vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into compilade/imatrix-batched-chunks

Browse source
This commit is contained in:
Francis Couture-Harpin 2025-02-09 12:06:15 -05:00
commit db502ddd0e
762 changed files with 149462 additions and 91773 deletions
Download patch file Download diff file Expand all files Collapse all files

48
examples/CMakeLists.txt
View file

@ -6,22 +6,26 @@ find_package(Threads REQUIRED)
# ...
# flags
llama_add_compile_flags()
# examples
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN)
else()
add_subdirectory(cvector-generator)
add_subdirectory(baby-llama)
add_subdirectory(batched-bench)
add_subdirectory(batched)
add_subdirectory(benchmark)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(embedding)
add_subdirectory(eval-callback)
add_subdirectory(export-lora)
add_subdirectory(gbnf-validator)
if (NOT WIN32)
# disabled on Windows because it uses internal functions not exported with LLAMA_API
add_subdirectory(gbnf-validator)
endif()
add_subdirectory(gguf-hash)
add_subdirectory(gguf-split)
add_subdirectory(gguf)
@ -29,27 +33,41 @@ else()
add_subdirectory(imatrix)
add_subdirectory(infill)
add_subdirectory(llama-bench)
add_subdirectory(llava)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(main)
add_subdirectory(parallel)
add_subdirectory(passkey)
add_subdirectory(perplexity)
add_subdirectory(quantize-stats)
add_subdirectory(quantize)
add_subdirectory(retrieval)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (LLAMA_BUILD_SERVER)
add_subdirectory(server)
endif()
if (GGML_SYCL)
add_subdirectory(sycl)
add_subdirectory(server)
endif()
add_subdirectory(save-load-state)
add_subdirectory(run)
add_subdirectory(simple)
add_subdirectory(simple-chat)
add_subdirectory(speculative)
add_subdirectory(speculative-simple)
add_subdirectory(tokenize)
add_subdirectory(tts)
add_subdirectory(gen-docs)
if (NOT GGML_BACKEND_DL)
# these examples use the backends directly and cannot be built with dynamic loading
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(cvector-generator)
add_subdirectory(export-lora)
if (NOT WIN32)
# disabled on Windows because it uses internal functions not exported with LLAMA_API
add_subdirectory(quantize-stats)
endif()
add_subdirectory(llava)
if (GGML_RPC)
add_subdirectory(rpc)
endif()
if (GGML_SYCL)
add_subdirectory(sycl)
endif()
endif()
endif()

5
examples/baby-llama/CMakeLists.txt
View file

@ -1,5 +0,0 @@
set(TARGET llama-baby-llama)
add_executable(${TARGET} baby-llama.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

1639
examples/baby-llama/baby-llama.cpp
View file

File diff suppressed because it is too large Load diff

61
examples/base-translate.sh
View file

@ -1,61 +0,0 @@
#!/bin/bash
#
# Few-shot translation example.
# Requires a base model (i.e. no fine-tuned or instruct models).
#
# Usage:
#
# cd llama.cpp
# make -j
#
# ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
#
if [ $# -lt 2 ]; then
echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
exit 1
fi
eargs=""
if [ $# -gt 2 ]; then
eargs="${@:3}"
fi
ftmp="__llama.cpp_example_tmp__.txt"
trap "rm -f $ftmp" EXIT
echo "Translate from English to French:
===
sea otter, peppermint, plush girafe:
sea otter => loutre de mer
peppermint => menthe poivrée
plush girafe => girafe peluche
===
violin
violin => violon
===
phone, computer, mouse, keyboard:
phone => téléphone
computer => ordinateur
mouse => souris
keyboard => clavier
===
" > $ftmp
echo "$2
" >> $ftmp
model=$1
# generate the most likely continuation until the string "===" is found
./llama-cli -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

2
examples/batched-bench/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-batched-bench)
add_executable(${TARGET} batched-bench.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

86
examples/batched-bench/batched-bench.cpp
View file

@ -1,47 +1,28 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
// mutates the input string
static std::vector<int> parse_list(char * p) {
std::vector<int> ret;
char * q = p;
while (*p) {
if (*p == ',') {
*p = '\0';
ret.push_back(std::atoi(q));
q = p + 1;
}
++p;
}
ret.push_back(std::atoi(q));
return ret;
}
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_BENCH, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
return 1;
}
common_init();
int is_pp_shared = params.is_pp_shared;
std::vector<int> n_pp = params.n_pp;
@ -55,21 +36,21 @@ int main(int argc, char ** argv) {
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
return 1;
}
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
llama_context_params ctx_params = common_context_params_to_llama(params);
// ensure enough sequences are available
ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
@ -93,12 +74,11 @@ int main(int argc, char ** argv) {
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
@ -111,21 +91,21 @@ int main(int argc, char ** argv) {
// warm up
{
for (int i = 0; i < 16; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
common_batch_add(batch, 0, i, { 0 }, false);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
if (!params.batched_bench_output_jsonl) {
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
LOG("\n");
LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG("\n");
LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
}
for ( int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
@ -141,11 +121,11 @@ int main(int argc, char ** argv) {
continue;
}
llama_batch_clear(batch);
common_batch_clear(batch);
for (int i = 0; i < pp; ++i) {
for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
llama_batch_add(batch, 0, i, { j }, false);
common_batch_add(batch, 0, i, { j }, false);
}
}
batch.logits[batch.n_tokens - 1] = true;
@ -155,7 +135,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_clear(ctx);
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -170,14 +150,14 @@ int main(int argc, char ** argv) {
const auto t_tg_start = ggml_time_us();
for (int i = 0; i < tg; ++i) {
llama_batch_clear(batch);
common_batch_clear(batch);
for (int j = 0; j < pl; ++j) {
llama_batch_add(batch, 0, pp + i, { j }, true);
common_batch_add(batch, 0, pp + i, { j }, true);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
@ -195,30 +175,30 @@ int main(int argc, char ** argv) {
const float speed = n_kv / t;
if(params.batched_bench_output_jsonl) {
LOG_TEE(
LOG(
"{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
"\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
);
} else {
LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
LOG("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
}
}
}
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

12
examples/batched.swift/Sources/main.swift
View file

@ -23,12 +23,12 @@ defer {
}
let model_params = llama_model_default_params()
guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else {
guard let model = llama_model_load_from_file(modelPath.cString(using: .utf8), model_params) else {
print("Failed to load model")
exit(1)
}
defer {
llama_free_model(model)
llama_model_free(model)
}
var tokens = tokenize(text: prompt, add_bos: true)
@ -140,10 +140,8 @@ while n_cur <= n_len {
let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
llama_sampler_accept(smpl, new_token_id)
// is it an end of stream? -> mark the stream as finished
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
i_batch[i] = -1
// print("")
if n_parallel > 1 {
@ -202,8 +200,8 @@ let t_main_end = ggml_time_us()
print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
llama_perf_print(UnsafeRawPointer(context), LLAMA_PERF_TYPE_CONTEXT)
llama_perf_print(UnsafeRawPointer(smpl), LLAMA_PERF_TYPE_SAMPLER_CHAIN)
llama_perf_sampler_print(smpl)
llama_perf_context_print(context)
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let utf8Count = text.utf8.count

2
examples/batched/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-batched)
add_executable(${TARGET} batched.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

102
examples/batched/batched.cpp
View file

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
@ -7,22 +9,22 @@
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
params.prompt = "Hello my name is";
params.n_predict = 32;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
return 1;
}
common_init();
// number of parallel batches
int n_parallel = params.n_parallel;
@ -37,66 +39,67 @@ int main(int argc, char ** argv) {
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: error: unable to load model\n" , __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(model, params.prompt, true);
tokens_list = common_tokenize(vocab, params.prompt, true);
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
// initialize the context
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
llama_context_params ctx_params = common_context_params_to_llama(params);
ctx_params.n_ctx = n_kv_req;
ctx_params.n_batch = std::max(n_predict, n_parallel);
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
auto sparams = llama_sampler_chain_default_params();
sparams.no_perf = false;
llama_sampler * smpl = llama_sampler_chain_init(sparams);
llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sampling.top_k));
llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sampling.top_p, params.sampling.min_keep));
llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sampling.temp));
llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sampling.seed));
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
const int n_ctx = llama_n_ctx(ctx);
LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__, n_kv_req);
LOG_TEE("%s: either reduce n_parallel or increase n_ctx\n", __func__);
LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__, n_kv_req);
LOG_ERR("%s: either reduce n_parallel or increase n_ctx\n", __func__);
return 1;
}
// print the prompt token-by-token
fprintf(stderr, "\n");
LOG("\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", common_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
// create a llama_batch
// we use this object to submit token data for decoding
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
@ -108,30 +111,30 @@ int main(int argc, char ** argv) {
// evaluate the initial prompt
for (size_t i = 0; i < tokens_list.size(); ++i) {
llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
common_batch_add(batch, tokens_list[i], i, seq_ids, false);
}
GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
if (llama_model_has_encoder(model)) {
if (llama_encode(ctx, batch)) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
decoder_start_token_id = llama_token_bos(model);
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = llama_vocab_bos(vocab);
}
llama_batch_clear(batch);
llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
common_batch_clear(batch);
common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
}
// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -142,7 +145,7 @@ int main(int argc, char ** argv) {
//}
if (n_parallel > 1) {
LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
}
// main loop
@ -161,7 +164,7 @@ int main(int argc, char ** argv) {
while (n_cur <= n_predict) {
// prepare the next batch
llama_batch_clear(batch);
common_batch_clear(batch);
// sample the next token for each parallel sequence / stream
for (int32_t i = 0; i < n_parallel; ++i) {
@ -172,14 +175,12 @@ int main(int argc, char ** argv) {
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
llama_sampler_accept(smpl, new_token_id);
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_predict) {
i_batch[i] = -1;
LOG_TEE("\n");
LOG("\n");
if (n_parallel > 1) {
LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
}
continue;
@ -187,16 +188,15 @@ int main(int argc, char ** argv) {
// if there is only one stream, we print immediately to stdout
if (n_parallel == 1) {
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
}
streams[i] += llama_token_to_piece(ctx, new_token_id);
streams[i] += common_token_to_piece(ctx, new_token_id);
i_batch[i] = batch.n_tokens;
// push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_cur, { i }, true);
common_batch_add(batch, new_token_id, n_cur, { i }, true);
n_decode += 1;
}
@ -210,29 +210,27 @@ int main(int argc, char ** argv) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
if (n_parallel > 1) {
LOG_TEE("\n");
LOG("\n");
for (int32_t i = 0; i < n_parallel; ++i) {
LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
}
}
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
LOG_TEE("\n");
llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_sampler_print(smpl);
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
@ -240,7 +238,7 @@ int main(int argc, char ** argv) {
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();

6
examples/benchmark/CMakeLists.txt
View file

@ -1,6 +0,0 @@
set(TARGET llama-bench-matmult)
add_executable(${TARGET} benchmark-matmult.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
target_include_directories(${TARGET} PRIVATE ../../common)
target_compile_features(${TARGET} PRIVATE cxx_std_11)

275
examples/benchmark/benchmark-matmult.cpp
View file

@ -1,275 +0,0 @@
#include "common.h"
#include "ggml.h"
#include <locale.h>
#include <assert.h>
#include <math.h>
#include <cstring>
#include <cstdio>
#include <cinttypes>
#include <unordered_map>
#include <queue>
#include <string.h>
#include <cassert>
#include <fstream>
#include <string>
#include <iterator>
#include <algorithm>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
if (plan.work_size > 0) {
buf.resize(plan.work_size);
plan.work_data = buf.data();
}
ggml_graph_compute(graph, &plan);
}
static float tensor_sum_elements(const ggml_tensor * tensor) {
double sum = 0;
if (tensor->type == GGML_TYPE_F32) {
for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) {
sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
}
}
}
return sum;
}
static void tensor_dump(const ggml_tensor * tensor, const char * name) {
printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
tensor->type, ggml_type_name(tensor->type),
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
float sum = tensor_sum_elements(tensor);
printf("Sum of tensor %s is %6.2f\n", name, sum);
}
#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
struct benchmark_params_struct {
int n_threads = 1;
int32_t n_iterations = 10;
};
static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -i N, --iter N number of iterations to use during computation (default: %d)\n", params.n_iterations);
fprintf(stderr, "\n");
}
int main(int argc, char ** argv) {
struct benchmark_params_struct benchmark_params;
bool invalid_param = false;
std::string arg;
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg == "-t" || arg == "--threads") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_threads = std::stoi(argv[i]);
} else if (arg == "-i" || arg == "--iter") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_iterations = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") {
print_usage(argc, argv, benchmark_params);
exit(0);
}
}
if (invalid_param) {
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
print_usage(argc, argv, benchmark_params);
exit(1);
}
print_build_info();
printf("Starting Test\n");
// create the ggml context
struct ggml_context * ctx;
//const int sizex = 4096;
//const int sizey = 11008;
#undef VERBOSE_DEBUGGING
#ifndef VERBOSE_DEBUGGING
const int sizey = 4096;
const int sizex = 11008;
const int sizez = 128;
#else
/* Working - let's increase size */
const int sizey = 1;
const int sizex = (8*32);
const int sizez = 1;
/*const int sizey = 1;
const int sizex = 3*(8*32);
const int sizez = 1;*/
#endif
//printf("Memsize required = %i\n", sizex*sizex);
// TODO: perform the bench for all types or for a user specified type
const ggml_type qtype = GGML_TYPE_Q4_1;
size_t ctx_size = 0;
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
ctx_size += ggml_row_size(qtype, sizex*sizey);
ctx_size += ggml_row_size(qtype, sizex*sizey);
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
ctx_size += 1024*1024*16;
printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
struct ggml_init_params params = {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/* no_alloc =*/ 0
};
ctx = ggml_init(params);
if (!ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return 1;
}
printf("Creating new tensors\n");
// printf("Creating new tensor m1\n");
struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m11, 1.0f);
// printf("Creating new tensor m1\n");
struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m12, 1.5f);
// printf("Creating new tensor m2\n");
struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
ggml_set_f32(m2, 2.0f);
printf("\n------ Test 1 - Matrix Mult via F32 code\n");
// printf("Creating new tensor m11xm2\n");
struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
// printf("Creating compute graph\n");
struct ggml_cgraph * gf = ggml_new_graph(ctx);
ggml_build_forward_expand(gf, m11xm2);
printf("n_threads=%i\n", benchmark_params.n_threads);
TENSOR_DUMP(m11);
TENSOR_DUMP(m2);
std::vector<uint8_t> work_buffer;
ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
TENSOR_DUMP(gf->nodes[0]);
printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
int32_t nelements = sizex*sizey;
// Set up a the benchmark matrices
// printf("Creating new tensor q11 & Running quantize\n");
struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
// Set up a the compute graph
// printf("Creating new tensor q31\n");
struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
// printf("Creating compute graph\n");
struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
ggml_build_forward_expand(gf31, q31);
// Set up a second graph computation to make sure we override the CPU cache lines
// printf("Creating new tensor q12 & Running quantize\n");
struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
// printf("Creating new tensor q32\n");
struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
//printf("Creating compute graph\n");
struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
ggml_build_forward_expand(gf32, q32);
printf("n_threads=%i\n", benchmark_params.n_threads);
const int dimx = sizex;
const int dimy = sizey;
const int dimz = sizez;
long long int flops_per_dot_product = dimy + dimy;
long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
// Let's use the F32 result from above as a reference for the quantized multiplication
float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
printf("=====================================================================================\n");
double gflops_sum = 0;
for (int i=0;i<benchmark_params.n_iterations ;i++) {
long long int start = ggml_time_us();
//printf("Running ggml_graph_compute\n");
ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
long long int stop = ggml_time_us();
long long int usec = stop-start;
double gflops = (double)(flops_per_matrix)/usec/1000.0;
gflops_sum += gflops;
printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
i,
benchmark_params.n_threads,
sizex, sizey, sizez, flops_per_matrix,
usec,gflops);
#ifdef VERBOSE_DEBUGGING
TENSOR_DUMP("res",gf31.nodes[0])
#endif
// Check that the matrix multiplication result is in the right ballpark
// We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; // Let's accept an epsilon of 10^-6
if (delta > allowed_delta) {
printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
sum_of_F32_reference,
sum_of_Q4_result,
delta,
allowed_delta
);
exit(0);
}
// Running a different graph computation to make sure we override the CPU cache lines
ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
}
printf("\n");
printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
printf("=====================================================================================\n");
}

12
examples/chat-persistent.sh
View file

@ -23,8 +23,9 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
'|'\
'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
CTX_SIZE=2048
@ -129,15 +130,12 @@ while read -e line; do
printf ' '
# HACK get num tokens from debug message
# TODO get both messages in one go
if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
exit 1
fi
n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
if ((n_tokens > CTX_ROTATE_POINT)); then
tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"

2
examples/convert-llama2c-to-ggml/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-convert-llama2c-to-ggml)
add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

