)
common/sampling.cpp
+9 -1
View File
@@ -209,7 +209,15 @@ struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const st
GGML_ASSERT(false && "unknown mirostat version");
}
} else {
llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
if (params.n_probs > 0) {
// some use cases require to sample greedily, but still obtain the probabilities of the top tokens
// ref: https://github.com/ggerganov/llama.cpp/pull/9605
//
// the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
// it is much faster, since we avoid sorting all tokens and should give a good approximation
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
}