5
examples/convert-llama2c-to-ggml/README.md
View file

@ -2,11 +2,8 @@
This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository:
To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository.
`$ make -j`
After successful compilation, following usage options are available:
```
usage: ./llama-convert-llama2c-to-ggml [options]

123
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -1,4 +1,6 @@
#include "ggml.h"
#include "gguf.h"
#include "llama.h"
#include "common.h"
#include "log.h"
@ -9,6 +11,7 @@
#include <climits>
#include <cstring>
#include <cstdarg>
#include <cinttypes>
#include <ctime>
#include <random>
#include <stdexcept>
@ -105,43 +108,43 @@ static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_
const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
try {
w->token_embedding_table.resize(p->vocab_size * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
w->rms_att_weight.resize(p->n_layers * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
w->rms_ffn_weight.resize(p->n_layers * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
w->wq.resize(p->n_layers * p->dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
w->wo.resize(p->n_layers * p->dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
w->rms_final_weight.resize(p->dim);
LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
LOG_INF("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
if (shared_weights) {
w->wcls = {};
} else {
w->wcls.resize(p->vocab_size * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
}
}
catch (std::length_error &) {
@ -173,7 +176,7 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
fseek(f, 0, SEEK_END);
auto end = ftell(f);
if (curr != end) {
LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", __func__, curr, end);
LOG_ERR("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", __func__, curr, end);
return 1;
}
@ -181,26 +184,26 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
}
static void print_sample_weights(TransformerWeights *w){
LOG("----- Quick print of first of the weight vales of all the variables\n");
LOG("%f\n", w->token_embedding_table[0]);
LOG("%f\n", w->rms_att_weight[0]);
LOG("%f\n", w->rms_ffn_weight[0]);
LOG_INF("----- Quick print of first of the weight vales of all the variables\n");
LOG_INF("%f\n", w->token_embedding_table[0]);
LOG_INF("%f\n", w->rms_att_weight[0]);
LOG_INF("%f\n", w->rms_ffn_weight[0]);
LOG("%f\n", w->wq[0]);
LOG("%f\n", w->wk[0]);
LOG("%f\n", w->wv[0]);
LOG("%f\n", w->wo[0]);
LOG("%f\n", w->w1[0]);
LOG("%f\n", w->w2[0]);
LOG("%f\n", w->w3[0]);
LOG("%f\n", w->rms_att_weight[0]);
if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
LOG_INF("%f\n", w->wq[0]);
LOG_INF("%f\n", w->wk[0]);
LOG_INF("%f\n", w->wv[0]);
LOG_INF("%f\n", w->wo[0]);
LOG_INF("%f\n", w->w1[0]);
LOG_INF("%f\n", w->w2[0]);
LOG_INF("%f\n", w->w3[0]);
LOG_INF("%f\n", w->rms_att_weight[0]);
if (!w->wcls.empty()) LOG_INF("%f\n", w->wcls[0]);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////
//////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
struct llama_vocab {
struct my_llama_vocab {
using id = int32_t;
using token = std::string;
using ttype = llama_token_type;
@ -318,20 +321,20 @@ struct train_params {
};
static void print_params(struct my_llama_hparams * params) {
LOG("%s: n_vocab: %u\n", __func__, params->n_vocab);
LOG("%s: n_ctx: %u\n", __func__, params->n_ctx);
LOG("%s: n_embd: %u\n", __func__, params->n_embd);
LOG("%s: n_mult: %u\n", __func__, params->n_mult);
LOG("%s: n_head: %u\n", __func__, params->n_head);
LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
LOG("%s: n_ff: %u\n", __func__, params->n_ff);
LOG("%s: n_layer: %u\n", __func__, params->n_layer);
LOG("%s: n_rot: %u\n", __func__, params->n_rot);
LOG_INF("%s: n_vocab: %u\n", __func__, params->n_vocab);
LOG_INF("%s: n_ctx: %u\n", __func__, params->n_ctx);
LOG_INF("%s: n_embd: %u\n", __func__, params->n_embd);
LOG_INF("%s: n_mult: %u\n", __func__, params->n_mult);
LOG_INF("%s: n_head: %u\n", __func__, params->n_head);
LOG_INF("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
LOG_INF("%s: n_ff: %u\n", __func__, params->n_ff);
LOG_INF("%s: n_layer: %u\n", __func__, params->n_layer);
LOG_INF("%s: n_rot: %u\n", __func__, params->n_rot);
}
static void print_tensor_info(const struct ggml_context * ctx) {
for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
LOG("%s: Allocating ", __func__);
LOG_INF("%s: Allocating ", __func__);
int64_t total = 1;
int i = 0;
for (; i < ggml_n_dims(t); ++i) {
@ -433,12 +436,12 @@ static void print_matrix(struct ggml_tensor * probs) {
}
}
struct llama_file {
struct my_llama_file {
// use FILE * so we don't have to re-open the file to mmap
FILE * fp;
size_t size;
llama_file(const char * fname, const char * mode) {
my_llama_file(const char * fname, const char * mode) {
fp = std::fopen(fname, mode);
if (fp == NULL) {
size = 0;
@ -499,7 +502,7 @@ struct llama_file {
return std::string(chars.data(), len);
}
~llama_file() {
~my_llama_file() {
if (fp) {
std::fclose(fp);
}
@ -507,7 +510,7 @@ struct llama_file {
};
static bool is_ggml_file(const char * filename) {
llama_file file(filename, "rb");
my_llama_file file(filename, "rb");
if (file.size < 4) {
return false;
}
@ -524,9 +527,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
return out.str();
}
static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
if (is_ggml_file(filename)) {
LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
struct ggml_context * ctx_data = NULL;
struct gguf_init_params params = {
@ -574,21 +577,21 @@ static void load_vocab(const char * filename, const Config * config, struct llam
gguf_free(ctx);
} else {
// assume llama2.c vocabulary
LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
llama_file file(filename, "rb");
LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
my_llama_file file(filename, "rb");
if (!file.fp) {
die_fmt("%s: %s", strerror(errno), filename);
}
const int n_vocab = config->vocab_size;
/* uint32_t max_token_length = */ file.read_u32(); // unused
vocab->id_to_token.resize(n_vocab);
for (llama_vocab::id id=0; id<n_vocab; ++id) {
for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
float_t score = file.read_f32();
uint32_t len = file.read_u32();
std::string text = file.read_string(len);
unsigned char byte_val;
llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
if (id == UNKNOWN_TOKEN_ID) {
text = "<unk>";
type = LLAMA_TOKEN_TYPE_UNKNOWN;
@ -630,7 +633,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
}
static void save_as_llama_model(
struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
struct my_llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
) {
// convert AK weights into GG weights one by one.
// w->token_embedding_table -> model->tok_embeddings
@ -670,7 +673,7 @@ static void save_as_llama_model(
std::vector<const char*> tokens;
std::vector<float> scores;
std::vector<llama_token_type> token_types;
for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
tokens.push_back(token_data.text.c_str());
scores.push_back(token_data.score);
token_types.push_back(token_data.type);
@ -688,8 +691,8 @@ static void save_as_llama_model(
gguf_set_val_u32(ctx, KV_TOKENIZER_UNK_ID, UNKNOWN_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_BOS_ID, BOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_EOS_ID, EOS_TOKEN_ID);
gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, -1);
gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, -1);
gguf_set_val_u32(ctx, KV_TOKENIZER_SEP_ID, LLAMA_TOKEN_NULL);
gguf_set_val_u32(ctx, KV_TOKENIZER_PAD_ID, LLAMA_TOKEN_NULL);
gguf_set_val_u32(ctx, KV_CONTEXT_LENGTH, model->hparams.n_ctx);
gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
@ -871,23 +874,25 @@ static std::string basename(const std::string &path) {
}
int main(int argc, char ** argv) {
common_init();
struct train_params params = get_default_train_params();
if (!params_parse(argc, argv, &params)) {
return 1;
}
log_set_target(stdout);
Config config;
TransformerWeights weights = {};
{
LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
LOG_INF("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
FILE * file = fopen(params.fn_llama2c_model, "rb");
if (!file) {
LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
LOG_ERR("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
return 1;
}
// read in the config header
if (fread(&config, sizeof(Config), 1, file) != 1) {
LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
LOG_ERR("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
return 1;
}
auto shared_weights = config.vocab_size > 0;
@ -896,17 +901,17 @@ int main(int argc, char ** argv) {
// read in the Transformer weights
alloc_weights(&weights, &config, shared_weights);
if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
LOG_ERR("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
return 1;
}
fclose(file);
}
struct llama_vocab vocab;
struct my_llama_vocab vocab;
load_vocab(params.fn_vocab_model, &config, &vocab);
struct my_llama_model model;
model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
model.hparams.n_vocab = config.vocab_size; //llama_vocab_n_vocab(lctx);
model.hparams.n_ctx = params.n_ctx;
model.hparams.n_embd = config.dim; //params.n_embd;
model.hparams.n_ff = config.hidden_dim;
@ -929,7 +934,7 @@ int main(int argc, char ** argv) {
model.name = basename(params.fn_llama2c_model);
save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
LOG_INF("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
ggml_free(model.ctx);
return 0;

26
examples/convert_legacy_llama.py
View file

@ -840,6 +840,8 @@ class OutputFile:
self.gguf.add_base_model_version(key, base_model_entry["version"])
if "organization" in base_model_entry:
self.gguf.add_base_model_organization(key, base_model_entry["organization"])
if "description" in base_model_entry:
self.gguf.add_base_model_description(key, base_model_entry["description"])
if "url" in base_model_entry:
self.gguf.add_base_model_url(key, base_model_entry["url"])
if "doi" in base_model_entry:
@ -849,12 +851,32 @@ class OutputFile:
if "repo_url" in base_model_entry:
self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
if metadata.datasets is not None:
self.gguf.add_dataset_count(len(metadata.datasets))
for key, dataset_entry in enumerate(metadata.datasets):
if "name" in dataset_entry:
self.gguf.add_dataset_name(key, dataset_entry["name"])
if "author" in dataset_entry:
self.gguf.add_dataset_author(key, dataset_entry["author"])
if "version" in dataset_entry:
self.gguf.add_dataset_version(key, dataset_entry["version"])
if "organization" in dataset_entry:
self.gguf.add_dataset_organization(key, dataset_entry["organization"])
if "description" in dataset_entry:
self.gguf.add_dataset_description(key, dataset_entry["description"])
if "url" in dataset_entry:
self.gguf.add_dataset_url(key, dataset_entry["url"])
if "doi" in dataset_entry:
self.gguf.add_dataset_doi(key, dataset_entry["doi"])
if "uuid" in dataset_entry:
self.gguf.add_dataset_uuid(key, dataset_entry["uuid"])
if "repo_url" in dataset_entry:
self.gguf.add_dataset_repo_url(key, dataset_entry["repo_url"])
if metadata.tags is not None:
self.gguf.add_tags(metadata.tags)
if metadata.languages is not None:
self.gguf.add_languages(metadata.languages)
if metadata.datasets is not None:
self.gguf.add_datasets(metadata.datasets)
def add_meta_arch(self, params: Params) -> None:
# Metadata About The Neural Architecture Itself

2
examples/cvector-generator/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-cvector-generator)
add_executable(${TARGET} cvector-generator.cpp pca.hpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

48
examples/cvector-generator/cvector-generator.cpp
View file

@ -1,6 +1,9 @@
#include "ggml.h"
#include "gguf.h"
#include "arg.h"
#include "common.h"
#include "llama.h"
#include "ggml.h"
#include "pca.hpp"
#include "mean.hpp"
@ -12,14 +15,15 @@
#include "ggml-metal.h"
#endif
#include <algorithm>
#include <climits>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <iostream>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#include <climits>
//////////////////////////////////////////////////
@ -29,7 +33,7 @@ template <class Iter>
static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
std::string ret;
for (; begin != end; ++begin) {
ret += llama_token_to_piece(ctx, *begin);
ret += common_token_to_piece(ctx, *begin);
}
return ret;
@ -269,9 +273,11 @@ struct tokenized_prompt {
size_t max_seq_len;
tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
tokens_pos = common_tokenize(ctx, pos, add_bos, true);
tokens_neg = common_tokenize(ctx, neg, add_bos, true);
max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
padding_seq(ctx, tokens_pos, max_seq_len);
padding_seq(ctx, tokens_neg, max_seq_len);
@ -279,7 +285,7 @@ struct tokenized_prompt {
void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
// TODO: customize padding token
std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
llama_token pad_tok = pad_tokens.back();
while (tokens.size() < len) {
tokens.push_back(pad_tok);
@ -337,7 +343,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
llama_kv_cache_clear(ctx);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
@ -368,7 +374,7 @@ static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const
* Load prompt files and completion file.
* Then format each pair of prompt + completion to make an entry.
*/
static int prepare_entries(gpt_params & params, train_context & ctx_train) {
static int prepare_entries(common_params & params, train_context & ctx_train) {
// load prompts
std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
@ -386,10 +392,9 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
return 1;
}
@ -412,14 +417,15 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa);
// load the model to get hparams
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
// int n_ctx = llama_n_ctx(ctx);
int n_layers = llama_n_layer(model);
int n_embd = llama_n_embd(model);
int n_layers = llama_model_n_layer(model);
int n_embd = llama_model_n_embd(model);
// get model hint param (a.k.a model arch name)
char model_hint[128];
llama_model_meta_val_str(model, "general.architecture", model_hint, 128);
@ -473,8 +479,6 @@ int main(int argc, char ** argv) {
// done with the model, we can now free it to make gain some memory
printf("Done evaluate prompts, unload model...\n");
llama_free(ctx);
llama_free_model(model);
bool use_pca = params.cvector_dimre_method == DIMRE_METHOD_PCA;

2
examples/cvector-generator/mean.hpp
View file

@ -15,7 +15,7 @@ static void run(
for (size_t il = 0; il < v_input.size(); ++il) {
// prepare output vector
struct ggml_tensor * ctrl_out = v_output[il];
ggml_format_name(ctrl_out, "direction.%ld", il+1);
ggml_format_name(ctrl_out, "direction.%zu", il+1);
// calculate mean vector
struct ggml_tensor * t_layer = v_input[il];

18
examples/cvector-generator/pca.hpp
View file

@ -12,12 +12,9 @@
#include <cstdio>
#include <ctime>
#include <random>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#define DEBUG_POS 5
@ -207,13 +204,6 @@ static ggml_status compute_piter(
ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
}
// TODO: enable GPU support when support for GGML_OP_SQRT is added
//#ifdef GGML_USE_METAL
// if (ggml_backend_is_metal(model.backend)) {
// ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
// }
//#endif
ggml_status res = ggml_backend_graph_compute(model.backend, gf);
if (res == GGML_STATUS_SUCCESS) {
auto extract_i = [](std::string prefix, std::string str) -> int {
@ -229,8 +219,8 @@ static ggml_status compute_piter(
result.eigenvectors.resize(params.n_batch);
result.distances.resize(params.n_batch);
// get output nodes
for (int i = 0; i < gf->n_nodes; ++i) {
auto node = gf->nodes[i];
for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
auto node = ggml_graph_node(gf, i);
int iter = -1;
// find b_tensor (without copying data from device)
if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {
@ -312,7 +302,7 @@ static void run_pca(
// prepare output vector
struct ggml_tensor * ctrl_out = v_output[il];
ggml_format_name(ctrl_out, "direction.%ld", il+1);
ggml_format_name(ctrl_out, "direction.%zu", il+1);
// run power_iteration
params.i_layer = il;

2
examples/deprecation-warning/deprecation-warning.cpp
View file

@ -12,7 +12,7 @@ int main(int argc, char** argv) {
}
// Get only the program name from the full path
auto pos = filename.find_last_of('/');
auto pos = filename.find_last_of("/\\");
if (pos != std::string::npos) {
filename = filename.substr(pos+1);
}

2
examples/embedding/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-embedding)
add_executable(${TARGET} embedding.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

151
examples/embedding/embedding.cpp
View file

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <ctime>
@ -26,7 +28,7 @@ static std::vector<std::string> split_lines(const std::string & s, const std::st
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
size_t n_tokens = tokens.size();
for (size_t i = 0; i < n_tokens; i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, true);
common_batch_add(batch, tokens[i], i, { seq_id }, true);
}
}
@ -38,16 +40,16 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
// encoder-only model
if (llama_encode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to encode\n", __func__);
LOG_ERR("%s : failed to encode\n", __func__);
}
} else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
// decoder-only model
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
LOG_ERR("%s : failed to decode\n", __func__);
}
}
@ -72,58 +74,58 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
}
float * out = output + embd_pos * n_embd;
llama_embd_normalize(embd, out, n_embd, embd_norm);
common_embd_normalize(embd, out, n_embd, embd_norm);
}
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EMBEDDING);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
return 1;
}
common_init();
params.embedding = true;
// For non-causal models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch;
print_build_info();
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
llama_backend_init();
llama_numa_init(params.numa);
// load the model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
LOG_ERR("%s: computing embeddings in encoder-decoder models is not supported\n", __func__);
return 1;
}
if (n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
// split the prompt into lines
@ -136,9 +138,9 @@ int main(int argc, char ** argv) {
// tokenize the prompts and trim
std::vector<std::vector<int32_t>> inputs;
for (const auto & prompt : prompts) {
auto inp = ::llama_tokenize(ctx, prompt, true, false);
auto inp = common_tokenize(ctx, prompt, true, true);
if (inp.size() > n_batch) {
fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch);
return 1;
}
@ -148,21 +150,21 @@ int main(int argc, char ** argv) {
// check if the last token is SEP
// it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
for (auto & inp : inputs) {
if (inp.empty() || inp.back() != llama_token_sep(model)) {
fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
fprintf(stderr, "%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
}
}
// tokenization stats
if (params.verbose_prompt) {
for (int i = 0; i < (int) inputs.size(); i++) {
fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
for (int j = 0; j < (int) inputs[i].size(); j++) {
fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
LOG("%6d -> '%s'\n", inputs[i][j], common_token_to_piece(ctx, inputs[i][j]).c_str());
}
fprintf(stderr, "\n\n");
LOG("\n\n");
}
}
@ -181,7 +183,7 @@ int main(int argc, char ** argv) {
}
// allocate output
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
std::vector<float> embeddings(n_embd_count * n_embd, 0);
float * emb = embeddings.data();
@ -200,7 +202,7 @@ int main(int argc, char ** argv) {
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
s = 0;
llama_batch_clear(batch);
common_batch_clear(batch);
}
// add to batch
@ -213,57 +215,62 @@ int main(int argc, char ** argv) {
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
if (params.embd_out.empty()) {
fprintf(stdout, "\n");
LOG("\n");
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
for (int j = 0; j < n_embd_count; j++) {
fprintf(stdout, "embedding %d: ", j);
LOG("embedding %d: ", j);
for (int i = 0; i < std::min(3, n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, " ... ");
LOG(" ... ");
for (int i = n_embd - 3; i < n_embd; i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
LOG("\n");
}
} else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
for (int j = 0; j < n_embd_count; j++) {
// NOTE: if you change this log - update the tests in ci/run.sh
LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
}
} else {
// print the first part of the embeddings or for a single prompt, the full embedding
for (int j = 0; j < n_prompts; j++) {
fprintf(stdout, "embedding %d: ", j);
LOG("embedding %d: ", j);
for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
LOG("\n");
}
// print cosine similarity matrix
if (n_prompts > 1) {
fprintf(stdout, "\n");
printf("cosine similarity matrix:\n\n");
LOG("\n");
LOG("cosine similarity matrix:\n\n");
for (int i = 0; i < n_prompts; i++) {
fprintf(stdout, "%6.6s ", prompts[i].c_str());
LOG("%6.6s ", prompts[i].c_str());
}
fprintf(stdout, "\n");
LOG("\n");
for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f ", sim);
float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
LOG("%6.2f ", sim);
}
fprintf(stdout, "%1.10s", prompts[i].c_str());
fprintf(stdout, "\n");
LOG("%1.10s", prompts[i].c_str());
LOG("\n");
}
}
}
@ -272,48 +279,46 @@ int main(int argc, char ** argv) {
if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
const bool notArray = params.embd_out != "array";
fprintf(stdout, notArray ? "{\n \"object\": \"list\",\n \"data\": [\n" : "[");
LOG(notArray ? "{\n \"object\": \"list\",\n \"data\": [\n" : "[");
for (int j = 0;;) { // at least one iteration (one prompt)
if (notArray) fprintf(stdout, " {\n \"object\": \"embedding\",\n \"index\": %d,\n \"embedding\": ",j);
fprintf(stdout, "[");
if (notArray) LOG(" {\n \"object\": \"embedding\",\n \"index\": %d,\n \"embedding\": ",j);
LOG("[");
for (int i = 0;;) { // at least one iteration (n_embd > 0)
fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
LOG(params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
i++;
if (i < n_embd) fprintf(stdout, ","); else break;
if (i < n_embd) LOG(","); else break;
}
fprintf(stdout, notArray ? "]\n }" : "]");
LOG(notArray ? "]\n }" : "]");
j++;
if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
if (j < n_embd_count) LOG(notArray ? ",\n" : ","); else break;
}
fprintf(stdout, notArray ? "\n ]" : "]\n");
LOG(notArray ? "\n ]" : "]\n");
if (params.embd_out == "json+" && n_prompts > 1) {
fprintf(stdout, ",\n \"cosineSimilarity\": [\n");
LOG(",\n \"cosineSimilarity\": [\n");
for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
fprintf(stdout, " [");
LOG(" [");
for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f", sim);
float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
LOG("%6.2f", sim);
j++;
if (j < n_embd_count) fprintf(stdout, ", "); else break;
if (j < n_embd_count) LOG(", "); else break;
}
fprintf(stdout, " ]");
LOG(" ]");
i++;
if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
if (i < n_embd_count) LOG(",\n"); else break;
}
fprintf(stdout, "\n ]");
LOG("\n ]");
}
if (notArray) fprintf(stdout, "\n}\n");
if (notArray) LOG("\n}\n");
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
// clean up
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
return 0;

5
examples/eval-callback/CMakeLists.txt
View file

@ -2,8 +2,9 @@ set(TARGET llama-eval-callback)
add_executable(${TARGET} eval-callback.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TEST_TARGET test-eval-callback)
add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
add_test(NAME ${TEST_TARGET}
COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

79
examples/eval-callback/eval-callback.cpp
View file

@ -1,11 +1,11 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include "ggml.h"
#include <cstdio>
#include <random>
#include <string>
#include <tuple>
#include <vector>
/**
@ -31,22 +31,22 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
GGML_ASSERT(n > 0);
float sum = 0;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
printf(" [\n");
LOG(" [\n");
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
if (i2 == n && ne[2] > 2*n) {
printf(" ..., \n");
LOG(" ..., \n");
i2 = ne[2] - n;
}
printf(" [\n");
LOG(" [\n");
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
if (i1 == n && ne[1] > 2*n) {
printf(" ..., \n");
LOG(" ..., \n");
i1 = ne[1] - n;
}
printf(" [");
LOG(" [");
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
if (i0 == n && ne[0] > 2*n) {
printf("..., ");
LOG("..., ");
i0 = ne[0] - n;
}
size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
@ -64,16 +64,16 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
} else {
GGML_ABORT("fatal error");
}
printf("%12.4f", v);
LOG("%12.4f", v);
sum += v;
if (i0 < ne[0] - 1) printf(", ");
if (i0 < ne[0] - 1) LOG(", ");
}
printf("],\n");
LOG("],\n");
}
printf(" ],\n");
LOG(" ],\n");
}
printf(" ]\n");
printf(" sum = %f\n", sum);
LOG(" ]\n");
LOG(" sum = %f\n", sum);
}
}
@ -102,11 +102,11 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
}
printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_desc(t),
src0->name, ggml_ne_string(src0).c_str(),
src1 ? src1_str : "",
ggml_ne_string(t).c_str());
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_desc(t),
src0->name, ggml_ne_string(src0).c_str(),
src1 ? src1_str : "",
ggml_ne_string(t).c_str());
// copy the data from the GPU memory if needed
@ -126,13 +126,16 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
return true;
}
static bool run(llama_context * ctx, const gpt_params & params) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
static bool run(llama_context * ctx, const common_params & params) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
const bool add_bos = llama_vocab_get_add_bos(vocab);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
@ -142,14 +145,13 @@ static bool run(llama_context * ctx, const gpt_params & params) {
int main(int argc, char ** argv) {
callback_data cb_data;
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
print_build_info();
common_init();
llama_backend_init();
llama_numa_init(params.numa);
@ -161,19 +163,21 @@ int main(int argc, char ** argv) {
params.warmup = false;
// init
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n");
}
bool OK = run(ctx, params);
@ -181,11 +185,8 @@ int main(int argc, char ** argv) {
return 1;
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_free(ctx);
llama_free_model(model);
LOG("\n");
llama_perf_context_print(ctx);
llama_backend_free();

2
examples/export-lora/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-export-lora)
add_executable(${TARGET} export-lora.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

41
examples/export-lora/export-lora.cpp
View file

@ -1,11 +1,13 @@
#include "common.h"
#include "ggml.h"
#include "ggml-alloc.h"
#include "gguf.h"
#include "arg.h"
#include "common.h"
#include <map>
#include <vector>
#include <string>
#include <thread>
#include <fstream>
static bool g_verbose = false;
@ -127,7 +129,7 @@ struct lora_merge_ctx {
lora_merge_ctx(
std::string & base_fname,
std::vector<llama_lora_adapter_info> & lora_files,
std::vector<common_adapter_lora_info> & lora_files,
std::string & outfile,
int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@ -264,8 +266,8 @@ struct lora_merge_ctx {
fout.write((const char *)data.data(), data.size());
}
printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
printf("%s : merged %zu tensors with lora adapters\n", __func__, n_merged);
printf("%s : wrote %zu tensors to output file\n", __func__, trans.size());
}
void copy_tensor(struct ggml_tensor * base) {
@ -313,9 +315,9 @@ struct lora_merge_ctx {
// optionally dequantize it
printf("%s : + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
auto nels = ggml_nelements(inp_base);
ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
const auto * qtype = ggml_get_type_traits(base->type);
std::vector<uint8_t> dequant_buf(nels * sizeof(float));
qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
qtype->to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
} else {
ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
@ -343,15 +345,25 @@ struct lora_merge_ctx {
gf = ggml_new_graph(ctx0);
struct ggml_tensor * cur = inp_base;
for (size_t i = 0; i < adapters.size(); ++i) {
struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32)));
struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
struct ggml_tensor * delta;
bool is_tok_embd = string_starts_with(name_base, "token_embd");
if (is_tok_embd) {
printf("%s : detected token embeddings tensor\n", __func__);
delta = ggml_mul_mat(ctx0,
ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32),
ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32));
} else {
delta = ggml_mul_mat(ctx0,
ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32))),
ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
}
// scale
const float alpha = adapters[i]->alpha;
const float rank = (float) inp_b[i]->ne[0];
const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
delta = ggml_scale(ctx0, delta, scale);
cur = ggml_add(ctx0, delta, cur);
printf("%s : + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
printf("%s : + merging from adapter[%zu] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
printf("%s : input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
}
cur = ggml_cast(ctx0, cur, out->type);
@ -369,7 +381,7 @@ struct lora_merge_ctx {
// write data to output file
{
auto result = gf->nodes[gf->n_nodes - 1];
auto * result = ggml_graph_node(gf, -1);
size_t len = ggml_nbytes(result);
if (read_buf.size() < len) {
read_buf.resize(len);
@ -399,14 +411,13 @@ static void print_usage(int, char ** argv) {
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
return 1;
}
g_verbose = (params.verbosity == 1);
g_verbose = (params.verbosity > 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
ctx.run_merge();

2
examples/gbnf-validator/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-gbnf-validator)
add_executable(${TARGET} gbnf-validator.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

13
examples/gbnf-validator/gbnf-validator.cpp
View file

@ -11,19 +11,15 @@
static bool llama_grammar_validate(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
const auto cpts = unicode_cpts_from_utf8(input_str);
const llama_grammar_rules & rules = llama_grammar_get_rules (grammar);
llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
auto & stacks_cur = llama_grammar_get_stacks(grammar);
size_t pos = 0;
for (const auto & cpt : cpts) {
const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
llama_grammar_accept(grammar, cpt);
if (stacks_cur.empty()) {
error_pos = pos;
error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
stacks_cur = stacks_prev;
return false;
}
++pos;
@ -80,9 +76,10 @@ int main(int argc, char** argv) {
grammar_str = buffer.str();
}
llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root", false, nullptr, 0, nullptr, 0);
if (grammar == nullptr) {
throw std::runtime_error("Failed to initialize llama_grammar");
fprintf(stdout, "Failed to initialize llama_grammar\n");
return 1;
}
// Read the input file
std::string input_str;

2
examples/gen-docs/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-gen-docs)
add_executable(${TARGET} gen-docs.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

92
examples/gen-docs/gen-docs.cpp
View file

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include <fstream>
@ -5,42 +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 common_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<common_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 options = gpt_params_parser_init(params, ex);
common_params params;
auto ctx_arg = common_params_parser_init(params, ex);
file << "| Argument | Explanation |\n";
file << "| -------- | ----------- |\n";
for (auto & opt : 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() ? ", " : "");
}
std::vector<common_arg *> common_options;
std::vector<common_arg *> sparam_options;
std::vector<common_arg *> specific_options;
for (auto & opt : ctx_arg.options) {
// 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);
}
// 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";
}
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 **) {

9
examples/gguf-hash/CMakeLists.txt
View file

@ -4,12 +4,19 @@ install(TARGETS ${TARGET} RUNTIME)
# clibs dependencies
include_directories(deps/)
add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
target_link_libraries(${TARGET} PRIVATE xxhash)
add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
target_link_libraries(${TARGET} PRIVATE sha1)
if (NOT MSVC)
# disable warnings in 3rd party code
target_compile_options(sha1 PRIVATE -w)
endif()
add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
target_link_libraries(${TARGET} PRIVATE sha256)
target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

1
examples/gguf-hash/gguf-hash.cpp
View file

@ -1,4 +1,5 @@
#include "ggml.h"
#include "gguf.h"
#include <cstdlib> /* abort() */
#include <cstddef>

2
examples/gguf-split/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-gguf-split)
add_executable(${TARGET} gguf-split.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

114
examples/gguf-split/gguf-split.cpp
View file

@ -1,18 +1,19 @@
#include "ggml.h"
#include "gguf.h"
#include "llama.h"
#include "common.h"
#include <algorithm>
#include <cmath>
#include <cinttypes>
#include <climits>
#include <cstdio>
#include <cstdlib>
#include <stdexcept>
#include <cstring>
#include <fstream>
#include <string>
#include <vector>
#include <stdio.h>
#include <string.h>
#include <climits>
#include <stdexcept>
#if defined(_WIN32)
#include <windows.h>
#ifndef PATH_MAX
@ -22,12 +23,20 @@
#endif
enum split_operation : uint8_t {
SPLIT_OP_SPLIT,
SPLIT_OP_MERGE,
OP_NONE,
OP_SPLIT,
OP_MERGE,
};
enum split_mode : uint8_t {
MODE_NONE,
MODE_TENSOR,
MODE_SIZE,
};
struct split_params {
split_operation operation = SPLIT_OP_SPLIT;
split_operation operation = OP_NONE;
split_mode mode = MODE_NONE;
size_t n_bytes_split = 0;
int n_split_tensors = 128;
std::string input;
@ -87,59 +96,52 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
}
bool arg_found = false;
bool is_op_set = false;
bool is_mode_set = false;
if (arg == "-h" || arg == "--help") {
split_print_usage(argv[0]);
exit(0);
}
if (arg == "--version") {
} else if (arg == "--version") {
fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
exit(0);
}
if (arg == "--dry-run") {
} else if (arg == "--dry-run") {
arg_found = true;
params.dry_run = true;
}
if (arg == "--no-tensor-first-split") {
} else if (arg == "--no-tensor-first-split") {
arg_found = true;
params.no_tensor_first_split = true;
}
if (is_op_set) {
throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
}
if (arg == "--merge") {
} else if (arg == "--merge") {
arg_found = true;
is_op_set = true;
params.operation = SPLIT_OP_MERGE;
}
if (arg == "--split") {
if (params.operation != OP_NONE && params.operation != OP_MERGE) {
throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
}
params.operation = OP_MERGE;
} else if (arg == "--split") {
arg_found = true;
is_op_set = true;
params.operation = SPLIT_OP_SPLIT;
}
if (is_mode_set) {
throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
}
if (arg == "--split-max-tensors") {
if (params.operation != OP_NONE && params.operation != OP_SPLIT) {
throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
}
params.operation = OP_SPLIT;
} else if (arg == "--split-max-tensors") {
if (++arg_idx >= argc) {
invalid_param = true;
break;
}
arg_found = true;
is_mode_set = true;
if (params.mode != MODE_NONE && params.mode != MODE_TENSOR) {
throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
}
params.mode = MODE_TENSOR;
params.n_split_tensors = atoi(argv[arg_idx]);
}
if (arg == "--split-max-size") {
} else if (arg == "--split-max-size") {
if (++arg_idx >= argc) {
invalid_param = true;
break;
}
arg_found = true;
is_mode_set = true;
if (params.mode != MODE_NONE && params.mode != MODE_SIZE) {
throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
}
params.mode = MODE_SIZE;
params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
}
@ -148,11 +150,20 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
}
}
// the operation is split if not specified
if (params.operation == OP_NONE) {
params.operation = OP_SPLIT;
}
// the split mode is by tensor if not specified
if (params.mode == MODE_NONE) {
params.mode = MODE_TENSOR;
}
if (invalid_param) {
throw std::invalid_argument("error: invalid parameter for argument: " + arg);
}
if (argc - arg_idx < 2) {
if (argc - arg_idx != 2) {
throw std::invalid_argument("error: bad arguments");
}
@ -265,17 +276,19 @@ struct split_strategy {
}
bool should_split(int i_tensor, size_t next_size) {
if (params.n_bytes_split > 0) {
if (params.mode == MODE_SIZE) {
// split by max size per file
return next_size > params.n_bytes_split;
} else {
} else if (params.mode == MODE_TENSOR) {
// split by number of tensors per file
return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
}
// should never happen
GGML_ABORT("invalid mode");
}
void print_info() {
printf("n_split: %ld\n", ctx_outs.size());
printf("n_split: %zu\n", ctx_outs.size());
int i_split = 0;
for (auto & ctx_out : ctx_outs) {
// re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
@ -285,7 +298,7 @@ struct split_strategy {
total_size += ggml_nbytes(t);
}
total_size = total_size / 1000 / 1000; // convert to megabytes
printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
printf("split %05d: n_tensors = %" PRIi64 ", total_size = %zuM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
i_split++;
}
}
@ -389,10 +402,17 @@ static void gguf_merge(const split_params & split_params) {
int n_split = 1;
int total_tensors = 0;
auto * ctx_out = gguf_init_empty();
// avoid overwriting existing output file
if (std::ifstream(split_params.output.c_str())) {
fprintf(stderr, "%s: output file %s already exists\n", __func__, split_params.output.c_str());
exit(EXIT_FAILURE);
}
std::ofstream fout(split_params.output.c_str(), std::ios::binary);
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
auto * ctx_out = gguf_init_empty();
std::vector<uint8_t> read_data;
std::vector<ggml_context *> ctx_metas;
std::vector<gguf_context *> ctx_ggufs;
@ -552,9 +572,9 @@ int main(int argc, const char ** argv) {
split_params_parse(argc, argv, params);
switch (params.operation) {
case SPLIT_OP_SPLIT: gguf_split(params);
case OP_SPLIT: gguf_split(params);
break;
case SPLIT_OP_MERGE: gguf_merge(params);
case OP_MERGE: gguf_merge(params);
break;
default: split_print_usage(argv[0]);
exit(EXIT_FAILURE);

10
examples/gguf-split/tests.sh
View file

@ -41,7 +41,7 @@ echo PASS
echo
# 2b. Test the sharded model is loading properly
$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --n-predict 32
echo PASS
echo
@ -51,7 +51,7 @@ echo PASS
echo
# 3b. Test the merged model is loading properly
$MAIN --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge.gguf --n-predict 32
echo PASS
echo
@ -61,7 +61,7 @@ echo PASS
echo
# 4b. Test the sharded model is loading properly
$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00007.gguf --n-predict 32
echo PASS
echo
@ -71,7 +71,7 @@ echo
#echo
# 5b. Test the merged model is loading properly
#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
#echo PASS
#echo
@ -81,7 +81,7 @@ echo PASS
echo
# 6b. Test the sharded model is loading properly
$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --n-predict 32
echo PASS
echo

2
examples/gguf/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-gguf)
add_executable(${TARGET} gguf.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

16
examples/gguf/gguf.cpp
View file

@ -1,10 +1,9 @@
#include "ggml.h"
#include "gguf.h"
#include <cstdio>
#include <cinttypes>
#include <string>
#include <sstream>
#include <fstream>
#include <vector>
#undef MIN
@ -135,9 +134,10 @@ static bool gguf_ex_read_0(const std::string & fname) {
for (int i = 0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name (ctx, i);
const size_t size = gguf_get_tensor_size (ctx, i);
const size_t offset = gguf_get_tensor_offset(ctx, i);
printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
}
}
@ -182,9 +182,10 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
for (int i = 0; i < n_tensors; ++i) {
const char * name = gguf_get_tensor_name (ctx, i);
const size_t size = gguf_get_tensor_size (ctx, i);
const size_t offset = gguf_get_tensor_offset(ctx, i);
printf("%s: tensor[%d]: name = %s, offset = %zu\n", __func__, i, name, offset);
printf("%s: tensor[%d]: name = %s, size = %zu, offset = %zu\n", __func__, i, name, size, offset);
}
}
@ -199,7 +200,8 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
printf("%s: tensor[%d]: n_dims = %d, ne = (%d, %d, %d, %d), name = %s, data = %p\n",
__func__, i, ggml_n_dims(cur), int(cur->ne[0]), int(cur->ne[1]), int(cur->ne[2]), int(cur->ne[3]), cur->name, cur->data);
// print first 10 elements
const float * data = (const float *) cur->data;
@ -215,7 +217,7 @@ static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
const float * data = (const float *) cur->data;
for (int j = 0; j < ggml_nelements(cur); ++j) {
if (data[j] != 100 + i) {
fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
fprintf(stderr, "%s: tensor[%d], data[%d]: found %f, expected %f\n", __func__, i, j, data[j], float(100 + i));
gguf_free(ctx);
return false;
}
@ -245,6 +247,8 @@ int main(int argc, char ** argv) {
check_data = false;
}
srand(123456);
const std::string fname(argv[1]);
const std::string mode (argv[2]);

2
examples/gritlm/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-gritlm)
add_executable(${TARGET} gritlm.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

56
examples/gritlm/gritlm.cpp
View file

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include "llama.h"
@ -10,24 +11,25 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
std::vector<std::vector<float>> result;
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
for (uint64_t i = 0; i < sentences.size(); i++) {
llama_batch_clear(batch);
common_batch_clear(batch);
const std::string input_string = instruction + sentences[i];
std::vector<llama_token> inputs = llama_tokenize(model, input_string, true, false);
std::vector<llama_token> inputs = common_tokenize(vocab, input_string, true, false);
const int32_t n_toks = inputs.size();
// GritLM seems to have EOS = ""
// https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
// inputs.push_back(llama_token_eos(model));
// inputs.push_back(llama_vocab_eos(vocab));
// we want to ignore instruction tokens for mean pooling
const int32_t n_inst = llama_tokenize(model, instruction, true, false).size();
const int32_t n_inst = common_tokenize(vocab, instruction, true, false).size();
#ifdef GRIT_DEBUG
// debug tokens - should be matching as referenced in the GritLM sample
@ -39,7 +41,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
// add input to batch (this increments n_tokens)
for (int32_t j = 0; j < n_toks; j++) {
llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
}
// clear previous kv_cache values (irrelevant for embeddings)
@ -51,7 +53,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
llama_decode(ctx, batch);
// get embedding dimensions
uint64_t n_embd = llama_n_embd(model);
uint64_t n_embd = llama_model_n_embd(model);
// allocate embedding output
std::vector<float> emb_unorm(n_embd, 0.0f);
@ -74,7 +76,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
}
std::vector<float> emb_norm(emb_unorm.size());
llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd, 2);
result.push_back(emb_norm);
#ifdef GRIT_DEBUG
@ -96,7 +98,9 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
std::string result;
const llama_model * model = llama_get_model(ctx);
llama_token eos_token = llama_token_eos(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_token eos_token = llama_vocab_eos(vocab);
llama_kv_cache_clear(ctx);
llama_set_embeddings(ctx, false);
@ -104,16 +108,16 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true);
std::vector<llama_token> inputs = common_tokenize(vocab, prompt, false, true);
int32_t i_current_token = 0;
while (true) {
llama_batch_clear(bat);
common_batch_clear(bat);
{
const int32_t n_inputs = inputs.size();
for (int32_t i = 0; i < n_inputs; i++) {
llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
common_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
}
}
inputs.clear();
@ -121,13 +125,12 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
llama_decode(ctx, bat);
llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
llama_sampler_accept(smpl, token);
if (token == eos_token) {
break;
}
std::string piece = llama_token_to_piece(ctx, token);
std::string piece = common_token_to_piece(ctx, token);
if (stream) {
std::printf("%s", piece.c_str());
std::fflush(stdout);
@ -152,22 +155,23 @@ static std::string gritlm_instruction(const std::string & instruction) {
}
int main(int argc, char * argv[]) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
llama_model_params mparams = llama_model_params_from_gpt_params(params);
llama_context_params cparams = llama_context_params_from_gpt_params(params);
common_init();
llama_model_params mparams = common_model_params_to_llama(params);
llama_context_params cparams = common_context_params_to_llama(params);
llama_backend_init();
llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams);
// create generation context
llama_context * ctx = llama_new_context_with_model(model, cparams);
llama_context * ctx = llama_init_from_model(model, cparams);
auto sparams = llama_sampler_chain_default_params();
@ -196,12 +200,12 @@ int main(int argc, char * argv[]) {
const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
const std::vector<std::vector<float>> q_rep = encode(ctx, queries, gritlm_instruction(instruction));
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
@ -218,7 +222,7 @@ int main(int argc, char * argv[]) {
llama_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();
return 0;

2
examples/imatrix/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-imatrix)
add_executable(${TARGET} imatrix.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

2
examples/imatrix/README.md
View file

@ -25,8 +25,6 @@ For faster computation, make sure to use GPU offloading via the `-ngl` argument
## Example
```bash
GGML_CUDA=1 make -j
# generate importance matrix (imatrix.dat)
./llama-imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99

176
examples/imatrix/imatrix.cpp
View file

@ -1,5 +1,8 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include "gguf.h"
#include <cmath>
#include <cstdio>
@ -17,12 +20,12 @@
#endif
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s \\\n"
" -m model.gguf -f some-text.txt [-o imatrix.gguf] [--process-output] [--verbosity 1] \\\n"
LOG("\nexample usage:\n");
LOG("\n %s \\\n"
" -m model.gguf -f some-text.txt [-o imatrix.gguf] [--process-output] \\\n"
" [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
" [--in-file imatrix-prev-0.gguf --in-file imatrix-prev-1.gguf ...]\n" , argv[0]);
LOG_TEE("\n");
LOG("\n");
}
static bool str_remove_suffix(std::string & str, const std::string & suffix) {
@ -45,13 +48,13 @@ struct Stats {
class IMatrixCollector {
public:
IMatrixCollector() = default;
void set_params(gpt_params params) { m_params = std::move(params); }
void set_params(common_params params) { m_params = std::move(params); }
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
void save_imatrix(int32_t n_chunk = -1) const;
bool load_imatrix(const char * file_name);
private:
std::unordered_map<std::string, Stats> m_stats;
gpt_params m_params;
common_params m_params;
std::mutex m_mutex;
int32_t m_last_chunk = 0;
std::vector<float> m_src1_data;
@ -136,16 +139,14 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
e.counts.resize(n_as, 0);
}
else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
exit(1); //GGML_ABORT("fatal error");
}
else if (e.counts.size() != (size_t)n_as) {
fprintf(stderr, "Oops: inconsistent expert count for %s (%d vs %d)\n", wname.c_str(), (int)e.counts.size(), (int)n_as);
LOG_ERR("Oops: inconsistent expert count for %s (%d vs %d)\n", wname.c_str(), (int)e.counts.size(), (int)n_as);
exit(1); //GGML_ABORT("fatal error");
}
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
}
LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
// loop over all possible experts, regardless if they are used or not in the batch
for (int ex = 0; ex < n_as; ++ex) {
size_t e_start = ex*src1->ne[0];
@ -167,7 +168,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
for (int j = 0; j < (int)src1->ne[0]; ++j) {
e.values[e_start + j] = std::fma(x[j], x[j], e.values[e_start + j]);
if (!std::isfinite((float)e.values[e_start + j])) {
fprintf(stderr, "%f detected in %s\n", (float)e.values[e_start + j], wname.c_str());
LOG_ERR("%f detected in %s\n", (float)e.values[e_start + j], wname.c_str());
exit(1);
}
}
@ -192,16 +193,14 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
e.counts.resize(1, 0);
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
exit(1); //GGML_ABORT("fatal error");
}
else if (e.counts.size() != 1) {
fprintf(stderr, "Oops: inconsistent expert count for %s (%d vs %d)\n", wname.c_str(), (int)e.counts.size(), 1);
LOG_ERR("Oops: inconsistent expert count for %s (%d vs %d)\n", wname.c_str(), (int)e.counts.size(), 1);
exit(1); //GGML_ABORT("fatal error");
}
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
}
LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_chunk, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
// TODO: higher dimensions
for (int row = 0; row < (int)src1->ne[1]; ++row) {
const float * x = data + row * src1->ne[0];
@ -209,7 +208,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
for (int j = 0; j < (int)src1->ne[0]; ++j) {
e.values[j] = std::fma(x[j], x[j], e.values[j]);
if (!std::isfinite((float)e.values[j])) {
fprintf(stderr, "%f detected in %s\n", (float)e.values[j], wname.c_str());
LOG_ERR("%f detected in %s\n", (float)e.values[j], wname.c_str());
exit(1);
}
}
@ -263,17 +262,17 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
}
if (n_zeros != 0 && is_first) {
fprintf(stderr, "\n");
LOG_INF("\n");
is_first = false;
}
if (n_zeros == n_all) {
fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
continue;
}
if (n_zeros > 0) {
fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
LOG_WRN("%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
continue;
}
@ -283,7 +282,7 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
}
if (to_store.size() < m_stats.size()) {
fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
}
// deterministic tensor name order
@ -328,9 +327,8 @@ void IMatrixCollector::save_imatrix(int32_t n_chunk) const {
gguf_write_to_file(ctx_gguf, fname.c_str(), false);
if (m_params.verbosity > 0) {
fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());
}
LOGV(1, "\n");
LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_chunk, fname.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
@ -348,7 +346,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
}
const int32_t n_entries = gguf_get_n_tensors(ctx_gguf);
if (n_entries < 1) {
fprintf(stderr, "%s: no data in file %s\n", __func__, file_name);
LOG_ERR("%s: no data in file %s\n", __func__, file_name);
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
@ -375,7 +373,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
// counts
sums_counts_for[name].second = cur;
} else {
fprintf(stderr, "%s: invalid imatrix tensor name: %s\n", __func__, name.c_str());
LOG_ERR("%s: invalid imatrix tensor name: %s\n", __func__, name.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
@ -388,7 +386,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
const struct ggml_tensor * counts = sc.second.second;
if (!sums || !counts) {
fprintf(stderr, "%s: mismatched sums and counts for %s\n", __func__, name.c_str());
LOG_ERR("%s: mismatched sums and counts for %s\n", __func__, name.c_str());
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
@ -400,7 +398,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
if (e.values.empty()) {
e.values.resize(nval, 0);
} else if ((size_t) nval != e.values.size()) {
fprintf(stderr, "%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size());
LOG_ERR("%s: mismatched sums size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) nval, e.values.size());
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
@ -413,7 +411,7 @@ bool IMatrixCollector::load_imatrix(const char * file_name) {
// broadcast, when loading an old imatrix
e.counts.resize(ncounts, e.counts[0]);
} else if ((size_t) ncounts != e.counts.size()) {
fprintf(stderr, "%s: mismatched counts size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) ncounts, e.counts.size());
LOG_ERR("%s: mismatched counts size for %s: %zu != %zu\n", __func__, name.c_str(), (size_t) ncounts, e.counts.size());
gguf_free(ctx_gguf);
ggml_free(ctx);
return false;
@ -511,31 +509,34 @@ static void process_logits(
}
}
static bool compute_imatrix(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
static bool compute_imatrix(llama_context * ctx, const common_params & params, const int32_t n_ctx) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
auto tim1 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
auto tim2 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
if (params.i_chunk > 0) {
if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
return false;
}
fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
}
if (int(tokens.size()) < 2*n_ctx) {
fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
n_ctx);
fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);
LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());
return false;
}
@ -550,15 +551,13 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
const int n_chunk_max = tokens.size() / n_ctx;
const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_vocab = llama_vocab_n_tokens(vocab);
const int n_batch = params.n_batch;
int count = 0;
double nll = 0.0;
double nll2 = 0.0;
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
const int num_batches = (n_ctx + n_batch - 1) / n_batch;
const int n_seq = std::max(1, n_batch / n_ctx);
@ -572,7 +571,9 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
logits.reserve((size_t)n_ctx * n_vocab);
}
fprintf(stderr, "%s: computing over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
LOG_INF("%s: computing over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
for (int i = 0; i < n_chunk; i += n_seq) {
const int start = i * n_ctx;
@ -590,7 +591,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
const int batch_size = std::min(end - batch_start, n_batch);
// clear the batch
llama_batch_clear(batch);
common_batch_clear(batch);
for (int seq = 0; seq < n_seq_batch; seq++) {
int seq_start = batch_start + seq*n_ctx;
@ -600,23 +601,23 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
// add BOS token for the first batch of each chunk
if (add_bos && j == 0) {
tokens[seq_start] = llama_token_bos(llama_get_model(ctx));
tokens[seq_start] = llama_vocab_bos(vocab);
}
for (int k = 0; k < batch_size; ++k) {
// NOTE: specifying all logits to get activations for the output.weight tensor
// and also for the perplexity calculation.
// TODO: only get outputs when (params.process_output || params.compute_ppl)
// (not possible when this skips FFN computation of the last layer)
llama_batch_add(batch, tokens[seq_start + k], j*n_batch + k, { seq }, true);
common_batch_add(batch, tokens[seq_start + k], j*n_batch + k, { seq }, true);
}
// restore the original token in case it was set to BOS
tokens[seq_start] = token_org;
}
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
llama_batch_free(batch);
return false;
}
@ -631,13 +632,13 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
llama_synchronize(ctx);
const auto t_end = std::chrono::high_resolution_clock::now();
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total*n_chunk/n_seq);
LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total * n_chunk / n_seq);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
LOG("%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
LOG("%.2f minutes\n", total_seconds / 60.0);
}
if (params.compute_ppl) {
@ -655,14 +656,15 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
count += n_ctx - first - 1;
printf("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
LOG("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
}
fflush(stdout);
logits.clear();
}
}
printf("\n");
LOG("\n");
if (params.compute_ppl) {
nll2 /= count;
@ -671,31 +673,34 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, cons
nll2 -= nll * nll;
if (nll2 > 0) {
nll2 = sqrt(nll2/(count-1));
printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
} else {
printf("Unexpected negative standard deviation of log(prob)\n");
LOG("Unexpected negative standard deviation of log(prob)\n");
}
}
llama_batch_free(batch);
return true;
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
params.n_ctx = 512;
params.logits_all = true;
params.verbosity = 1;
params.escape = false;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_IMATRIX, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
return 1;
}
common_init();
const int32_t n_ctx = params.n_ctx;
if (n_ctx <= 0) {
fprintf(stderr, "%s: imatrix tool requires '--ctx-size' > 0\n", __func__);
LOG_ERR("%s: imatrix tool requires '--ctx-size' > 0\n", __func__);
return 1;
}
@ -712,15 +717,15 @@ int main(int argc, char ** argv) {
g_collector.set_params(params);
for (const auto & in_file : params.in_files) {
printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
if (!g_collector.load_imatrix(in_file.c_str())) {
fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());
return 1;
}
}
if (params.in_files.size() > 1) {
printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
g_collector.save_imatrix();
}
@ -734,38 +739,45 @@ int main(int argc, char ** argv) {
params.warmup = false;
// init
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, params.n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
if (!compute_imatrix(ctx, params, n_ctx)) {
return 1;
if (params.prompt.empty()) {
if (params.in_files.empty()) {
LOG_ERR("Error: No prompt provided and no precomputed matrices (--in-file) to combine.\n");
return 1;
}
LOG_INF("No prompt provided; combining precomputed matrices only.\n");
} else {
if (!compute_imatrix(ctx, params, n_ctx)) {
return 1;
}
}
g_collector.save_imatrix();
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_free(ctx);
llama_free_model(model);
LOG("\n");
llama_perf_context_print(ctx);
llama_backend_free();

2
examples/infill/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-infill)
add_executable(${TARGET} infill.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

2
examples/infill/README.md
View file

@ -14,7 +14,7 @@ In this section, we cover the most commonly used options for running the `infill
- `-m FNAME, --model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.bin`).
- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
- `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference.
- `--spm-infill`: Use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this.
## Input Prompts

328
examples/infill/infill.cpp
View file

@ -1,6 +1,8 @@
#include "arg.h"
#include "common.h"
#include "console.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cassert>
@ -33,58 +35,14 @@
static llama_context ** g_ctx;
static llama_model ** g_model;
static gpt_sampler ** g_smpl;
static gpt_params * g_params;
static common_sampler ** g_smpl;
static common_params * g_params;
static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens
) {
if (params.logdir.empty()) {
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: infill\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Generation Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_perf_dump_yaml(logfile, ctx);
fclose(logfile);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
@ -92,9 +50,13 @@ static void sigint_handler(int signo) {
is_interacting = true;
} else {
console::cleanup();
printf("\n");
gpt_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
LOG("\n");
common_perf_print(*g_ctx, *g_smpl);
// make sure all logs are flushed
LOG("Interrupted by user\n");
common_log_pause(common_log_main());
_exit(130);
}
}
@ -102,142 +64,135 @@ static void sigint_handler(int signo) {
#endif
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
g_params = &params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_INFILL);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
return 1;
}
auto & sparams = params.sparams;
common_init();
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("infill", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
auto & sparams = params.sampling;
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
printf("\n************\n");
printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
print_build_info();
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
LOG("%s: llama backend init\n", __func__);
LOG_INF("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = nullptr;
llama_context * ctx = nullptr;
gpt_sampler * smpl = nullptr;
common_sampler * smpl = nullptr;
g_model = &model;
g_ctx = &ctx;
g_smpl = &smpl;
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
llama_init_result llama_init = llama_init_from_gpt_params(params);
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model;
ctx = llama_init.context;
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
LOG_DBG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
}
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
const bool add_bos = llama_add_bos_token(model);
GGML_ASSERT(!llama_add_eos_token(model));
LOG("add_bos: %d\n", add_bos);
const bool add_bos = llama_vocab_get_add_bos(vocab);
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
std::vector<llama_token> embd_inp;
std::vector<llama_token> embd_end;
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
GGML_ASSERT(llama_token_prefix(model) >= 0);
GGML_ASSERT(llama_token_suffix(model) >= 0);
GGML_ASSERT(llama_vocab_fim_pre(vocab) >= 0);
GGML_ASSERT(llama_vocab_fim_suf(vocab) >= 0);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;
if (add_bos) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
const llama_token middle_token = llama_token_middle(model);
const llama_token middle_token = llama_vocab_fim_mid(vocab);
if (middle_token >= 0) {
embd_inp.push_back(middle_token);
}
LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_DBG("add_bos: %d\n", add_bos);
LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(model));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
embd_inp.push_back(llama_vocab_bos(vocab));
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
}
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
return 1;
}
@ -246,9 +201,8 @@ int main(int argc, char ** argv) {
params.n_keep = (int)embd_inp.size();
}
LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
@ -256,21 +210,21 @@ int main(int argc, char ** argv) {
}
if (params.verbose_prompt) {
LOG_TEE("\n");
LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
LOG_INF("\n");
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (params.n_keep > 0) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
LOG_CNT("'\n");
}
LOG_TEE("\n");
LOG_INF("\n");
}
if (params.interactive) {
@ -287,30 +241,30 @@ int main(int argc, char ** argv) {
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
LOG_TEE("%s: interactive mode on.\n", __func__);
LOG_INF("%s: interactive mode on.\n", __func__);
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
LOG_INF("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
}
}
LOG_TEE("sampling: \n%s\n", sparams.print().c_str());
LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
LOG_TEE("\n\n");
smpl = common_sampler_init(model, sparams);
LOG_TEE("\n##### Infill mode #####\n\n");
if (params.infill) {
printf("\n************\n");
printf("no need to specify '--infill', always running infill\n");
printf("************\n\n");
}
LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());
LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
LOG_INF("\n");
LOG_INF("\n##### Infill mode #####\n\n");
if (params.interactive) {
const char *control_message;
if (params.multiline_input) {
@ -321,11 +275,11 @@ int main(int argc, char ** argv) {
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
LOG_TEE("== Running in interactive mode. ==\n");
LOG_INF("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
LOG_INF( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
LOG_INF( "%s\n", control_message);
is_interacting = params.interactive_first;
}
@ -345,8 +299,6 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
smpl = gpt_sampler_init(model, sparams);
while (n_remain != 0 || params.interactive) {
// predict
if (!embd.empty()) {
@ -360,9 +312,8 @@ int main(int argc, char ** argv) {
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
// infinite text generation via context swapping
@ -371,14 +322,14 @@ int main(int argc, char ** argv) {
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() > n_ctx) {
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep - 1;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
@ -386,9 +337,9 @@ int main(int argc, char ** argv) {
n_past -= n_discard;
LOG("after swap: n_past = %d\n", n_past);
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
}
@ -400,16 +351,16 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
LOG_DBG("n_past = %d\n", n_past);
}
}
@ -417,11 +368,11 @@ int main(int argc, char ** argv) {
embd.clear();
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
const llama_token id = common_sampler_sample(smpl, ctx, -1);
gpt_sampler_accept(smpl, id, true);
common_sampler_accept(smpl, id, true);
// LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(id);
@ -431,16 +382,16 @@ int main(int argc, char ** argv) {
// decrement remaining sampling budget
--n_remain;
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
while ((int) embd_inp.size() > n_consumed) {
embd.push_back(embd_inp[n_consumed]);
// push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules
gpt_sampler_accept(smpl, embd_inp[n_consumed], false);
common_sampler_accept(smpl, embd_inp[n_consumed], false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
@ -452,8 +403,8 @@ int main(int argc, char ** argv) {
// display text
if (input_echo) {
for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id);
printf("%s", token_str.c_str());
const std::string token_str = common_token_to_piece(ctx, id);
LOG("%s", token_str.c_str());
if (embd.size() > 1) {
input_tokens.push_back(id);
@ -462,7 +413,6 @@ int main(int argc, char ** argv) {
output_ss << token_str;
}
}
fflush(stdout);
}
// reset color to default if we there is no pending user input
if (input_echo && (int) embd_inp.size() == n_consumed) {
@ -472,13 +422,12 @@ int main(int argc, char ** argv) {
// if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) {
// deal with eot token in infill mode
if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
if ((common_sampler_last(smpl) == llama_vocab_eot(vocab) || is_interacting) && params.interactive){
if (is_interacting && !params.interactive_first) {
// print an eot token
printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
fflush(stdout);
printf("\n");
LOG("\n");
console::set_display(console::user_input);
std::string buffer;
std::string line;
@ -513,16 +462,16 @@ int main(int argc, char ** argv) {
}
// tokenize new prefix and suffix
std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
inp_pfx.insert(inp_pfx.begin(), llama_vocab_fim_pre(vocab));
inp_sfx.insert(inp_sfx.begin(), llama_vocab_fim_suf(vocab));
embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
embd_end = params.spm_infill ? inp_pfx : inp_sfx;
if (add_bos) {
embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
embd_inp.insert(embd_inp.begin(), llama_vocab_bos(vocab));
}
embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
@ -534,35 +483,33 @@ int main(int argc, char ** argv) {
n_remain = params.n_predict;
n_past = 0;
n_consumed = 0;
// LOG_TEE("took new input\n");
is_interacting = false;
}
// deal with end of generation tokens in interactive mode
else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
LOG("found EOS token\n");
else if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
LOG_DBG("found EOS token\n");
if (params.interactive) {
is_interacting = true;
printf("\n");
LOG("\n");
console::set_display(console::user_input);
fflush(stdout);
}
}
if (n_past > 0 && is_interacting && !params.interactive) {
LOG("waiting for user input\n");
LOG_DBG("waiting for user input\n");
if (params.input_prefix_bos) {
LOG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_vocab_bos(vocab));
}
std::string buffer;
if (!params.input_prefix.empty()) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
buffer += params.input_prefix;
printf("%s", buffer.c_str());
LOG("%s", buffer.c_str());
}
std::string line;
@ -580,30 +527,30 @@ int main(int argc, char ** argv) {
if (buffer.length() > 1) {
// append input suffix if any
if (!params.input_suffix.empty()) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
buffer += params.input_suffix;
printf("%s", params.input_suffix.c_str());
LOG("%s", params.input_suffix.c_str());
}
LOG("buffer: '%s'\n", buffer.c_str());
LOG_DBG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
const auto line_inp = ::llama_tokenize(ctx, buffer, false);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
const auto line_inp = common_tokenize(ctx, buffer, false);
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i];
output_tokens.push_back(token);
output_ss << llama_token_to_piece(ctx, token);
output_ss << common_token_to_piece(ctx, token);
}
n_remain -= line_inp.size();
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
LOG_DBG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
@ -611,14 +558,14 @@ int main(int argc, char ** argv) {
if (n_past > 0) {
if (is_interacting) {
gpt_sampler_reset(smpl);
common_sampler_reset(smpl);
}
is_interacting = false;
}
}
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !params.interactive) {
break;
}
@ -630,23 +577,14 @@ int main(int argc, char ** argv) {
}
}
if (!params.interactive && n_remain <= 0) {
printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
fflush(stdout);
LOG("%s", common_token_to_piece(ctx, llama_vocab_eot(vocab)).c_str());
}
LOG_TEE("\n");
gpt_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
LOG("\n");
common_perf_print(ctx, smpl);
llama_free(ctx);
llama_free_model(model);
gpt_sampler_free(smpl);
common_sampler_free(smpl);
llama_backend_free();
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
return 0;
}

2
examples/json_schema_to_grammar.py
View file

@ -540,7 +540,7 @@ class SchemaConverter:
return self._add_rule(
name,
to_rule(transform()) if self._raw_pattern \
else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space")
def _resolve_ref(self, ref):

2
examples/llama-bench/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-bench)
add_executable(${TARGET} llama-bench.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

952
examples/llama-bench/llama-bench.cpp
View file

File diff suppressed because it is too large Load diff

2
examples/llama.android/llama/build.gradle.kts
View file

@ -18,6 +18,8 @@ android {
}
externalNativeBuild {
cmake {
arguments += "-DLLAMA_BUILD_COMMON=ON"
arguments += "-DGGML_LLAMAFILE=OFF"
arguments += "-DCMAKE_BUILD_TYPE=Release"
cppFlags += listOf()
arguments += listOf()

40
examples/llama.android/llama/src/main/cpp/llama-android.cpp
View file

@ -87,7 +87,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
auto path_to_model = env->GetStringUTFChars(filename, 0);
LOGi("Loading model from %s", path_to_model);
auto model = llama_load_model_from_file(path_to_model, model_params);
auto model = llama_model_load_from_file(path_to_model, model_params);
env->ReleaseStringUTFChars(filename, path_to_model);
if (!model) {
@ -102,7 +102,7 @@ Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring fi
extern "C"
JNIEXPORT void JNICALL
Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
llama_free_model(reinterpret_cast<llama_model *>(model));
llama_model_free(reinterpret_cast<llama_model *>(model));
}
extern "C"
@ -186,11 +186,11 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
for (nri = 0; nri < nr; nri++) {
LOGi("Benchmark prompt processing (pp)");
llama_batch_clear(*batch);
common_batch_clear(*batch);
const int n_tokens = pp;
for (i = 0; i < n_tokens; i++) {
llama_batch_add(*batch, 0, i, { 0 }, false);
common_batch_add(*batch, 0, i, { 0 }, false);
}
batch->logits[batch->n_tokens - 1] = true;
@ -210,9 +210,9 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
const auto t_tg_start = ggml_time_us();
for (i = 0; i < tg; i++) {
llama_batch_clear(*batch);
common_batch_clear(*batch);
for (j = 0; j < pl; j++) {
llama_batch_add(*batch, 0, i, { j }, true);
common_batch_add(*batch, 0, i, { j }, true);
}
LOGi("llama_decode() text generation: %d", i);
@ -283,9 +283,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
nullptr,
nullptr,
nullptr,
0,
0,
0,
};
if (embd) {
@ -308,7 +305,9 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
extern "C"
JNIEXPORT void JNICALL
Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
//llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
delete batch;
}
extern "C"
@ -348,6 +347,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
jlong context_pointer,
jlong batch_pointer,
jstring jtext,
jboolean format_chat,
jint n_len
) {
@ -357,7 +357,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
const auto context = reinterpret_cast<llama_context *>(context_pointer);
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
const auto tokens_list = llama_tokenize(context, text, 1);
bool parse_special = (format_chat == JNI_TRUE);
const auto tokens_list = common_tokenize(context, text, true, parse_special);
auto n_ctx = llama_n_ctx(context);
auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@ -369,14 +370,14 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
}
for (auto id : tokens_list) {
LOGi("%s", llama_token_to_piece(context, id).c_str());
LOGi("token: `%s`-> %d ", common_token_to_piece(context, id).c_str(), id);
}
llama_batch_clear(*batch);
common_batch_clear(*batch);
// evaluate the initial prompt
for (auto i = 0; i < tokens_list.size(); i++) {
llama_batch_add(*batch, tokens_list[i], i, { 0 }, false);
common_batch_add(*batch, tokens_list[i], i, { 0 }, false);
}
// llama_decode will output logits only for the last token of the prompt
@ -406,6 +407,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
const auto model = llama_get_model(context);
const auto vocab = llama_model_get_vocab(model);
if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
@ -414,14 +416,12 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
// sample the most likely token
const auto new_token_id = llama_sampler_sample(sampler, context, -1);
llama_sampler_accept(sampler, new_token_id);
const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
return nullptr;
}
auto new_token_chars = llama_token_to_piece(context, new_token_id);
auto new_token_chars = common_token_to_piece(context, new_token_id);
cached_token_chars += new_token_chars;
jstring new_token = nullptr;
@ -433,8 +433,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
new_token = env->NewStringUTF("");
}
llama_batch_clear(*batch);
llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
common_batch_clear(*batch);
common_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
env->CallVoidMethod(intvar_ncur, la_int_var_inc);

5
examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt
View file

@ -65,6 +65,7 @@ class LLamaAndroid {
context: Long,
batch: Long,
text: String,
formatChat: Boolean,
nLen: Int
): Int
@ -115,10 +116,10 @@ class LLamaAndroid {
}
}
fun send(message: String): Flow<String> = flow {
fun send(message: String, formatChat: Boolean = false): Flow<String> = flow {
when (val state = threadLocalState.get()) {
is State.Loaded -> {
val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
val ncur = IntVar(completion_init(state.context, state.batch, message, formatChat, nlen))
while (ncur.value <= nlen) {
val str = completion_loop(state.context, state.batch, state.sampler, nlen, ncur)
if (str == null) {

17
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View file

@ -46,15 +46,14 @@ actor LlamaContext {
let sparams = llama_sampler_chain_default_params()
self.sampling = llama_sampler_chain_init(sparams)
llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
}
deinit {
llama_sampler_free(sampling)
llama_batch_free(batch)
llama_model_free(model)
llama_free(context)
llama_free_model(model)
llama_backend_free()
}
@ -66,7 +65,7 @@ actor LlamaContext {
model_params.n_gpu_layers = 0
print("Running on simulator, force use n_gpu_layers = 0")
#endif
let model = llama_load_model_from_file(path, model_params)
let model = llama_model_load_from_file(path, model_params)
guard let model else {
print("Could not load model at \(path)")
throw LlamaError.couldNotInitializeContext
@ -152,9 +151,7 @@ actor LlamaContext {
new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
llama_sampler_accept(sampling, new_token_id)
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
if llama_vocab_is_eog(model, new_token_id) || n_cur == n_len {
print("\n")
is_done = true
let new_token_str = String(cString: temporary_invalid_cchars + [0])
@ -213,20 +210,20 @@ actor LlamaContext {
llama_kv_cache_clear(context)
let t_pp_start = ggml_time_us()
let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
if llama_decode(context, batch) != 0 {
print("llama_decode() failed during prompt")
}
llama_synchronize(context)
let t_pp_end = ggml_time_us()
let t_pp_end = DispatchTime.now().uptimeNanoseconds / 1000;
// bench text generation
llama_kv_cache_clear(context)
let t_tg_start = ggml_time_us()
let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
for i in 0..<tg {
llama_batch_clear(&batch)
@ -241,7 +238,7 @@ actor LlamaContext {
llama_synchronize(context)
}
let t_tg_end = ggml_time_us()
let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
llama_kv_cache_clear(context)

8
examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj
View file

@ -7,6 +7,7 @@
objects = {
/* Begin PBXBuildFile section */
1809696D2D05A39F00400EE8 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = 1809696C2D05A39F00400EE8 /* llama */; };
549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; };
7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
@ -17,7 +18,6 @@
8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
/* End PBXBuildFile section */
@ -42,7 +42,7 @@
isa = PBXFrameworksBuildPhase;
buildActionMask = 2147483647;
files = (
DF810E132B4A5BA200301144 /* llama in Frameworks */,
1809696D2D05A39F00400EE8 /* llama in Frameworks */,
549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
);
@ -151,7 +151,7 @@
);
name = llama.swiftui;
packageProductDependencies = (
DF810E122B4A5BA200301144 /* llama */,
1809696C2D05A39F00400EE8 /* llama */,
);
productName = llama.swiftui;
productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@ -429,7 +429,7 @@
/* End XCConfigurationList section */
/* Begin XCSwiftPackageProductDependency section */
DF810E122B4A5BA200301144 /* llama */ = {
1809696C2D05A39F00400EE8 /* llama */ = {
isa = XCSwiftPackageProductDependency;
productName = llama;
};

898
examples/llama.vim
View file

@ -1,135 +1,783 @@
" Requires an already running llama.cpp server
" To install either copy or symlink to ~/.vim/autoload/llama.vim
" Then start with either :call llama#doLlamaGen(),
" or add a keybind to your vimrc such as
" nnoremap Z :call llama#doLlamaGen()<CR>
" Similarly, you could add an insert mode keybind with
" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
" LLM-based text completion using llama.cpp
"
" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc
" let g:llama_api_url = "192.168.1.10:8080"
" llama_overrides can also be set through buffer/window scopes. For instance
" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
" Could be added to your .vimrc to automatically set a lower temperature when
" editing a python script
" Additionally, an override dict can be stored at the top of a file
" !*{"stop": ["User:"]}
" Could be added to the start of your chatlog.txt to set the stopping token
" These parameter dicts are merged together from lowest to highest priority:
" server default -> g:llama_overrides -> w:llama_overrides ->
" b:llama_overrides -> in file (!*) overrides
" requires:
"
" - neovim or vim
" - curl
" - llama.cpp server instance
" - FIM-compatible model
"
" sample config:
"
" - Tab - accept the current suggestion
" - Shift+Tab - accept just the first line of the suggestion
" - Ctrl+F - toggle FIM completion manually
"
" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
"
" start the llama.cpp server with a FIM-compatible model. for example:
"
" $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
"
" --batch-size [512, model max context]
"
" adjust the batch size to control how much of the provided local context will be used during the inference
" lower values will use smaller part of the context around the cursor, which will result in faster processing
"
" --ubatch-size [64, 2048]
"
" chunks the batch into smaller chunks for faster processing
" depends on the specific hardware. use llama-bench to profile and determine the best size
"
" --cache-reuse (ge:llama_config.n_predict, 1024]
"
" this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
" using non-zero value enables context reuse on the server side which dramatically improves the performance at
" large contexts. a value of 256 should be good for all cases
"
" run this once to initialise llama.vim:
"
" :call llama#init()
"
" more info: https://github.com/ggerganov/llama.cpp/pull/9787
"
" Sublists (like logit_bias and stop) are overridden, not merged
" Example override:
" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
if !exists("g:llama_api_url")
let g:llama_api_url= "127.0.0.1:8080"
endif
if !exists("g:llama_overrides")
let g:llama_overrides = {}
endif
const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
let s:linedict = {}
func s:callbackHandler(bufn, channel, msg)
if len(a:msg) < 3
return
elseif a:msg[0] == "d"
let l:msg = a:msg[6:-1]
else
let l:msg = a:msg
endif
let l:decoded_msg = json_decode(l:msg)
let l:newtext = split(l:decoded_msg['content'], "\n", 1)
if len(l:newtext) > 0
call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
else
echo "nothing genned"
endif
if len(newtext) > 1
let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
endif
if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
echo "Finished generation"
endif
endfunction
" colors (adjust to your liking)
highlight llama_hl_hint guifg=#ff772f ctermfg=202
highlight llama_hl_info guifg=#77ff2f ctermfg=119
func llama#doLlamaGen()
if exists("b:job")
if job_status(b:job) == "run"
call job_stop(b:job)
return
endif
endif
" general parameters:
"
" endpoint: llama.cpp server endpoint
" n_prefix: number of lines before the cursor location to include in the local prefix
" n_suffix: number of lines after the cursor location to include in the local suffix
" n_predict: max number of tokens to predict
" t_max_prompt_ms: max alloted time for the prompt processing (TODO: not yet supported)
" t_max_predict_ms: max alloted time for the prediction
" show_info: show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
" auto_fim: trigger FIM completion automatically on cursor movement
" max_line_suffix: do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
"
" ring buffer of chunks, accumulated with time upon:
"
" - completion request
" - yank
" - entering a buffer
" - leaving a buffer
" - writing a file
"
" parameters for the ring-buffer with extra context:
"
" ring_n_chunks: max number of chunks to pass as extra context to the server (0 to disable)
" ring_chunk_size: max size of the chunks (in number of lines)
" note: adjust these numbers so that you don't overrun your context
" at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
" ring_scope: the range around the cursor position (in number of lines) for gathering chunks after FIM
" ring_update_ms: how often to process queued chunks in normal mode
"
let s:default_config = {
\ 'endpoint': 'http://127.0.0.1:8012/infill',
\ 'n_prefix': 256,
\ 'n_suffix': 64,
\ 'n_predict': 128,
\ 't_max_prompt_ms': 500,
\ 't_max_predict_ms': 3000,
\ 'show_info': 2,
\ 'auto_fim': v:true,
\ 'max_line_suffix': 8,
\ 'ring_n_chunks': 64,
\ 'ring_chunk_size': 64,
\ 'ring_scope': 1024,
\ 'ring_update_ms': 1000,
\ }
let l:cbuffer = bufnr("%")
let s:linedict[l:cbuffer] = line('$')
let l:buflines = getbufline(l:cbuffer, 1, 1000)
let l:querydata = copy(s:querydata)
call extend(l:querydata, g:llama_overrides)
if exists("w:llama_overrides")
call extend(l:querydata, w:llama_overrides)
endif
if exists("b:llama_overrides")
call extend(l:querydata, b:llama_overrides)
endif
if l:buflines[0][0:1] == '!*'
let l:userdata = json_decode(l:buflines[0][2:-1])
call extend(l:querydata, l:userdata)
let l:buflines = l:buflines[1:-1]
endif
let l:querydata.prompt = join(l:buflines, "\n")
let l:curlcommand = copy(s:curlcommand)
if exists("g:llama_api_key")
call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key])
endif
let l:curlcommand[2] = json_encode(l:querydata)
let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
endfunction
let g:llama_config = get(g:, 'llama_config', s:default_config)
" Echos the tokkenization of the provided string , or cursor to end of word
" Onus is placed on the user to include the preceding space
func llama#tokenizeWord(...)
if (a:0 > 0)
let l:input = a:1
else
exe "normal \"*ye"
let l:input = @*
endif
let l:querydata = {"content": l:input}
let l:curlcommand = copy(s:curlcommand)
let l:curlcommand[2] = json_encode(l:querydata)
let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
endfunction
func s:tokenizeWordCallback(plaintext, channel, msg)
echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
endfunction
" Echos the token count of the entire buffer (or provided string)
" Example usage :echo llama#tokenCount()
func llama#tokenCount(...)
if (a:0 > 0)
let l:buflines = a:1
else
let l:buflines = getline(1,1000)
if l:buflines[0][0:1] == '!*'
let l:buflines = l:buflines[1:-1]
function! s:get_indent(str)
let l:count = 0
for i in range(len(a:str))
if a:str[i] == "\t"
let l:count += &tabstop - 1
else
break
endif
let l:buflines = join(l:buflines, "\n")
endif
let l:querydata = {"content": l:buflines}
let l:curlcommand = copy(s:curlcommand)
let l:curlcommand[2] = json_encode(l:querydata)
let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
endfor
return l:count
endfunction
func s:tokenCountCallback(channel, msg)
let resp = json_decode(a:msg)
echo len(resp.tokens)
function! s:rand(i0, i1) abort
return a:i0 + rand() % (a:i1 - a:i0 + 1)
endfunction
function! llama#init()
if !executable('curl')
echohl WarningMsg
echo 'llama.vim requires the "curl" command to be available'
echohl None
return
endif
let s:pos_x = 0 " cursor position upon start of completion
let s:pos_y = 0
let s:line_cur = ''
let s:line_cur_prefix = ''
let s:line_cur_suffix = ''
let s:ring_chunks = [] " current set of chunks used as extra context
let s:ring_queued = [] " chunks that are queued to be sent for processing
let s:ring_n_evict = 0
let s:hint_shown = v:false
let s:pos_y_pick = -9999 " last y where we picked a chunk
let s:pos_dx = 0
let s:content = []
let s:can_accept = v:false
let s:timer_fim = -1
let s:t_fim_start = reltime() " used to measure total FIM time
let s:t_last_move = reltime() " last time the cursor moved
let s:current_job = v:null
let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
let s:ghost_text_vim = has('textprop')
if s:ghost_text_vim
let s:hlgroup_hint = 'llama_hl_hint'
let s:hlgroup_info = 'llama_hl_info'
if empty(prop_type_get(s:hlgroup_hint))
call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
endif
if empty(prop_type_get(s:hlgroup_info))
call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
endif
endif
augroup llama
autocmd!
autocmd InsertEnter * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
autocmd InsertLeavePre * call llama#fim_cancel()
autocmd CursorMoved * call s:on_move()
autocmd CursorMovedI * call s:on_move()
autocmd CompleteChanged * call llama#fim_cancel()
if g:llama_config.auto_fim
autocmd CursorMovedI * call llama#fim(v:true)
endif
" gather chunks upon yanking
autocmd TextYankPost * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
" gather chunks upon entering/leaving a buffer
autocmd BufEnter * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
autocmd BufLeave * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
" gather chunk upon saving the file
autocmd BufWritePost * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
augroup END
silent! call llama#fim_cancel()
" init background update of the ring buffer
if g:llama_config.ring_n_chunks > 0
call s:ring_update()
endif
endfunction
" compute how similar two chunks of text are
" 0 - no similarity, 1 - high similarity
" TODO: figure out something better
function! s:chunk_sim(c0, c1)
let l:lines0 = len(a:c0)
let l:lines1 = len(a:c1)
let l:common = 0
for l:line0 in a:c0
for l:line1 in a:c1
if l:line0 == l:line1
let l:common += 1
break
endif
endfor
endfor
return 2.0 * l:common / (l:lines0 + l:lines1)
endfunction
" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
"
" no_mod - do not pick chunks from buffers with pending changes
" do_evict - evict chunks that are very similar to the new one
"
function! s:pick_chunk(text, no_mod, do_evict)
" do not pick chunks from buffers with pending changes or buffers that are not files
if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
return
endif
" if the extra context option is disabled - do nothing
if g:llama_config.ring_n_chunks <= 0
return
endif
" don't pick very small chunks
if len(a:text) < 3
return
endif
if len(a:text) + 1 < g:llama_config.ring_chunk_size
let l:chunk = a:text
else
let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
let l:chunk = a:text[l:l0:l:l1]
endif
let l:chunk_str = join(l:chunk, "\n") . "\n"
" check if this chunk is already added
let l:exist = v:false
for i in range(len(s:ring_chunks))
if s:ring_chunks[i].data == l:chunk
let l:exist = v:true
break
endif
endfor
for i in range(len(s:ring_queued))
if s:ring_queued[i].data == l:chunk
let l:exist = v:true
break
endif
endfor
if l:exist
return
endif
" evict queued chunks that are very similar to the new one
for i in range(len(s:ring_queued) - 1, 0, -1)
if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
if a:do_evict
call remove(s:ring_queued, i)
let s:ring_n_evict += 1
else
return
endif
endif
endfor
" also from s:ring_chunks
for i in range(len(s:ring_chunks) - 1, 0, -1)
if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
if a:do_evict
call remove(s:ring_chunks, i)
let s:ring_n_evict += 1
else
return
endif
endif
endfor
" TODO: become parameter ?
if len(s:ring_queued) == 16
call remove(s:ring_queued, 0)
endif
call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
"let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
endfunction
" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
" called every g:llama_config.ring_update_ms
function! s:ring_update()
call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
" update only if in normal mode or if the cursor hasn't moved for a while
if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
return
endif
if len(s:ring_queued) == 0
return
endif
" move the first queued chunk to the ring buffer
if len(s:ring_chunks) == g:llama_config.ring_n_chunks
call remove(s:ring_chunks, 0)
endif
call add(s:ring_chunks, remove(s:ring_queued, 0))
"let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
" send asynchronous job with the new extra context so that it is ready for the next FIM
let l:extra_context = []
for l:chunk in s:ring_chunks
call add(l:extra_context, {
\ 'text': l:chunk.str,
\ 'time': l:chunk.time,
\ 'filename': l:chunk.filename
\ })
endfor
" no samplers needed here
let l:request = json_encode({
\ 'input_prefix': "",
\ 'input_suffix': "",
\ 'input_extra': l:extra_context,
\ 'prompt': "",
\ 'n_predict': 1,
\ 'temperature': 0.0,
\ 'stream': v:false,
\ 'samplers': ["temperature"],
\ 'cache_prompt': v:true,
\ 't_max_prompt_ms': 1,
\ 't_max_predict_ms': 1
\ })
let l:curl_command = [
\ "curl",
\ "--silent",
\ "--no-buffer",
\ "--request", "POST",
\ "--url", g:llama_config.endpoint,
\ "--header", "Content-Type: application/json",
\ "--data", l:request
\ ]
" no callbacks because we don't need to process the response
if s:ghost_text_nvim
call jobstart(l:curl_command, {})
elseif s:ghost_text_vim
call job_start(l:curl_command, {})
endif
endfunction
" necessary for 'inoremap <expr>'
function! llama#fim_inline(is_auto) abort
call llama#fim(a:is_auto)
return ''
endfunction
" the main FIM call
" takes local context around the cursor and sends it together with the extra context to the server for completion
function! llama#fim(is_auto) abort
" we already have a suggestion for the current cursor position
if s:hint_shown && !a:is_auto
call llama#fim_cancel()
return
endif
call llama#fim_cancel()
" avoid sending repeated requests too fast
if reltimefloat(reltime(s:t_fim_start)) < 0.6
if s:timer_fim != -1
call timer_stop(s:timer_fim)
let s:timer_fim = -1
endif
let s:t_fim_start = reltime()
let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
return
endif
let s:t_fim_start = reltime()
let s:content = []
let s:can_accept = v:false
let s:pos_x = col('.') - 1
let s:pos_y = line('.')
let l:max_y = line('$')
let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
let s:line_cur = getline('.')
let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
return
endif
let l:prefix = ""
\ . join(l:lines_prefix, "\n")
\ . "\n"
let l:prompt = ""
\ . s:line_cur_prefix
let l:suffix = ""
\ . s:line_cur_suffix
\ . "\n"
\ . join(l:lines_suffix, "\n")
\ . "\n"
" prepare the extra context data
let l:extra_context = []
for l:chunk in s:ring_chunks
call add(l:extra_context, {
\ 'text': l:chunk.str,
\ 'time': l:chunk.time,
\ 'filename': l:chunk.filename
\ })
endfor
" the indentation of the current line
let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
let l:request = json_encode({
\ 'input_prefix': l:prefix,
\ 'input_suffix': l:suffix,
\ 'input_extra': l:extra_context,
\ 'prompt': l:prompt,
\ 'n_predict': g:llama_config.n_predict,
\ 'n_indent': l:indent,
\ 'top_k': 40,
\ 'top_p': 0.99,
\ 'stream': v:false,
\ 'samplers': ["top_k", "top_p", "infill"],
\ 'cache_prompt': v:true,
\ 't_max_prompt_ms': g:llama_config.t_max_prompt_ms,
\ 't_max_predict_ms': g:llama_config.t_max_predict_ms
\ })
let l:curl_command = [
\ "curl",
\ "--silent",
\ "--no-buffer",
\ "--request", "POST",
\ "--url", g:llama_config.endpoint,
\ "--header", "Content-Type: application/json",
\ "--data", l:request
\ ]
if s:current_job != v:null
if s:ghost_text_nvim
call jobstop(s:current_job)
elseif s:ghost_text_vim
call job_stop(s:current_job)
endif
endif
" send the request asynchronously
if s:ghost_text_nvim
let s:current_job = jobstart(l:curl_command, {
\ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
\ 'on_exit': function('s:fim_on_exit'),
\ 'stdout_buffered': v:true
\ })
elseif s:ghost_text_vim
let s:current_job = job_start(l:curl_command, {
\ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
\ 'exit_cb': function('s:fim_on_exit')
\ })
endif
" TODO: per-file location
let l:delta_y = abs(s:pos_y - s:pos_y_pick)
" gather some extra context nearby and process it in the background
" only gather chunks if the cursor has moved a lot
" TODO: something more clever? reranking?
if a:is_auto && l:delta_y > 32
" expand the prefix even further
call s:pick_chunk(getline(max([1, s:pos_y - g:llama_config.ring_scope]), max([1, s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
" pick a suffix chunk
call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]), min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
let s:pos_y_pick = s:pos_y
endif
endfunction
" if first_line == v:true accept only the first line of the response
function! llama#fim_accept(first_line)
" insert the suggestion at the cursor location
if s:can_accept && len(s:content) > 0
call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
if len(s:content) > 1
if !a:first_line
call append(s:pos_y, s:content[1:-1])
endif
endif
" move the cursor to the end of the accepted text
if !a:first_line && len(s:content) > 1
call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
else
call cursor(s:pos_y, s:pos_x + len(s:content[0]))
endif
endif
call llama#fim_cancel()
endfunction
function! llama#fim_cancel()
let s:hint_shown = v:false
" clear the virtual text
let l:bufnr = bufnr('%')
if s:ghost_text_nvim
let l:id_vt_fim = nvim_create_namespace('vt_fim')
call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim, 0, -1)
elseif s:ghost_text_vim
call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
call prop_remove({'type': s:hlgroup_info, 'all': v:true})
endif
" remove the mappings
silent! iunmap <buffer> <Tab>
silent! iunmap <buffer> <S-Tab>
silent! iunmap <buffer> <Esc>
endfunction
function! s:on_move()
let s:t_last_move = reltime()
call llama#fim_cancel()
endfunction
" callback that processes the FIM result from the server and displays the suggestion
function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
if s:ghost_text_nvim
let l:raw = join(a:data, "\n")
elseif s:ghost_text_vim
let l:raw = a:data
endif
if len(l:raw) == 0
return
endif
if a:pos_x != col('.') - 1 || a:pos_y != line('.')
return
endif
" show the suggestion only in insert mode
if mode() !=# 'i'
return
endif
let s:pos_x = a:pos_x
let s:pos_y = a:pos_y
let s:can_accept = v:true
let l:has_info = v:false
if s:can_accept && v:shell_error
if !a:is_auto
call add(s:content, "<| curl error: is the server on? |>")
endif
let s:can_accept = v:false
endif
let l:n_prompt = 0
let l:t_prompt_ms = 1.0
let l:s_prompt = 0
let l:n_predict = 0
let l:t_predict_ms = 1.0
let l:s_predict = 0
" get the generated suggestion
if s:can_accept
let l:response = json_decode(l:raw)
for l:part in split(get(l:response, 'content', ''), "\n", 1)
call add(s:content, l:part)
endfor
" remove trailing new lines
while len(s:content) > 0 && s:content[-1] == ""
call remove(s:content, -1)
endwhile
let l:generation_settings = get(l:response, 'generation_settings', {})
let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
let l:n_cached = get(l:response, 'tokens_cached', 0)
let l:truncated = get(l:response, 'truncated', v:false)
" if response.timings is available
if len(get(l:response, 'timings', {})) > 0
let l:has_info = v:true
let l:timings = get(l:response, 'timings', {})
let l:n_prompt = get(l:timings, 'prompt_n', 0)
let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
let l:s_prompt = get(l:timings, 'prompt_per_second', 0)
let l:n_predict = get(l:timings, 'predicted_n', 0)
let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
let l:s_predict = get(l:timings, 'predicted_per_second', 0)
endif
endif
if len(s:content) == 0
call add(s:content, "")
let s:can_accept = v:false
endif
if len(s:content) == 0
return
endif
" NOTE: the following is logic for discarding predictions that repeat existing text
" the code is quite ugly and there is very likely a simpler and more canonical way to implement this
"
" still, I wonder if there is some better way that avoids having to do these special hacks?
" on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
" start generating whatever we have given it via the extra context. but on the other hand, it's not very
" helpful to re-generate the same code that is already there
" truncate the suggestion if the first line is empty
if len(s:content) == 1 && s:content[0] == ""
let s:content = [""]
endif
" ... and the next lines are repeated
if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
let s:content = [""]
endif
" truncate the suggestion if it repeats the suffix
if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
let s:content = [""]
endif
" find the first non-empty line (strip whitespace)
let l:cmp_y = s:pos_y + 1
while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
let l:cmp_y += 1
endwhile
if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
" truncate the suggestion if it repeats the next line
if len(s:content) == 1
let s:content = [""]
endif
" ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
let s:content = [""]
endif
" ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
let s:content = [""]
endif
endif
" keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
"let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
"for i in range(1, len(s:content) - 1)
" if strlen(matchstr(s:content[i], '^\s*')) < l:indent
" let s:content = s:content[:i - 1]
" break
" endif
"endfor
let s:pos_dx = len(s:content[-1])
let s:content[-1] .= s:line_cur_suffix
call llama#fim_cancel()
" display virtual text with the suggestion
let l:bufnr = bufnr('%')
if s:ghost_text_nvim
let l:id_vt_fim = nvim_create_namespace('vt_fim')
endif
" construct the info message
if g:llama_config.show_info > 0 && l:has_info
let l:prefix = ' '
if l:truncated
let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
\ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
\ l:n_cached, l:n_ctx
\ )
else
let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
\ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
\ l:n_cached, l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
\ l:n_prompt, l:t_prompt_ms, l:s_prompt,
\ l:n_predict, l:t_predict_ms, l:s_predict,
\ 1000.0 * reltimefloat(reltime(s:t_fim_start))
\ )
endif
if g:llama_config.show_info == 1
" display the info in the statusline
let &statusline = l:info
let l:info = ''
endif
endif
" display the suggestion and append the info to the end of the first line
if s:ghost_text_nvim
call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
\ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
\ 'virt_text_win_col': virtcol('.') - 1
\ })
call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
\ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
\ 'virt_text_win_col': virtcol('.')
\ })
elseif s:ghost_text_vim
let l:new_suffix = s:content[0]
if !empty(l:new_suffix)
call prop_add(s:pos_y, s:pos_x + 1, {
\ 'type': s:hlgroup_hint,
\ 'text': l:new_suffix
\ })
endif
for line in s:content[1:]
call prop_add(s:pos_y, 0, {
\ 'type': s:hlgroup_hint,
\ 'text': line,
\ 'text_padding_left': s:get_indent(line),
\ 'text_align': 'below'
\ })
endfor
if !empty(l:info)
call prop_add(s:pos_y, 0, {
\ 'type': s:hlgroup_info,
\ 'text': l:info,
\ 'text_padding_left': col('$'),
\ 'text_wrap': 'truncate'
\ })
endif
endif
" setup accept shortcuts
inoremap <buffer> <Tab> <C-O>:call llama#fim_accept(v:false)<CR>
inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
let s:hint_shown = v:true
endfunction
function! s:fim_on_exit(job_id, exit_code, event = v:null)
if a:exit_code != 0
echom "Job failed with exit code: " . a:exit_code
endif
let s:current_job = v:null
endfunction

13
examples/llava/CMakeLists.txt
View file

@ -11,7 +11,7 @@ target_include_directories(llava PUBLIC .)
target_include_directories(llava PUBLIC ../..)
target_include_directories(llava PUBLIC ../../common)
target_compile_features(llava PRIVATE cxx_std_11)
target_compile_features(llava PRIVATE cxx_std_17)
add_library(llava_static STATIC $<TARGET_OBJECTS:llava>)
if (BUILD_SHARED_LIBS)
@ -35,11 +35,18 @@ add_executable(${TARGET} llava-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-minicpmv-cli)
add_executable(${TARGET} minicpmv-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-qwen2vl-cli)
add_executable(${TARGET} qwen2vl-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)

10
examples/llava/MobileVLM-README.md
View file

@ -39,7 +39,7 @@ python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
```sh
python ./examples/llava/convert_image_encoder_to_gguf \
python ./examples/llava/convert_image_encoder_to_gguf.py \
-m path/to/clip-vit-large-patch14-336 \
--llava-projector path/to/MobileVLM-1.7B/llava.projector \
--output-dir path/to/MobileVLM-1.7B \
@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf \
```
```sh
python ./examples/llava/convert_image_encoder_to_gguf \
python ./examples/llava/convert_image_encoder_to_gguf.py \
-m path/to/clip-vit-large-patch14-336 \
--llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
--output-dir path/to/MobileVLM-1.7B_V2 \
@ -57,12 +57,12 @@ python ./examples/llava/convert_image_encoder_to_gguf \
4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
```sh
python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B
python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B --skip-unknown
```
5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
5. Use `quantize` to convert LLaMA part's DataType from `fp32` to `q4_k`
```sh
./llama-quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
./llama-quantize path/to/MobileVLM-1.7B/ggml-model-F32.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
```
Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.

46
examples/llava/README-minicpmo2.6.md Normal file
View file

@ -0,0 +1,46 @@
## MiniCPM-o 2.6
Currently, this readme only supports minicpm-omni's image capabilities, and we will update the full-mode support as soon as possible.
### Prepare models and code
Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch model from huggingface to "MiniCPM-o-2_6" folder.
Clone llama.cpp:
```bash
git clone git@github.com:OpenBMB/llama.cpp.git
cd llama.cpp
git checkout minicpm-omni
```
### Usage of MiniCPM-o 2.6
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us)
```bash
python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-o-2_6
python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-o-2_6 --minicpmv-projector ../MiniCPM-o-2_6/minicpmv.projector --output-dir ../MiniCPM-o-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 4
python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
# quantize int4 version
./llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
```
Build llama.cpp using `CMake`:
https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md
```bash
cmake -B build
cmake --build build --config Release
```
Inference on Linux or Mac
```
# run f16 version
./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# run quantized int4 version
./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# or run in interactive mode
./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
```

500
examples/llava/clip.cpp
View file

File diff suppressed because it is too large Load diff

10
examples/llava/clip.h
View file

@ -45,6 +45,7 @@ CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity
CLIP_API void clip_free(struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
@ -55,11 +56,13 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
CLIP_API int clip_n_mmproj_embd (const struct clip_ctx * ctx);
CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip);
CLIP_API struct clip_image_size * clip_image_size_init();
CLIP_API struct clip_image_u8 * clip_image_u8_init ();
@ -86,6 +89,9 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
#ifdef __cplusplus
}

4
examples/llava/convert_image_encoder_to_gguf.py
View file

@ -274,7 +274,7 @@ fout.add_bool("clip.use_gelu", use_gelu)
if has_llava_projector:
model.vision_model.encoder.layers.pop(-1) # pyright: ignore[reportAttributeAccessIssue]
model.vision_model.encoder.layers.pop(-1)
projector = torch.load(args.llava_projector)
for name, data in projector.items():
name = get_tensor_name(name)
@ -288,7 +288,7 @@ if has_llava_projector:
print("Projector tensors added\n")
state_dict = model.state_dict() # pyright: ignore[reportAttributeAccessIssue]
state_dict = model.state_dict()
for name, data in state_dict.items():
if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
# we don't need this

132
examples/llava/llava-cli.cpp
View file

@ -1,14 +1,16 @@
#include "ggml.h"
#include "arg.h"
#include "base64.hpp"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "base64.hpp"
#include "ggml.h"
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
@ -18,8 +20,8 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
if (n_eval > n_batch) {
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
@ -35,21 +37,25 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
return true;
}
static const char * sample(struct gpt_sampler * smpl,
static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true);
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = llama_token_to_piece(ctx_llama, id);
ret = common_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past);
return ret.c_str();
@ -74,7 +80,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
size_t img_base64_str_start, img_base64_str_end;
find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
return NULL;
}
@ -88,7 +94,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
if (!embed) {
LOG_TEE("%s: could not load image from base64 string.\n", __func__);
LOG_ERR("%s: could not load image from base64 string.\n", __func__);
return NULL;
}
@ -113,23 +119,23 @@ struct llava_context {
};
static void print_usage(int, char ** argv) {
LOG_TEE("\n example usage:\n");
LOG_TEE("\n %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG_TEE("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
LOG("\n example usage:\n");
LOG("\n %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) {
static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
// load and preprocess the image
llava_image_embed * embed = NULL;
auto prompt = params->prompt;
if (prompt_contains_image(prompt)) {
if (!params->image.empty()) {
LOG_TEE("using base64 encoded image instead of command line image path\n");
LOG_INF("using base64 encoded image instead of command line image path\n");
}
embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
if (!embed) {
LOG_TEE("%s: can't load image from prompt\n", __func__);
LOG_ERR("%s: can't load image from prompt\n", __func__);
return NULL;
}
params->prompt = remove_image_from_prompt(prompt);
@ -144,7 +150,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
return embed;
}
static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) {
static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
int n_past = 0;
const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
@ -155,18 +161,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
// new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<image>").length());
LOG_TEE("system_prompt: %s\n", system_prompt.c_str());
LOG_INF("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
LOG_TEE("user_prompt: %s\n", user_prompt.c_str());
LOG_INF("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
@ -174,9 +180,9 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
user_prompt = prompt + "\nASSISTANT:";
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
@ -187,11 +193,11 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
// generate the response
LOG_TEE("\n");
LOG("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
if (!smpl) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
@ -201,7 +207,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
printf("%s", tmp);
LOG("%s", tmp);
if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
@ -209,25 +215,25 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
fflush(stdout);
}
gpt_sampler_free(smpl);
printf("\n");
common_sampler_free(smpl);
LOG("\n");
}
static struct llama_model * llava_init(gpt_params * params) {
static struct llama_model * llava_init(common_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = llama_model_params_from_gpt_params(*params);
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt;
@ -237,18 +243,17 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return NULL;
}
auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->ctx_clip = ctx_clip;
@ -263,76 +268,65 @@ static void llava_free(struct llava_context * ctx_llava) {
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
int main(int argc, char ** argv) {
ggml_time_init();
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_LLAVA, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("llava", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
common_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
auto model = llava_init(&params);
auto * model = llava_init(&params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
return 1;
}
if (prompt_contains_image(params.prompt)) {
auto ctx_llava = llava_init_context(&params, model);
auto * ctx_llava = llava_init_context(&params, model);
auto image_embed = load_image(ctx_llava, &params, "");
auto * image_embed = load_image(ctx_llava, &params, "");
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
} else {
for (auto & image : params.image) {
auto ctx_llava = llava_init_context(&params, model);
auto * ctx_llava = llava_init_context(&params, model);
auto image_embed = load_image(ctx_llava, &params, image);
auto * image_embed = load_image(ctx_llava, &params, image);
if (!image_embed) {
std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
return 1;
}
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
}
}
llama_free_model(model);
llama_model_free(model);
return 0;
}

152
examples/llava/llava.cpp
View file

@ -1,13 +1,27 @@
#include "clip.h"
#include "common.h"
#include "llama.h"
#include "llava.h"
#include "base64.hpp"
#include "llama.h"
#include <algorithm>
#include <cerrno>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <vector>
#include <numeric>
#if defined(LLAVA_LOG_OFF)
# define LOG_INF(...)
# define LOG_WRN(...)
# define LOG_ERR(...)
# define LOG_DBG(...)
#else // defined(LLAVA_LOG_OFF)
# define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
# define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
# define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
# define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
#endif // defined(LLAVA_LOG_OFF)
// RGB uint8 image
struct clip_image_u8 {
@ -54,7 +68,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int>& ori
int downscaled_height = static_cast<int>(original_height * scale);
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
int wasted_resolution = (width * height) - effective_resolution;
// LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
// LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
max_effective_resolution = effective_resolution;
min_wasted_resolution = wasted_resolution;
@ -184,7 +198,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
// ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
ggml_build_forward_expand(gf, flatten);
ggml_graph_compute_with_ctx(model.ctx, gf, 1);
struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor* result = ggml_graph_node(gf, -1);
memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
// append without newline tokens (default behavior in llava_arch when not using unpad ):
@ -202,7 +216,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
return true;
}
static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) {
int width = image->nx;
int height = image->ny;
int num_patches = (height / patch_size) * (width / patch_size);
@ -236,7 +250,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
LOG_TEE("%s: unable to preprocess image\n", __func__);
LOG_ERR("%s: unable to preprocess image\n", __func__);
delete[] img_res_v.data;
return false;
}
@ -245,39 +259,44 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
if (clip_is_minicpmv(ctx_clip)) {
if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
struct clip_image_size * load_image_size = clip_image_size_init();
for (size_t i = 0; i < img_res_v.size; i++) {
const int64_t t_img_enc_step_start_us = ggml_time_us();
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
int patch_size=14;
load_image_size->width = img_res_v.data[i].nx;
load_image_size->height = img_res_v.data[i].ny;
clip_add_load_image_size(ctx_clip, load_image_size);
bool encoded = false;
int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
if (has_minicpmv_projector == 2) {
encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
}
else if (has_minicpmv_projector == 3) {
if (clip_is_qwen2vl(ctx_clip)) {
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
}
else {
encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
}
if (!encoded) {
LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
const int64_t t_img_enc_steop_batch_us = ggml_time_us();
LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
int n_img_pos_out = 0;
for (size_t i = 0; i < image_embd_v.size(); i++) {
std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
n_img_pos_out += clip_n_patches(ctx_clip);
std::memcpy(
image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
image_embd_v[i],
clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
}
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
@ -287,7 +306,10 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
load_image_size->width = img->nx;
load_image_size->height = img->ny;
clip_add_load_image_size(ctx_clip, load_image_size);
LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
delete[] img_res_v.data;
img_res_v.size = 0;
img_res_v.data = nullptr;
}
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
@ -295,7 +317,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
delete[] img_res_v.data;
if (!encoded) {
LOG_TEE("Unable to encode image\n");
LOG_ERR("Unable to encode image\n");
return false;
}
@ -309,12 +331,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
if (!encoded) {
LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_TEE("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
const int32_t * image_grid = clip_image_grid(ctx_clip);
@ -347,22 +369,22 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
// clip_image_save_to_bmp(*tmp, "image_feature.bmp");
}
LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
const int64_t t_img_enc_end_us = ggml_time_us();
float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
return true;
}
bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
// make sure that the correct mmproj was used, i.e., compare apples to apples
int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
int n_llama_embd = llama_model_n_embd(llama_get_model(ctx_llama));
auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
if (n_image_embd != n_llama_embd) {
LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
return false;
}
return true;
@ -373,15 +395,21 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
if (clip_is_minicpmv(ctx_clip)) {
num_max_patches = 10;
}
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
float * image_embd;
if (clip_is_qwen2vl(ctx_clip)) {
// qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
} else {
image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
}
if (!image_embd) {
LOG_TEE("Unable to allocate memory for image embeddings\n");
LOG_ERR("Unable to allocate memory for image embeddings\n");
return false;
}
int n_img_pos;
if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
LOG_TEE("%s: cannot encode image, aborting\n", __func__);
LOG_ERR("%s: cannot encode image, aborting\n", __func__);
free(image_embd);
return false;
}
@ -391,17 +419,51 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
return true;
}
struct llava_embd_batch {
std::vector<llama_pos> pos;
std::vector<int32_t> n_seq_id;
std::vector<llama_seq_id> seq_id_0;
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
logits .resize(n_tokens);
seq_id_0.resize(1);
seq_id_0[0] = seq_id;
seq_ids [n_tokens] = nullptr;
batch = {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
/*logits =*/ logits.data(),
};
for (int i = 0; i < n_tokens; i++) {
batch.pos [i] = pos_0 + i;
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
};
bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
int n_embd = llama_model_n_embd(llama_get_model(ctx_llama));
for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
int n_eval = image_embed->n_image_pos - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
if (llama_decode(ctx_llama, batch)) {
LOG_TEE("%s : failed to eval\n", __func__);
float * embd = image_embed->embed+i*n_embd;
llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
if (llama_decode(ctx_llama, llava_batch.batch)) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
*n_past += n_eval;
@ -413,7 +475,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
clip_image_u8 * img = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
clip_image_u8_free(img);
LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
return NULL;
}
@ -422,7 +484,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
if (!image_embed_result) {
clip_image_u8_free(img);
LOG_TEE("%s: coulnd't embed the image\n", __func__);
LOG_ERR("%s: couldn't embed the image\n", __func__);
return NULL;
}
@ -436,7 +498,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
auto file = fopen(path, "rb");
if (file == NULL) {
LOG_TEE("%s: can't read file %s\n", __func__, path);
LOG_ERR("%s: can't read file %s\n", __func__, path);
return false;
}
@ -446,7 +508,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
if (buffer == NULL) {
LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
perror("Memory allocation error");
fclose(file);
return false;
@ -454,10 +516,16 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
errno = 0;
size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
if (ferror(file)) {
die_fmt("read error: %s", strerror(errno));
LOG_ERR("read error: %s", strerror(errno));
free(buffer);
fclose(file);
return false;
}
if (ret != (size_t) fileSize) {
die("unexpectedly reached end of file");
LOG_ERR("unexpectedly reached end of file");
free(buffer);
fclose(file);
return false;
}
fclose(file); // Close the file
@ -471,7 +539,7 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
long image_bytes_length;
auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
if (!loaded) {
LOG_TEE("%s: failed to load %s\n", __func__, image_path);
LOG_ERR("%s: failed to load %s\n", __func__, image_path);
return NULL;
}

153
examples/llava/minicpmv-cli.cpp
View file

@ -1,13 +1,18 @@
#include "ggml.h"
#include "arg.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "ggml.h"
#include <algorithm>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
#include <iostream> // TODO: remove me
struct llava_context {
struct clip_ctx * ctx_clip = NULL;
@ -16,53 +21,47 @@ struct llava_context {
};
static void show_additional_info(int /*argc*/, char ** argv) {
LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n");
LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
static struct llama_model * llava_init(gpt_params * params) {
static struct llama_model * llava_init(common_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = llama_model_params_from_gpt_params(*params);
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
llama_context_params ctx_params = common_context_params_to_llama(*params);
if (params->n_ctx < 2048) {
// warn user here, "Image processing requires at least 2048 context, setting context to 2048"
LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
ctx_params.n_ctx = 2048;
} else {
ctx_params.n_ctx = params->n_ctx;
}
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return NULL;
}
auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->model = model;
@ -76,18 +75,18 @@ static void llava_free(struct llava_context * ctx_llava) {
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
static struct clip_ctx * clip_init_context(gpt_params * params) {
static struct clip_ctx * clip_init_context(common_params * params) {
const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
return ctx_clip;
}
@ -98,8 +97,8 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
if (n_eval > n_batch) {
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
@ -115,7 +114,7 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
}
@ -123,14 +122,14 @@ static void process_eval_image_embed(struct llava_context * ctx_llava, const str
float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
slice_embed->embed = image_embed;
slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
llava_image_embed_free(slice_embed);
}
static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, common_params * params, int &n_past) {
std::string system_prompt;
int idx = 0;
int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
@ -141,7 +140,10 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
else if (has_minicpmv_projector == 3) {
system_prompt = "<|im_start|>user\n";
}
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
else if (has_minicpmv_projector == 4) {
system_prompt = "<|im_start|>user\n";
}
LOG_INF("%s: image token past: %d\n", __func__, n_past);
eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
@ -160,61 +162,65 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
}
eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
}
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
LOG_INF("%s: image token past: %d\n", __func__, n_past);
}
static const char * sample(struct gpt_sampler * smpl,
static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
gpt_sampler_accept(smpl, id, true);
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = llama_token_to_piece(ctx_llama, id);
ret = common_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past);
return ret.c_str();
}
static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
auto ctx_clip = clip_init_context(params);
auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
static struct llava_context * minicpmv_init(common_params * params, const std::string & fname, int &n_past){
auto * ctx_clip = clip_init_context(params);
auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
if (!embeds) {
std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
return NULL;
}
// process the prompt
if (params->prompt.empty() && params->interactive == false) {
LOG_TEE("prompt should be given or interactive mode should be on");
LOG_ERR("prompt should be given or interactive mode should be on");
return NULL;
}
auto model = llava_init(params);
auto * model = llava_init(params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
return NULL;
}
const int64_t t_llava_init_start_us = ggml_time_us();
auto ctx_llava = llava_init_context(params, model);
auto * ctx_llava = llava_init_context(params, model);
ctx_llava->ctx_clip = ctx_clip;
const int64_t t_llava_init_end_us = ggml_time_us();
float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
const int64_t t_process_image_start_us = ggml_time_us();
process_image(ctx_llava, embeds, params, n_past);
const int64_t t_process_image_end_us = ggml_time_us();
float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
llava_image_embed_free(embeds);
return ctx_llava;
}
static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
std::string user_prompt = prompt;
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (!is_first) {
@ -224,6 +230,9 @@ static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_par
else if (has_minicpmv_projector == 3) {
user_prompt = "<|im_start|>user\n" + prompt;
}
else if (has_minicpmv_projector == 4) {
user_prompt = "<|im_start|>user\n" + prompt;
}
}
eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
@ -233,16 +242,19 @@ static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_par
else if (has_minicpmv_projector == 3) {
eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
}
else if (has_minicpmv_projector == 4) {
eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
}
// generate the response
LOG_TEE("\n");
LOG_INF("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
return smpl;
}
static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){
static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
return tmp;
@ -251,19 +263,13 @@ static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampl
int main(int argc, char ** argv) {
ggml_time_init();
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, show_additional_info);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("llava", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
common_init();
if (params.mmproj.empty() || (params.image.empty())) {
show_additional_info(argc, argv);
@ -272,21 +278,23 @@ int main(int argc, char ** argv) {
for (auto & image : params.image) {
int n_past = 0;
auto ctx_llava = minicpmv_init(&params, image, n_past);
auto * ctx_llava = minicpmv_init(&params, image, n_past);
if (!params.prompt.empty()) {
LOG_TEE("<user>%s\n", params.prompt.c_str());
LOG_TEE("<assistant>");
auto smpl = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
LOG("<user>%s\n", params.prompt.c_str());
LOG("<assistant>");
auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
std::string response;
bool have_tmp = false;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, smpl, n_past);
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0){
if(!have_tmp)continue;
else break;
if (!have_tmp) {
continue;
}
break;
}
if (strstr(tmp, "###")) break; // Yi-VL behavior
have_tmp = true;
@ -295,30 +303,29 @@ int main(int argc, char ** argv) {
fflush(stdout);
}
gpt_sampler_free(smpl);
common_sampler_free(smpl);
}else {
while (true) {
LOG_TEE("<user>");
LOG("<user>");
std::string prompt;
std::getline(std::cin, prompt);
LOG_TEE("<assistant>");
auto smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
LOG("<assistant>");
auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
std::string response;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, smpl, n_past);
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
printf("%s", tmp);// mistral llava-1.6
if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout);
}
gpt_sampler_free(smpl);
common_sampler_free(smpl);
}
}
printf("\n");
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
ctx_llava->model = NULL;
llava_free(ctx_llava);

15
examples/llava/minicpmv-convert-image-encoder-to-gguf.py
View file

@ -501,7 +501,7 @@ default_image_mean = [0.48145466, 0.4578275, 0.40821073]
default_image_std = [0.26862954, 0.26130258, 0.27577711]
ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3', default=2)
ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4', default=2)
# with proper
args = ap.parse_args()
@ -545,12 +545,19 @@ if args.use_f32:
minicpmv_version = args.minicpmv_version
emb_dim = 4096
block_count = 26
if minicpmv_version == 1:
emb_dim = 2304
block_count = 26
elif minicpmv_version == 2:
emb_dim = 4096
block_count = 27
elif minicpmv_version == 3:
emb_dim = 3584
block_count = 27
elif minicpmv_version == 4:
emb_dim = 3584
block_count = 27
default_vision_config = {
"hidden_size": 1152,
@ -567,6 +574,9 @@ model = Idefics2VisionTransformer(vision_config)
if minicpmv_version == 3:
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(vision_config)
elif minicpmv_version == 4:
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(vision_config)
processor = None
# if model.attn_pool is not None:
@ -587,7 +597,7 @@ elif args.minicpmv_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_minicpmv_projector = True
minicpmv_version = 3
minicpmv_version = 4
elif args.vision_only:
fname_middle = "vision-"
has_text_encoder = False
@ -625,7 +635,6 @@ if has_vision_encoder:
fout.add_uint32("clip.vision.projection_dim", 0)
fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
block_count = 26
fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
if processor is not None:

2
examples/llava/minicpmv-surgery.py
View file

@ -8,7 +8,7 @@ ap.add_argument("-m", "--model", help="Path to MiniCPM-V model")
args = ap.parse_args()
# find the model part that includes the the multimodal projector weights
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True, torch_dtype=torch.bfloat16)
checkpoint = model.state_dict()
# get a list of mm tensor names

165
examples/llava/qwen2_vl_surgery.py Normal file
View file

@ -0,0 +1,165 @@
import argparse
from typing import Dict
import torch
import numpy as np
from gguf import *
from transformers import (
Qwen2VLForConditionalGeneration,
Qwen2VLProcessor,
AutoProcessor,
Qwen2VLConfig
)
VISION = "clip.vision"
def k(raw_key: str, arch: str) -> str:
return raw_key.format(arch=arch)
def to_gguf_name(name: str) -> str:
og = name
name = name.replace("text_model", "t").replace("vision_model", "v")
name = name.replace("blocks", "blk").replace("embeddings.", "")
name = name.replace("attn.", "attn_")
name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
# name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
name = name.replace("norm1", "ln1").replace("norm2", "ln2")
name = name.replace("merger.mlp", 'mm')
print(f"[to_gguf_name] {og} --> {name}")
return name
def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
vision_model = qwen2vl.visual
tensor_map = {}
for name, ten in vision_model.state_dict().items():
ten = ten.numpy()
if 'qkv' in name:
if ten.ndim == 2: # weight
c3, _ = ten.shape
else: # bias
c3 = ten.shape[0]
assert c3 % 3 == 0
c = c3 // 3
wq = ten[:c]
wk = ten[c: c * 2]
wv = ten[c * 2:]
tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
elif 'merger' in name:
if name.endswith("ln_q.weight"):
tensor_map['v.post_ln.weight'] = ten
elif name.endswith("ln_q.bias"):
tensor_map['v.post_ln.bias'] = ten
else:
# "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
tensor_map[to_gguf_name(name)] = ten
elif 'patch_embed.proj.weight' in name:
# NOTE: split Conv3D into Conv2Ds
c1, c2, kt, kh, kw = ten.shape
assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
else:
tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
for new_name, ten in tensor_map.items():
if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
tensor_map[new_name] = ten.astype(np.float32)
else:
tensor_map[new_name] = ten.astype(dtype)
tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32) # dummy tensor, just here as a placeholder
return tensor_map
def main(args):
if args.data_type == 'fp32':
dtype = torch.float32
np_dtype = np.float32
ftype = 0
elif args.data_type == 'fp16':
dtype = torch.float32
np_dtype = np.float16
ftype = 1
else:
raise ValueError()
local_model = False
model_path = ""
model_name = args.model_name
print("model_name: ", model_name)
qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
model_name, torch_dtype=dtype, device_map="cpu"
)
cfg: Qwen2VLConfig = qwen2vl.config # type: ignore[reportAssignmentType]
vcfg = cfg.vision_config
if os.path.isdir(model_name):
local_model = True
if model_name.endswith(os.sep):
model_name = model_name[:-1]
model_path = model_name
model_name = os.path.basename(model_name)
fname_out = f"{model_name.replace('/', '-').lower()}-vision.gguf"
fout = GGUFWriter(path=fname_out, arch="clip")
fout.add_description("image encoder for Qwen2VL")
fout.add_file_type(ftype)
fout.add_bool("clip.has_text_encoder", False)
fout.add_bool("clip.has_vision_encoder", True)
fout.add_bool("clip.has_qwen2vl_merger", True)
fout.add_string("clip.projector_type", "qwen2vl_merger")
print(cfg.vision_config)
if 'silu' in cfg.vision_config.hidden_act.lower():
fout.add_bool("clip.use_silu", True)
fout.add_bool("clip.use_gelu", False)
elif 'gelu' in cfg.vision_config.hidden_act.lower():
fout.add_bool("clip.use_silu", False)
fout.add_bool("clip.use_gelu", 'quick' not in cfg.vision_config.hidden_act.lower())
else:
raise ValueError()
tensor_map = find_vision_tensors(qwen2vl, np_dtype)
for name, data in tensor_map.items():
fout.add_tensor(name, data)
fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
fout.add_uint32("clip.vision.image_size", 14 * 40) # some reasonable size that is divable by (14*2)
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), 0) # not sure what this does, put 0 here as a placeholder
fout.add_name(model_name)
"""
HACK: Since vision rope related parameter aren't stored in the `Qwen2VLConfig,
it will be hardcoded in the `clip_image_build_graph` from `clip.cpp`.
"""
if local_model:
processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_path)
else:
processor: Qwen2VLProcessor = AutoProcessor.from_pretrained(model_name)
fout.add_array("clip.vision.image_mean", processor.image_processor.image_mean) # type: ignore[reportAttributeAccessIssue]
fout.add_array("clip.vision.image_std", processor.image_processor.image_std) # type: ignore[reportAttributeAccessIssue]
fout.write_header_to_file()
fout.write_kv_data_to_file()
fout.write_tensors_to_file()
fout.close()
print("save model as: ", fname_out)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
args = parser.parse_args()
main(args)

584
examples/llava/qwen2vl-cli.cpp Normal file
View file

@ -0,0 +1,584 @@
#include "arg.h"
#include "base64.hpp"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "ggml.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
#ifdef NDEBUG
#include "ggml-alloc.h"
#include "ggml-backend.h"
#endif
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {
int n_embd = llama_model_n_embd(llama_get_model(ctx_llama));
const int patch_size = 14 * 2;
const int ph = image_size->height / patch_size + (image_size->height % patch_size > 0);
const int pw = image_size->width / patch_size + (image_size->width % patch_size > 0);
auto img_tokens = image_embed->n_image_pos;
// llama_pos mrope_pos[img_tokens * 4];
std::vector<llama_pos> mrope_pos;
mrope_pos.resize(img_tokens * 4);
for (int y = 0; y < ph; y++)
{
for (int x = 0; x < pw; x++)
{
int i = y * pw + x;
mrope_pos[i] = *st_pos_id;
mrope_pos[i + img_tokens] = *st_pos_id + y;
mrope_pos[i + img_tokens * 2] = *st_pos_id + x;
mrope_pos[i + img_tokens * 3] = 0;
}
}
*st_pos_id += std::max(pw, ph);
int processed = 0;
std::vector<llama_pos> batch_mrope_pos;
batch_mrope_pos.resize(img_tokens * 4);
for (int i = 0; i < img_tokens; i += n_batch) {
int n_eval = img_tokens - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
// llama_pos batch_mrope_pos[n_eval * 4];
std::fill(batch_mrope_pos.begin(), batch_mrope_pos.end(), 0);
memcpy(batch_mrope_pos.data(), &mrope_pos[processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 1], &mrope_pos[img_tokens * 1 + processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 2], &mrope_pos[img_tokens * 2 + processed], n_eval * sizeof(llama_pos));
memcpy(&batch_mrope_pos[n_eval * 3], &mrope_pos[img_tokens * 3 + processed], n_eval * sizeof(llama_pos));
llama_batch batch = {
int32_t(n_eval), // n_tokens
nullptr, // token
(image_embed->embed+i*n_embd), // embed
batch_mrope_pos.data(), // pos
nullptr, // n_seq_id
nullptr, // seq_id
nullptr, // logits
};
if (llama_decode(ctx_llama, batch)) {
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
*n_past += n_eval;
processed += n_eval;
}
return true;
}
static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past, int * st_pos_id) {
int N = (int) tokens.size();
std::vector<llama_pos> pos;
for (int i = 0; i < N; i += n_batch) {
int n_eval = (int) tokens.size() - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
auto batch = llama_batch_get_one(&tokens[i], n_eval);
// TODO: add mrope pos ids somewhere else
pos.resize(batch.n_tokens * 4);
std::fill(pos.begin(), pos.end(), 0);
for (int j = 0; j < batch.n_tokens * 3; j ++) {
pos[j] = *st_pos_id + (j % batch.n_tokens);
}
batch.pos = pos.data();
if (llama_decode(ctx_llama, batch)) {
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
*st_pos_id += n_eval;
}
return true;
}
static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past, int * st_pos_id) {
std::vector<llama_token> tokens;
tokens.push_back(id);
return eval_tokens(ctx_llama, tokens, 1, n_past, st_pos_id);
}
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, int * st_pos_id, bool add_bos){
std::string str2 = str;
std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
eval_tokens(ctx_llama, embd_inp, n_batch, n_past, st_pos_id);
return true;
}
static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past, int * st_pos_id) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past, st_pos_id);
return ret.c_str();
}
static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
static const char* IMG_BASE64_TAG_END = "\">";
static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
}
static bool prompt_contains_image(const std::string& prompt) {
size_t begin, end;
find_image_tag_in_prompt(prompt, begin, end);
return (begin != std::string::npos);
}
// replaces the base64 image tag in the prompt with `replacement`
static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
size_t img_base64_str_start, img_base64_str_end;
find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
return NULL;
}
auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
auto required_bytes = base64::required_encode_size(base64_str.size());
auto img_bytes = std::vector<unsigned char>(required_bytes);
base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
if (!embed) {
LOG_ERR("%s: could not load image from base64 string.\n", __func__);
return NULL;
}
return embed;
}
static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
size_t begin, end;
find_image_tag_in_prompt(prompt, begin, end);
if (begin == std::string::npos || end == std::string::npos) {
return prompt;
}
auto pre = prompt.substr(0, begin);
auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
return pre + replacement + post;
}
struct llava_context {
struct clip_ctx * ctx_clip = NULL;
struct llama_context * ctx_llama = NULL;
struct llama_model * model = NULL;
};
static void print_usage(int, char ** argv) {
LOG("\n example usage:\n");
LOG("\n %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
// load and preprocess the image
llava_image_embed * embed = NULL;
auto prompt = params->prompt;
if (prompt_contains_image(prompt)) {
if (!params->image.empty()) {
LOG_INF("using base64 encoded image instead of command line image path\n");
}
embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
if (!embed) {
LOG_ERR("%s: can't load image from prompt\n", __func__);
return NULL;
}
params->prompt = remove_image_from_prompt(prompt);
} else {
embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
if (!embed) {
fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
return NULL;
}
}
return embed;
}
static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
int n_past = 0;
int cur_pos_id = 0;
const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
std::string system_prompt, user_prompt;
size_t image_pos = prompt.find("<|vision_start|>");
if (image_pos != std::string::npos) {
// new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<|vision_pad|>").length());
LOG_INF("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
LOG_INF("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
// llava-1.5 native mode
system_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|>";
user_prompt = "<|vision_end|>" + prompt + "<|im_end|>\n<|im_start|>assistant\n";
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, true);
if (image_embed != nullptr) {
auto image_size = clip_get_load_image_size(ctx_llava->ctx_clip);
qwen2vl_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past, &cur_pos_id, image_size);
}
eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, &cur_pos_id, false);
// generate the response
LOG("\n");
struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
if (!smpl) {
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past, &cur_pos_id);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
LOG("%s", tmp);
if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
fflush(stdout);
}
common_sampler_free(smpl);
LOG("\n");
}
static struct llama_model * llava_init(common_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return NULL;
}
auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->ctx_clip = ctx_clip;
ctx_llava->model = model;
return ctx_llava;
}
static void llava_free(struct llava_context * ctx_llava) {
if (ctx_llava->ctx_clip) {
clip_free(ctx_llava->ctx_clip);
ctx_llava->ctx_clip = NULL;
}
llama_free(ctx_llava->ctx_llama);
llama_model_free(ctx_llava->model);
llama_backend_free();
}
#ifndef NDEBUG
static void debug_test_mrope_2d() {
// 1. Initialize backend
ggml_backend_t backend = NULL;
std::string backend_name = "";
#ifdef GGML_USE_CUDA
fprintf(stderr, "%s: using CUDA backend\n", __func__);
backend = ggml_backend_cuda_init(0); // init device 0
backend_name = "cuda";
if (!backend) {
fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
}
#endif
// if there aren't GPU Backends fallback to CPU backend
if (!backend) {
backend = ggml_backend_cpu_init();
backend_name = "cpu";
}
// Calculate the size needed to allocate
size_t ctx_size = 0;
ctx_size += 2 * ggml_tensor_overhead(); // tensors
// no need to allocate anything else!
// 2. Allocate `ggml_context` to store tensor data
struct ggml_init_params params = {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_backend_alloc_ctx_tensors()
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 128, 12, 30);
ggml_set_name(inp_raw, "inp_raw");
ggml_set_input(inp_raw);
struct ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 30 * 4);
ggml_set_name(pos, "pos");
ggml_set_input(pos);
std::vector<float> dummy_q;
dummy_q.resize(128 * 12 * 30);
std::fill(dummy_q.begin(), dummy_q.end(), 0.1);
// memcpy(inp_raw->data, dummy_q.data(), 128 * 12 * 30 * ggml_element_size(inp_raw));
std::vector<int> pos_id;
pos_id.resize(30 * 4);
for (int i = 0; i < 30; i ++) {
pos_id[i] = i;
pos_id[i + 30] = i + 10;
pos_id[i + 60] = i + 20;
pos_id[i + 90] = i + 30;
}
int sections[4] = {32, 32, 0, 0};
// 4. Allocate a `ggml_backend_buffer` to store all tensors
ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
// 5. Copy tensor data from main memory (RAM) to backend buffer
ggml_backend_tensor_set(inp_raw, dummy_q.data(), 0, ggml_nbytes(inp_raw));
ggml_backend_tensor_set(pos, pos_id.data(), 0, ggml_nbytes(pos));
// 6. Create a `ggml_cgraph` for mul_mat operation
struct ggml_cgraph * gf = NULL;
struct ggml_context * ctx_cgraph = NULL;
// create a temporally context to build the graph
struct ggml_init_params params0 = {
/*.mem_size =*/ ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true, // the tensors will be allocated later by ggml_gallocr_alloc_graph()
};
ctx_cgraph = ggml_init(params0);
gf = ggml_new_graph(ctx_cgraph);
struct ggml_tensor * result0 = ggml_rope_multi(
ctx_cgraph, inp_raw, pos, nullptr,
128/2, sections, LLAMA_ROPE_TYPE_VISION, 32768, 1000000, 1,
0, 1, 32, 1);
// Add "result" tensor and all of its dependencies to the cgraph
ggml_build_forward_expand(gf, result0);
// 7. Create a `ggml_gallocr` for cgraph computation
ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
ggml_gallocr_alloc_graph(allocr, gf);
// 9. Run the computation
int n_threads = 1; // Optional: number of threads to perform some operations with multi-threading
if (ggml_backend_is_cpu(backend)) {
ggml_backend_cpu_set_n_threads(backend, n_threads);
}
ggml_backend_graph_compute(backend, gf);
// 10. Retrieve results (output tensors)
// in this example, output tensor is always the last tensor in the graph
struct ggml_tensor * result = result0;
// struct ggml_tensor * result = gf->nodes[gf->n_nodes - 1];
float * result_data = (float *)malloc(ggml_nbytes(result));
// because the tensor data is stored in device buffer, we need to copy it back to RAM
ggml_backend_tensor_get(result, result_data, 0, ggml_nbytes(result));
const std::string bin_file = "mrope_2d_" + backend_name +".bin";
std::ofstream outFile(bin_file, std::ios::binary);
if (outFile.is_open()) {
outFile.write(reinterpret_cast<const char*>(result_data), ggml_nbytes(result));
outFile.close();
std::cout << "Data successfully written to " + bin_file << std::endl;
} else {
std::cerr << "Error opening file!" << std::endl;
}
free(result_data);
// 11. Free memory and exit
ggml_free(ctx_cgraph);
ggml_gallocr_free(allocr);
ggml_free(ctx);
ggml_backend_buffer_free(buffer);
ggml_backend_free(backend);
}
static void debug_dump_img_embed(struct llava_context * ctx_llava) {
int n_embd = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
int ne = n_embd * 4;
float vals[56 * 56 * 3];
// float embd[ne];
std::vector<float> embd;
embd.resize(ne);
for (int i = 0; i < 56*56; i++)
{
for (int c = 0; c < 3; c++)
vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
}
clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
std::ofstream outFile("img_embed.bin", std::ios::binary);
if (outFile.is_open()) {
outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
outFile.close();
std::cout << "Data successfully written to mrope.bin" << std::endl;
} else {
std::cerr << "Error opening file!" << std::endl;
}
}
#endif
int main(int argc, char ** argv) {
ggml_time_init();
common_params params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
return 1;
}
common_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
auto * model = llava_init(&params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
return 1;
}
if (prompt_contains_image(params.prompt)) {
auto * ctx_llava = llava_init_context(&params, model);
auto * image_embed = load_image(ctx_llava, &params, "");
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
#ifndef NDEBUG
} else if (params.image[0].empty()) {
auto ctx_llava = llava_init_context(&params, model);
debug_test_mrope_2d();
debug_dump_img_embed(ctx_llava);
llama_perf_context_print(ctx_llava->ctx_llama);
ctx_llava->model = NULL;
llava_free(ctx_llava);
#endif
} else {
for (auto & image : params.image) {
auto * ctx_llava = llava_init_context(&params, model);
auto * image_embed = load_image(ctx_llava, &params, image);
if (!image_embed) {
LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
return 1;
}
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
}
}
llama_model_free(model);
return 0;
}

2
examples/lookahead/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-lookahead)
add_executable(${TARGET} lookahead.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

117
examples/lookahead/lookahead.cpp
View file

@ -1,4 +1,7 @@
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cstdio>
@ -34,54 +37,51 @@ struct ngram_container {
};
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
common_init();
const int W = 15; // lookahead window
const int N = 5; // n-gram size
const int G = 15; // max verification n-grams
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookahead", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
// load the target model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// Tokenize the prompt
std::vector<llama_token> inp;
std::vector<llama_token> all;
inp = ::llama_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx, params.prompt, true, true);
all = inp;
const int max_context_size = llama_n_ctx(ctx);
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", common_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
@ -91,8 +91,8 @@ int main(int argc, char ** argv) {
const auto t_enc_start = ggml_time_us();
// eval the prompt
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1));
for (int s = 1; s < W + G + 1; ++s) {
llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);
@ -117,7 +117,7 @@ int main(int argc, char ** argv) {
llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
// target model sampling context
struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
struct common_sampler * smpl = common_sampler_init(model, params.sampling);
// verification n-grams
std::vector<ngram_data> ngrams_cur(G);
@ -149,7 +149,7 @@ int main(int argc, char ** argv) {
}
// here we keep adding new n-grams as we go
ngram_container ngrams_observed(llama_n_vocab(model), N, G);
ngram_container ngrams_observed(llama_vocab_n_tokens(vocab), N, G);
// debug
struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, W + G + 1);
@ -158,14 +158,14 @@ int main(int argc, char ** argv) {
// sample first token
{
id = gpt_sampler_sample(smpl, ctx, 0);
id = common_sampler_sample(smpl, ctx, 0);
gpt_sampler_accept(smpl, id, true);
common_sampler_accept(smpl, id, true);
{
const std::string token_str = llama_token_to_piece(ctx, id);
const std::string token_str = common_token_to_piece(ctx, id);
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
fflush(stdout);
}
}
@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
common_kv_cache_dump_view_seqs(kvc_view, 40);
}
// build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
@ -203,10 +203,10 @@ int main(int argc, char ** argv) {
// V V V V V V
// id
{
llama_batch_clear(batch);
common_batch_clear(batch);
// current token - first token of the first level
llama_batch_add(batch, id, n_past, seq_id_all, true);
common_batch_add(batch, id, n_past, seq_id_all, true);
// verification n-grams - queue this before the lookahead tokens for less KV cache fragmentation
{
@ -231,7 +231,7 @@ int main(int argc, char ** argv) {
ngrams_cur[g].tokens [j + 1] = t;
ngrams_cur[g].i_batch[j + 1] = batch.n_tokens;
llama_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
common_batch_add(batch, t, n_past + j + 1, { W + 1 + g }, true);
}
}
}
@ -243,19 +243,19 @@ int main(int argc, char ** argv) {
seq_id_look[j] = i + j + 1;
}
llama_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
common_batch_add(batch, tokens_j[0][i], n_past + i, seq_id_look, false);
}
// fill the rest of the levels
for (int j = 1; j < N - 1; j++) {
for (int i = 0; i < W; i++) {
llama_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
common_batch_add(batch, tokens_j[j][i], n_past + j + i, { i + 1 }, j == N - 2);
}
}
}
if (llama_decode(ctx, batch) != 0) {
fprintf(stderr, "\n\n%s: error: llama_decode failed - increase KV cache size\n", __func__);
LOG_ERR("\n\n%s: llama_decode failed - increase KV cache size\n", __func__);
return 1;
}
@ -283,23 +283,23 @@ int main(int argc, char ** argv) {
}
// sample the next token
id = gpt_sampler_sample(smpl, ctx, i_batch);
id = common_sampler_sample(smpl, ctx, i_batch);
gpt_sampler_accept(smpl, id, true);
common_sampler_accept(smpl, id, true);
// print
{
const std::string token_str = llama_token_to_piece(ctx, id);
const std::string token_str = common_token_to_piece(ctx, id);
if (v == 0) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
} else {
// print light cyan
printf("\033[0;96m%s\033[0m", token_str.c_str());
LOG("\033[0;96m%s\033[0m", token_str.c_str());
}
fflush(stdout);
if (llama_token_is_eog(model, id)) {
if (llama_vocab_is_eog(vocab, id)) {
has_eos = true;
}
@ -329,21 +329,21 @@ int main(int argc, char ** argv) {
// print known n-grams starting with token id (debug)
if (0 && v == 0) {
if (ngrams_observed.cnt[id] > 0) {
printf("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str());
LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], common_token_to_piece(ctx, id).c_str());
}
for (int i = 0; i < ngrams_observed.cnt[id]; i++) {
printf(" - ngram %2d: ", i);
LOG(" - ngram %2d: ", i);
const int idx = id*(N - 1)*G + i*(N - 1);
for (int j = 0; j < N - 1; j++) {
const std::string token_str = llama_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
const std::string token_str = common_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
printf("\n");
LOG("\n");
}
}
@ -360,7 +360,7 @@ int main(int argc, char ** argv) {
if (v == 0) {
// sample from the last level
for (int i = 0; i < W; i++) {
tokens_j[N - 2][i] = gpt_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
tokens_j[N - 2][i] = common_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
}
} else {
for (int i = 0; i < W; i++) {
@ -454,34 +454,31 @@ int main(int argc, char ** argv) {
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("W = %2d\n", W);
LOG_TEE("N = %2d\n", N);
LOG_TEE("G = %2d\n", G);
LOG_TEE("\n");
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_accept = %d\n", n_accept);
LOG_INF("\n");
LOG_INF("W = %2d\n", W);
LOG_INF("N = %2d\n", N);
LOG_INF("G = %2d\n", G);
LOG_INF("\n");
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_accept = %d\n", n_accept);
LOG_TEE("\n");
gpt_perf_print(ctx, smpl);
LOG_INF("\n");
common_perf_print(ctx, smpl);
gpt_sampler_free(smpl);
common_sampler_free(smpl);
llama_kv_cache_view_free(&kvc_view);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

8
examples/lookup/CMakeLists.txt
View file

@ -2,22 +2,22 @@ set(TARGET llama-lookup)
add_executable(${TARGET} lookup.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-lookup-create)
add_executable(${TARGET} lookup-create.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-lookup-merge)
add_executable(${TARGET} lookup-merge.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-lookup-stats)
add_executable(${TARGET} lookup-stats.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

31
examples/lookup/lookup-create.cpp
View file

@ -1,20 +1,15 @@
#include "ggml.h"
#include "llama.h"
#include "arg.h"
#include "common.h"
#include "ngram-cache.h"
#include "llama.h"
#include <cstdint>
#include <fstream>
#include <iostream>
#include <string>
#include <unordered_map>
#include <vector>
int main(int argc, char ** argv){
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
@ -23,21 +18,23 @@ int main(int argc, char ** argv){
llama_numa_init(params.numa);
// load the model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model_ptr & model = llama_init.model;
llama_context_ptr & ctx = llama_init.context;
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
GGML_ASSERT(model != nullptr);
// tokenize the prompt
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx.get(), params.prompt, true, true);
fprintf(stderr, "%s: tokenization done\n", __func__);
llama_ngram_cache ngram_cache;
llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
common_ngram_cache ngram_cache;
common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
common_ngram_cache_save(ngram_cache, params.lookup_cache_static);
return 0;
}

8
examples/lookup/lookup-merge.cpp
View file

@ -33,15 +33,15 @@ int main(int argc, char ** argv){
}
fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
common_ngram_cache ngram_cache_merged = common_ngram_cache_load(args[0]);
for (size_t i = 1; i < args.size()-1; ++i) {
fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
common_ngram_cache ngram_cache = common_ngram_cache_load(args[i]);
llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
common_ngram_cache_merge(ngram_cache_merged, ngram_cache);
}
fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
llama_ngram_cache_save(ngram_cache_merged, args.back());
common_ngram_cache_save(ngram_cache_merged, args.back());
}

72
examples/lookup/lookup-stats.cpp
View file

@ -1,44 +1,45 @@
#include "ggml.h"
#include "arg.h"
#include "common.h"
#include "llama.h"
#include "log.h"
#include "ngram-cache.h"
#include "llama.h"
#include "ggml.h"
#include <cmath>
#include <cstdint>
#include <cstdio>
#include <cinttypes>
#include <fstream>
#include <string>
#include <vector>
#include <unordered_map>
int main(int argc, char ** argv){
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
const int n_draft = params.n_draft;
common_init();
const int n_draft = params.speculative.n_max;
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
// load the model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_context_ptr & ctx = llama_init.context;
// tokenize the prompt
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx.get(), params.prompt, true, true);
common_ngram_cache ngram_cache_context;
common_ngram_cache ngram_cache_dynamic;
common_ngram_cache ngram_cache_static;
llama_ngram_cache ngram_cache_context;
llama_ngram_cache ngram_cache_dynamic;
llama_ngram_cache ngram_cache_static;
int64_t t_draft_flat_us = 0;
int64_t t_draft_us = 0;
@ -47,16 +48,16 @@ int main(int argc, char ** argv){
if (!params.lookup_cache_static.empty()) {
try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) {
fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1);
}
}
if (!params.lookup_cache_dynamic.empty()) {
try {
ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
} catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
}
@ -64,7 +65,7 @@ int main(int argc, char ** argv){
}
const int n_input = inp.size();
const int n_ctx = llama_n_ctx(ctx);
const int n_ctx = llama_n_ctx(ctx.get());
int n_drafted = 0;
int n_accept = 0;
@ -85,7 +86,7 @@ int main(int argc, char ** argv){
{
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
common_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
t_draft_us += ggml_time_us() - t_start_draft_us;
}
@ -104,7 +105,7 @@ int main(int argc, char ** argv){
{
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us;
}
}
@ -114,7 +115,7 @@ int main(int argc, char ** argv){
pseudo_output.push_back(inp_slice[pseudo_output.size()]);
{
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us;
}
}
@ -128,32 +129,29 @@ int main(int argc, char ** argv){
const int64_t eta_min = eta_ms / (60*1000);
const int64_t eta_s = (eta_ms - 60*1000*eta_min) / 1000;
LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
LOG_INF("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
}
// After each chunk, update the dynamic ngram cache with the context ngram cache:
llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
ngram_cache_context.clear();
}
LOG_TEE("\n");
LOG("\n");
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_input - n_input % n_ctx);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_input - n_input % n_ctx);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_INF("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
llama_free(ctx);
llama_free_model(model);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

128
examples/lookup/lookup.cpp
View file

@ -1,7 +1,10 @@
#include "arg.h"
#include "ggml.h"
#include "llama.h"
#include "common.h"
#include "ngram-cache.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cstdint>
#include <cstdio>
@ -10,61 +13,58 @@
#include <vector>
int main(int argc, char ** argv){
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
common_init();
// max. number of additional tokens to draft if match is found
const int n_draft = params.n_draft;
const int n_draft = params.speculative.n_max;
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookup", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
// load the model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// tokenize the prompt
std::vector<llama_token> inp;
inp = ::llama_tokenize(ctx, params.prompt, true, true);
inp = common_tokenize(ctx, params.prompt, true, true);
llama_ngram_cache ngram_cache_context;
llama_ngram_cache ngram_cache_dynamic;
llama_ngram_cache ngram_cache_static;
common_ngram_cache ngram_cache_context;
common_ngram_cache ngram_cache_dynamic;
common_ngram_cache ngram_cache_static;
int64_t t_draft_flat_us = 0;
int64_t t_draft_us = 0;
{
// Fill up context ngram cache with tokens from user input:
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
if (!params.lookup_cache_static.empty()) {
try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
ngram_cache_static = common_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) {
fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1);
}
}
if (!params.lookup_cache_dynamic.empty()) {
try {
ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
ngram_cache_dynamic = common_ngram_cache_load(params.lookup_cache_dynamic);
} catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
}
@ -75,14 +75,14 @@ int main(int argc, char ** argv){
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", common_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
@ -91,8 +91,8 @@ int main(int argc, char ** argv){
const auto t_enc_start = ggml_time_us();
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0, 0));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1, n_input - 1, 0));
llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
llama_decode(ctx, llama_batch_get_one(&inp.back(), 1));
const auto t_enc_end = ggml_time_us();
@ -104,7 +104,7 @@ int main(int argc, char ** argv){
bool has_eos = false;
struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
struct common_sampler * smpl = common_sampler_init(model, params.sampling);
std::vector<llama_token> draft;
@ -119,26 +119,26 @@ int main(int argc, char ** argv){
// debug
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
common_kv_cache_dump_view_seqs(kvc_view, 40);
}
// print current draft sequence
LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
LOG_DBG("drafted %s\n", string_from(ctx, draft).c_str());
int i_dft = 0;
while (true) {
// sample from the target model
llama_token id = gpt_sampler_sample(smpl, ctx, i_dft);
llama_token id = common_sampler_sample(smpl, ctx, i_dft);
gpt_sampler_accept(smpl, id, true);
common_sampler_accept(smpl, id, true);
const std::string token_str = llama_token_to_piece(ctx, id);
const std::string token_str = common_token_to_piece(ctx, id);
if (!params.use_color) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
if (llama_token_is_eog(model, id)) {
if (llama_vocab_is_eog(vocab, id)) {
has_eos = true;
}
@ -146,7 +146,7 @@ int main(int argc, char ** argv){
// check if the target token matches the draft
if (i_dft < (int) draft.size() && id == draft[i_dft]) {
LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
LOG_DBG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
++n_accept;
++n_past;
++i_dft;
@ -154,25 +154,25 @@ int main(int argc, char ** argv){
{
// Update context ngram cache with the newly accepted token:
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us;
}
if (params.use_color) {
// color accepted draft token
printf("\033[34m%s\033[0m", token_str.c_str());
LOG("\033[34m%s\033[0m", token_str.c_str());
fflush(stdout);
}
continue;
}
if (params.use_color) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
fflush(stdout);
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
LOG_DBG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
draft.clear();
draft.push_back(id);
@ -180,7 +180,7 @@ int main(int argc, char ** argv){
{
// Update context ngram cache with the newly accepted token:
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
t_draft_us += ggml_time_us() - t_start_draft_us;
}
break;
@ -194,18 +194,18 @@ int main(int argc, char ** argv){
// clean the cache of draft tokens that weren't accepted
llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
llama_batch_clear(batch_tgt);
llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
common_batch_clear(batch_tgt);
common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
// Draft already contains a single token sampled from the model:
GGML_ASSERT(draft.size() == 1);
GGML_ASSERT(draft[0] == inp.back());
const int64_t t_start_draft_us = ggml_time_us();
llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
common_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
for (size_t i = 1; i < draft.size(); ++i) {
llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
}
t_draft_us += ggml_time_us() - t_start_draft_us;
@ -220,38 +220,34 @@ int main(int argc, char ** argv){
auto t_dec_end = ggml_time_us();
// Update dynamic ngram cache with context ngram cache and save it to disk:
llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
common_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_INF("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_TEE("\ntarget:\n\n");
llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG_INF("\ntarget:\n\n");
common_perf_print(ctx, smpl);
gpt_sampler_free(smpl);
common_sampler_free(smpl);
llama_batch_free(batch_tgt);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

32
examples/main-cmake-pkg/CMakeLists.txt
View file

@ -1,32 +0,0 @@
cmake_minimum_required(VERSION 3.12)
project("llama-cli-cmake-pkg" C CXX)
set(TARGET llama-cli-cmake-pkg)
find_package(Llama 0.0.1 REQUIRED)
# Bake common functionality in with target. Because applications
# using the relocatable Llama package should be outside of the
# source tree, llama-cli-cmake-pkg pretends the dependencies are built-in.
set(_common_path "${CMAKE_CURRENT_LIST_DIR}/../../common")
add_library(common OBJECT)
file(GLOB _common_files
"${_common_path}/*.h"
"${_common_path}/*.cpp"
)
target_sources(common PRIVATE ${_common_files})
# If the common project was part of "llama-cli-cmake-pkg" the transient
# defines would automatically be attached. Because the common func-
# tionality is separate, but dependent upon the defines, it must be
# explicitly extracted from the "llama" target.
#
get_target_property(_llama_transient_defines llama
INTERFACE_COMPILE_DEFINITIONS)
target_compile_definitions(common PRIVATE "${_llama_transient_defines}")
add_executable(${TARGET} ${CMAKE_CURRENT_LIST_DIR}/../main/main.cpp)
target_include_directories(${TARGET} PRIVATE ${_common_path})
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

31
examples/main-cmake-pkg/README.md
View file

@ -1,31 +0,0 @@
# llama.cpp/example/main-cmake-pkg
This program builds [llama-cli](../main) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree.
## Building
Because this example is "outside of the source tree", it is important to first build/install llama.cpp using CMake. An example is provided here, but please see the [llama.cpp build instructions](../..) for more detailed build instructions.
### Considerations
When hardware acceleration libraries are used (e.g. CUDA, Metal, etc.), CMake must be able to locate the associated CMake package.
### Build llama.cpp and install to C:\LlamaCPP directory
```cmd
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
cmake -B build -DBUILD_SHARED_LIBS=OFF -G "Visual Studio 17 2022" -A x64
cmake --build build --config Release
cmake --install build --prefix C:/LlamaCPP
```
### Build llama-cli-cmake-pkg
```cmd
cd ..\examples\main-cmake-pkg
cmake -B build -DBUILD_SHARED_LIBS=OFF -DCMAKE_PREFIX_PATH="C:/LlamaCPP/lib/cmake/Llama" -G "Visual Studio 17 2022" -A x64
cmake --build build --config Release
cmake --install build --prefix C:/MyLlamaApp
```

2
examples/main/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-cli)
add_executable(${TARGET} main.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

74
examples/main/README.md
View file

@ -66,10 +66,10 @@ In this section, we cover the most commonly used options for running the `llama-
- `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file (e.g [https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true](https://huggingface.co/ggml-org/gemma-1.1-7b-it-Q4_K_M-GGUF/resolve/main/gemma-1.1-7b-it.Q4_K_M.gguf?download=true)).
- `-i, --interactive`: Run the program in interactive mode, allowing you to provide input directly and receive real-time responses.
- `-n N, --n-predict N`: Set the number of tokens to predict when generating text. Adjusting this value can influence the length of the generated text.
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference.
- `-c N, --ctx-size N`: Set the size of the prompt context. The default is 4096, but if a LLaMA model was built with a longer context, increasing this value will provide better results for longer input/inference.
- `-mli, --multiline-input`: Allows you to write or paste multiple lines without ending each in '\'
- `-t N, --threads N`: Set the number of threads to use during generation. For optimal performance, it is recommended to set this value to the number of physical CPU cores your system has.
- - `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
- `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
## Input Prompts
@ -131,7 +131,7 @@ During text generation, LLaMA models have a limited context size, which means th
### Context Size
- `-c N, --ctx-size N`: Set the size of the prompt context (default: 0, 0 = loaded from model). The LLaMA models were built with a context of 2048-8192, which will yield the best results on longer input/inference.
- `-c N, --ctx-size N`: Set the size of the prompt context (default: 4096, 0 = loaded from model). If a LLaMA model was built with a longer context, increasing this value will yield the best results on longer input/inference.
### Extended Context Size
@ -161,6 +161,8 @@ A value of -1 will enable infinite text generation, even though we have a finite
If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
The `--no-context-shift` option allows you to stop the infinite text generation once the finite context window is full.
It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
### Temperature
@ -175,15 +177,34 @@ Example usage: `--temp 0`
- `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
- `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
- `--no-penalize-nl`: Disable penalization for newline tokens when applying the repeat penalty.
The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.
The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
Use the `--no-penalize-nl` option to disable newline penalization when applying the repeat penalty. This option is particularly useful for generating chat conversations, dialogues, code, poetry, or any text where newline tokens play a significant role in structure and formatting. Disabling newline penalization helps maintain the natural flow and intended formatting in these specific use cases.
### DRY Repetition Penalty
Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
### Top-K Sampling
@ -209,14 +230,6 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e
Example usage: `--min-p 0.05`
### Tail-Free Sampling (TFS)
- `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
Example usage: `--tfs 0.95`
### Locally Typical Sampling
- `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
@ -239,6 +252,19 @@ The `--mirostat-ent` option sets the Mirostat target entropy (tau), which repres
Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
### XTC Sampling
- `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
- `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
### Logit Bias
- `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
@ -282,15 +308,11 @@ These options help improve the performance and memory usage of the LLaMA models.
These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
### Memory Float 32
- `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
### Batch Size
- `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
- `-ub N`, `--ubatch-size N`: Physical batch size. This is the maximum number of tokens that may be processed at a time. Increasing this value may improve performance during prompt processing, at the expense of higher memory usage. Default: `512`.
- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
- `-b N`, `--batch-size N`: Logical batch size. Increasing this value above the value of the physical batch size may improve prompt processing performance when using multiple GPUs with pipeline parallelism. Default: `2048`.
### Prompt Caching
@ -306,14 +328,22 @@ These options help improve the performance and memory usage of the LLaMA models.
For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
## LoRA (Low-Rank Adaptation) adapters
- `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
- `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
## Additional Options
These options provide extra functionality and customization when running the LLaMA models:
- `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
- `--verbose-prompt`: Print the prompt before generating text.
- `--no-display-prompt`: Don't print prompt at generation.
- `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
- `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
- `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
- `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
- `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable or in an OS-specific local cache.

528
examples/main/main.cpp
View file

@ -1,11 +1,11 @@
#include "arg.h"
#include "common.h"
#include "console.h"
#include "log.h"
#include "sampling.h"
#include "llama.h"
#include "chat-template.hpp"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
@ -31,21 +31,25 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static const char * DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant";
static llama_context ** g_ctx;
static llama_model ** g_model;
static gpt_sampler ** g_smpl;
static gpt_params * g_params;
static common_sampler ** g_smpl;
static common_params * g_params;
static std::vector<llama_token> * g_input_tokens;
static std::ostringstream * g_output_ss;
static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static bool need_insert_eot = false;
static void print_usage(int, char ** argv) {
printf("\nexample usage:\n");
printf("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
printf("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
printf("\n");
static void print_usage(int argc, char ** argv) {
(void) argc;
LOG("\nexample usage:\n");
LOG("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
LOG("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
LOG("\n");
}
static bool file_exists(const std::string & path) {
@ -60,50 +64,6 @@ static bool file_is_empty(const std::string & path) {
return f.tellg() == 0;
}
static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model,
const std::vector<llama_token> & input_tokens, const std::string & output,
const std::vector<llama_token> & output_tokens
) {
if (params.logdir.empty()) {
return;
}
const std::string timestamp = string_get_sortable_timestamp();
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
const std::string logfile_path = params.logdir + timestamp + ".yml";
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
fprintf(logfile, "binary: main\n");
char model_desc[128];
llama_model_desc(model, model_desc, sizeof(model_desc));
yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
fprintf(logfile, "\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "# Generation Results #\n");
fprintf(logfile, "######################\n");
fprintf(logfile, "\n");
yaml_dump_string_multiline(logfile, "output", output.c_str());
yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
llama_perf_dump_yaml(logfile, ctx);
fclose(logfile);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
@ -112,49 +72,29 @@ static void sigint_handler(int signo) {
need_insert_eot = true;
} else {
console::cleanup();
printf("\n");
gpt_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
LOG("\n");
common_perf_print(*g_ctx, *g_smpl);
// make sure all logs are flushed
LOG("Interrupted by user\n");
common_log_pause(common_log_main());
_exit(130);
}
}
}
#endif
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
llama_chat_msg new_msg{role, content};
auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
chat_msgs.push_back({role, content});
LOG("formatted: %s\n", formatted.c_str());
return formatted;
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
g_params = &params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_MAIN, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
return 1;
}
auto & sparams = params.sparams;
common_init();
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("main", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
// TODO: Dump params ?
//LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
auto & sparams = params.sampling;
// save choice to use color for later
// (note for later: this is a slightly awkward choice)
@ -162,68 +102,70 @@ int main(int argc, char ** argv) {
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
LOG_WRN("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
LOG_WRN("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
print_build_info();
LOG_INF("%s: llama backend init\n", __func__);
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
LOG("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = nullptr;
llama_context * ctx = nullptr;
gpt_sampler * smpl = nullptr;
std::vector<llama_chat_msg> chat_msgs;
common_sampler * smpl = nullptr;
g_model = &model;
g_ctx = &ctx;
g_smpl = &smpl;
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
llama_init_result llama_init = llama_init_from_gpt_params(params);
std::vector<common_chat_msg> chat_msgs;
model = llama_init.model;
ctx = llama_init.context;
// load the model and apply lora adapter, if any
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
common_init_result llama_init = common_init_from_params(params);
model = llama_init.model.get();
ctx = llama_init.context.get();
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
LOG_ERR("%s: error: unable to load model\n", __func__);
return 1;
}
LOG("%s: llama threadpool init = n_threads = %d\n",
__func__,
(int) params.cpuparams.n_threads
);
const llama_vocab * vocab = llama_model_get_vocab(model);
auto chat_templates = common_chat_templates_from_model(model, params.chat_template);
LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
auto * ggml_threadpool_new_fn = (decltype(ggml_threadpool_new) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_new");
auto * ggml_threadpool_free_fn = (decltype(ggml_threadpool_free) *) ggml_backend_reg_get_proc_address(reg, "ggml_threadpool_free");
struct ggml_threadpool_params tpp_batch =
ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
struct ggml_threadpool_params tpp =
@ -233,108 +175,134 @@ int main(int argc, char ** argv) {
struct ggml_threadpool * threadpool_batch = NULL;
if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
threadpool_batch = ggml_threadpool_new(&tpp_batch);
threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
if (!threadpool_batch) {
LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
exit(1);
LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
return 1;
}
// Start the non-batch threadpool in the paused state
tpp.paused = true;
}
struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
struct ggml_threadpool * threadpool = ggml_threadpool_new_fn(&tpp);
if (!threadpool) {
LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
exit(1);
LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
return 1;
}
llama_attach_threadpool(ctx, threadpool, threadpool_batch);
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
}
// auto enable conversation mode if chat template is available
const bool has_chat_template = chat_templates.has_explicit_template && chat_templates.template_default;
if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) {
if (has_chat_template) {
LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__);
params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
} else {
params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
}
}
// in case user force-activate conversation mode (via -cnv) without proper chat template, we show a warning
if (params.conversation_mode && !has_chat_template) {
LOG_WRN("%s: chat template is not available or is not supported. This may cause the model to output suboptimal responses\n", __func__);
}
// print chat template example in conversation mode
if (params.conversation) {
if (params.conversation_mode) {
if (params.enable_chat_template) {
LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(*chat_templates.template_default, params.use_jinja).c_str());
} else {
LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
}
}
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
LOG_INF("\n");
}
std::string path_session = params.path_prompt_cache;
std::vector<llama_token> session_tokens;
if (!path_session.empty()) {
LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
LOG_INF("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
if (!file_exists(path_session)) {
LOG_TEE("%s: session file does not exist, will create.\n", __func__);
LOG_INF("%s: session file does not exist, will create.\n", __func__);
} else if (file_is_empty(path_session)) {
LOG_TEE("%s: The session file is empty. A new session will be initialized.\n", __func__);
LOG_INF("%s: The session file is empty. A new session will be initialized.\n", __func__);
} else {
// The file exists and is not empty
session_tokens.resize(n_ctx);
size_t n_token_count_out = 0;
if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
LOG_ERR("%s: failed to load session file '%s'\n", __func__, path_session.c_str());
return 1;
}
session_tokens.resize(n_token_count_out);
LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
LOG_INF("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
}
}
const bool add_bos = llama_add_bos_token(model);
const bool add_bos = llama_vocab_get_add_bos(vocab) && !params.use_jinja;
if (!llama_model_has_encoder(model)) {
GGML_ASSERT(!llama_add_eos_token(model));
GGML_ASSERT(!llama_vocab_get_add_eos(vocab));
}
LOG("add_bos: %d\n", add_bos);
LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
std::vector<llama_token> embd_inp;
auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) {
common_chat_msg new_msg{role, content, {}};
auto formatted = common_chat_format_single(*chat_templates.template_default, chat_msgs, new_msg, role == "user", g_params->use_jinja);
chat_msgs.push_back({role, content, {}});
LOG_DBG("formatted: '%s'\n", formatted.c_str());
return formatted;
};
{
auto prompt = (params.conversation && params.enable_chat_template && !params.prompt.empty())
? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
auto prompt = (params.conversation_mode && params.enable_chat_template)
// format the system prompt in conversation mode (fallback to default if empty)
? chat_add_and_format("system", params.prompt.empty() ? DEFAULT_SYSTEM_MESSAGE : params.prompt)
// otherwise use the prompt as is
: params.prompt;
if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n");
embd_inp = ::llama_tokenize(ctx, prompt, true, true);
LOG_DBG("tokenize the prompt\n");
embd_inp = common_tokenize(ctx, prompt, true, true);
} else {
LOG("use session tokens\n");
LOG_DBG("use session tokens\n");
embd_inp = session_tokens;
}
LOG("prompt: \"%s\"\n", log_tostr(prompt));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_DBG("prompt: \"%s\"\n", prompt.c_str());
LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
}
// Should not run without any tokens
if (embd_inp.empty()) {
if (add_bos) {
embd_inp.push_back(llama_token_bos(model));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
embd_inp.push_back(llama_vocab_bos(vocab));
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
} else {
LOG_TEE("error: input is empty\n");
LOG_ERR("input is empty\n");
return -1;
}
}
// Tokenize negative prompt
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
return 1;
}
@ -348,29 +316,28 @@ int main(int argc, char ** argv) {
n_matching_session_tokens++;
}
if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
LOG_TEE("%s: using full prompt from session file\n", __func__);
LOG_INF("%s: using full prompt from session file\n", __func__);
} else if (n_matching_session_tokens >= embd_inp.size()) {
LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
LOG_INF("%s: session file has exact match for prompt!\n", __func__);
} else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
__func__, n_matching_session_tokens, embd_inp.size());
LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
__func__, n_matching_session_tokens, embd_inp.size());
} else {
LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
__func__, n_matching_session_tokens, embd_inp.size());
LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
__func__, n_matching_session_tokens, embd_inp.size());
}
// remove any "future" tokens that we might have inherited from the previous session
llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
}
LOGLN(
"recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu",
log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
// if we will use the cache for the full prompt without reaching the end of the cache, force
// reevaluation of the last token to recalculate the cached logits
if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
session_tokens.resize(embd_inp.size() - 1);
}
@ -382,7 +349,7 @@ int main(int argc, char ** argv) {
params.n_keep += add_bos; // always keep the BOS token
}
if (params.conversation) {
if (params.conversation_mode) {
params.interactive_first = true;
}
@ -392,21 +359,20 @@ int main(int argc, char ** argv) {
}
if (params.verbose_prompt) {
LOG_TEE("\n");
LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (params.n_keep > add_bos) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
LOG_CNT("'\n");
}
LOG_TEE("\n");
LOG_INF("\n");
}
// ctrl+C handling
@ -426,54 +392,56 @@ int main(int argc, char ** argv) {
}
if (params.interactive) {
LOG_TEE("%s: interactive mode on.\n", __func__);
LOG_INF("%s: interactive mode on.\n", __func__);
if (!params.antiprompt.empty()) {
for (const auto & antiprompt : params.antiprompt) {
LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
LOG_INF("Reverse prompt: '%s'\n", antiprompt.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
auto tmp = common_tokenize(ctx, antiprompt, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
}
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
LOG_INF("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
auto tmp = common_tokenize(ctx, params.input_prefix, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
auto tmp = common_tokenize(ctx, params.input_suffix, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
}
smpl = gpt_sampler_init(model, sparams);
smpl = common_sampler_init(model, sparams);
if (!smpl) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
return 1;
}
LOG_TEE("sampling params: \n%s\n", sparams.print().c_str());
LOG_TEE(" sampler constr: \n%s\n", gpt_sampler_print(smpl).c_str());
LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
LOG_INF("sampler seed: %u\n", common_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", common_sampler_print(smpl).c_str());
LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
// group-attention state
// number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
@ -487,9 +455,9 @@ int main(int argc, char ** argv) {
GGML_ASSERT(ga_w % ga_n == 0 && "grp_attn_w must be a multiple of grp_attn_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
}
LOG_TEE("\n\n");
LOG_INF("\n");
if (params.interactive) {
const char * control_message;
@ -501,11 +469,15 @@ int main(int argc, char ** argv) {
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
LOG_TEE("== Running in interactive mode. ==\n");
LOG_INF("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
LOG_INF( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
LOG_INF( "%s", control_message);
if (params.conversation_mode && params.enable_chat_template && params.prompt.empty()) {
LOG_INF( " - Using default system message. To change it, set a different value via -p PROMPT or -f FILE argument.\n");
}
LOG_INF("\n");
is_interacting = params.interactive_first;
}
@ -531,26 +503,28 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
// tokenized antiprompts
std::vector<std::vector<llama_token>> antiprompt_ids;
// single-token antiprompts
std::vector<llama_token> antiprompt_token;
antiprompt_ids.reserve(params.antiprompt.size());
for (const std::string & antiprompt : params.antiprompt) {
antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
auto ids = ::common_tokenize(ctx, antiprompt, false, true);
if (ids.size() == 1) {
antiprompt_token.push_back(ids[0]);
}
}
if (llama_model_has_encoder(model)) {
int enc_input_size = embd_inp.size();
llama_token * enc_input_buf = embd_inp.data();
if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
decoder_start_token_id = llama_token_bos(model);
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = llama_vocab_bos(vocab);
}
embd_inp.clear();
@ -570,9 +544,8 @@ int main(int argc, char ** argv) {
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
if (ga_n == 1) {
@ -580,16 +553,22 @@ int main(int argc, char ** argv) {
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() >= n_ctx) {
if (!params.ctx_shift){
LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
break;
}
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
@ -597,11 +576,11 @@ int main(int argc, char ** argv) {
n_past -= n_discard;
LOG("after swap: n_past = %d\n", n_past);
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG("clear session path\n");
LOG_DBG("clear session path\n");
path_session.clear();
}
} else {
@ -611,10 +590,10 @@ int main(int argc, char ** argv) {
const int bd = (ga_w/ga_n)*(ga_n - 1);
const int dd = (ga_w/ga_n) - ib*bd - ga_w;
LOG("\n");
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
LOG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
LOG_DBG("\n");
LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
LOG_DBG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd);
llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);
@ -624,7 +603,7 @@ int main(int argc, char ** argv) {
ga_i += ga_w/ga_n;
LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
LOG_DBG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
}
}
@ -656,19 +635,19 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
LOG_DBG("n_past = %d\n", n_past);
// Display total tokens alongside total time
if (params.n_print > 0 && n_past % params.n_print == 0) {
LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
LOG_DBG("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
}
}
@ -686,14 +665,14 @@ int main(int argc, char ** argv) {
need_to_save_session = false;
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
LOG("saved session to %s\n", path_session.c_str());
LOG_DBG("saved session to %s\n", path_session.c_str());
}
const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
const llama_token id = common_sampler_sample(smpl, ctx, -1);
gpt_sampler_accept(smpl, id, /* apply_grammar= */ true);
common_sampler_accept(smpl, id, /* accept_grammar= */ true);
// LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(id);
@ -703,16 +682,16 @@ int main(int argc, char ** argv) {
// decrement remaining sampling budget
--n_remain;
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
while ((int) embd_inp.size() > n_consumed) {
embd.push_back(embd_inp[n_consumed]);
// push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules
gpt_sampler_accept(smpl, embd_inp[n_consumed], /* apply_grammar= */ false);
common_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
@ -724,10 +703,10 @@ int main(int argc, char ** argv) {
// display text
if (input_echo && display) {
for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id, params.special);
const std::string token_str = common_token_to_piece(ctx, id, params.special);
// Console/Stream Output
fprintf(stdout, "%s", token_str.c_str());
LOG("%s", token_str.c_str());
// Record Displayed Tokens To Log
// Note: Generated tokens are created one by one hence this check
@ -739,8 +718,6 @@ int main(int argc, char ** argv) {
output_tokens.push_back(id);
output_ss << token_str;
}
fflush(stdout);
}
}
@ -755,7 +732,7 @@ int main(int argc, char ** argv) {
// check for reverse prompt in the last n_prev tokens
if (!params.antiprompt.empty()) {
const int n_prev = 32;
const std::string last_output = gpt_sampler_prev_str(smpl, ctx, n_prev);
const std::string last_output = common_sampler_prev_str(smpl, ctx, n_prev);
is_antiprompt = false;
// Check if each of the reverse prompts appears at the end of the output.
@ -777,64 +754,61 @@ int main(int argc, char ** argv) {
}
// check for reverse prompt using special tokens
llama_token last_token = gpt_sampler_last(smpl);
for (std::vector<llama_token> ids : antiprompt_ids) {
if (ids.size() == 1 && last_token == ids[0]) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
break;
llama_token last_token = common_sampler_last(smpl);
if (std::find(antiprompt_token.begin(), antiprompt_token.end(), last_token) != antiprompt_token.end()) {
if (params.interactive) {
is_interacting = true;
}
is_antiprompt = true;
}
if (is_antiprompt) {
LOG("found antiprompt: %s\n", last_output.c_str());
LOG_DBG("found antiprompt: %s\n", last_output.c_str());
}
}
// deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
LOG("found an EOG token\n");
if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) {
LOG_DBG("found an EOG token\n");
if (params.interactive) {
if (!params.antiprompt.empty()) {
// tokenize and inject first reverse prompt
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
const auto first_antiprompt = common_tokenize(ctx, params.antiprompt.front(), false, true);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
is_antiprompt = true;
}
if (params.enable_chat_template) {
chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
chat_add_and_format("assistant", assistant_ss.str());
}
is_interacting = true;
printf("\n");
LOG("\n");
}
}
// if current token is not EOG, we add it to current assistant message
if (params.conversation) {
const auto id = gpt_sampler_last(smpl);
assistant_ss << llama_token_to_piece(ctx, id, false);
if (params.conversation_mode) {
const auto id = common_sampler_last(smpl);
assistant_ss << common_token_to_piece(ctx, id, false);
}
if (n_past > 0 && is_interacting) {
LOG("waiting for user input\n");
LOG_DBG("waiting for user input\n");
if (params.conversation) {
printf("\n> ");
if (params.conversation_mode) {
LOG("\n> ");
}
if (params.input_prefix_bos) {
LOG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_vocab_bos(vocab));
}
std::string buffer;
if (!params.input_prefix.empty() && !params.conversation) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
printf("%s", params.input_prefix.c_str());
if (!params.input_prefix.empty() && !params.conversation_mode) {
LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
LOG("%s", params.input_prefix.c_str());
}
// color user input only
@ -856,12 +830,12 @@ int main(int argc, char ** argv) {
// Entering a empty line lets the user pass control back
if (buffer.length() > 1) {
// append input suffix if any
if (!params.input_suffix.empty() && !params.conversation) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
printf("%s", params.input_suffix.c_str());
if (!params.input_suffix.empty() && !params.conversation_mode) {
LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
LOG("%s", params.input_suffix.c_str());
}
LOG("buffer: '%s'\n", buffer.c_str());
LOG_DBG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
@ -869,21 +843,21 @@ int main(int argc, char ** argv) {
string_process_escapes(buffer);
}
bool format_chat = params.conversation && params.enable_chat_template;
bool format_chat = params.conversation_mode && params.enable_chat_template;
std::string user_inp = format_chat
? chat_add_and_format(model, chat_msgs, "user", std::move(buffer))
? chat_add_and_format("user", std::move(buffer))
: std::move(buffer);
// TODO: one inconvenient of current chat template implementation is that we can't distinguish between user input and special tokens (prefix/postfix)
const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
const auto line_inp = ::llama_tokenize(ctx, user_inp, false, format_chat);
const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
const auto line_pfx = common_tokenize(ctx, params.input_prefix, false, true);
const auto line_inp = common_tokenize(ctx, user_inp, false, format_chat);
const auto line_sfx = common_tokenize(ctx, params.input_suffix, false, true);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
// if user stop generation mid-way, we must add EOT to finish model's last response
if (need_insert_eot && format_chat) {
llama_token eot = llama_token_eot(model);
embd_inp.push_back(eot == -1 ? llama_token_eos(model) : eot);
llama_token eot = llama_vocab_eot(vocab);
embd_inp.push_back(eot == LLAMA_TOKEN_NULL ? llama_vocab_eos(vocab) : eot);
need_insert_eot = false;
}
@ -894,16 +868,16 @@ int main(int argc, char ** argv) {
for (size_t i = original_size; i < embd_inp.size(); ++i) {
const llama_token token = embd_inp[i];
output_tokens.push_back(token);
output_ss << llama_token_to_piece(ctx, token);
output_ss << common_token_to_piece(ctx, token);
}
// reset assistant message
assistant_ss.str("");
n_remain -= line_inp.size();
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
LOG_DBG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
@ -911,15 +885,15 @@ int main(int argc, char ** argv) {
if (n_past > 0) {
if (is_interacting) {
gpt_sampler_reset(smpl);
common_sampler_reset(smpl);
}
is_interacting = false;
}
}
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
LOG_TEE(" [end of text]\n");
if (!embd.empty() && llama_vocab_is_eog(vocab, embd.back()) && !(params.interactive)) {
LOG(" [end of text]\n");
break;
}
@ -932,27 +906,19 @@ int main(int argc, char ** argv) {
}
if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
}
LOG_TEE("\n");
gpt_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
LOG("\n\n");
common_perf_print(ctx, smpl);
gpt_sampler_free(smpl);
llama_free(ctx);
llama_free_model(model);
common_sampler_free(smpl);
llama_backend_free();
ggml_threadpool_free(threadpool);
ggml_threadpool_free(threadpool_batch);
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
ggml_threadpool_free_fn(threadpool);
ggml_threadpool_free_fn(threadpool_batch);
return 0;
}

2
examples/parallel/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-parallel)
add_executable(${TARGET} parallel.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

117
examples/parallel/parallel.cpp
View file

@ -1,7 +1,10 @@
// A basic application simulating a server with multiple clients.
// The clients submit requests to the server and they are processed in parallel.
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cmath>
@ -51,7 +54,7 @@ static std::vector<std::string> k_prompts = {
struct client {
~client() {
if (smpl) {
gpt_sampler_free(smpl);
common_sampler_free(smpl);
}
}
@ -72,7 +75,7 @@ struct client {
std::string prompt;
std::string response;
struct gpt_sampler * smpl = nullptr;
struct common_sampler * smpl = nullptr;
};
static void print_date_time() {
@ -81,7 +84,9 @@ static void print_date_time() {
char buffer[80];
strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
LOG_INF("\n");
LOG_INF("\033[35mrun parameters as of %s\033[0m\n", buffer);
LOG_INF("\n");
}
// Define a split string function to ...
@ -98,13 +103,14 @@ static std::vector<std::string> split_string(const std::string& input, char deli
int main(int argc, char ** argv) {
srand(1234);
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
return 1;
}
common_init();
// number of simultaneous "clients" to simulate
const int32_t n_clients = params.n_parallel;
@ -119,41 +125,36 @@ int main(int argc, char ** argv) {
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("parallel", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
// load the target model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
const llama_vocab * vocab = llama_model_get_vocab(model);
// load the prompts from an external file if there are any
if (params.prompt.empty()) {
printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
LOG_INF("\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
} else {
// Output each line of the input params.prompts vector and copy to k_prompts
int index = 0;
printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
LOG_INF("\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
std::vector<std::string> prompts = split_string(params.prompt, '\n');
for (const auto& prompt : prompts) {
k_prompts.resize(index + 1);
k_prompts[index] = prompt;
index++;
printf("%3d prompt: %s\n", index, prompt.c_str());
LOG_INF("%3d prompt: %s\n", index, prompt.c_str());
}
}
fprintf(stderr, "\n\n");
fflush(stderr);
LOG_INF("\n\n");
const int n_ctx = llama_n_ctx(ctx);
@ -161,11 +162,11 @@ int main(int argc, char ** argv) {
for (size_t i = 0; i < clients.size(); ++i) {
auto & client = clients[i];
client.id = i;
client.smpl = gpt_sampler_init(model, params.sparams);
client.smpl = common_sampler_init(model, params.sampling);
}
std::vector<llama_token> tokens_system;
tokens_system = ::llama_tokenize(ctx, k_system, true);
tokens_system = common_tokenize(ctx, k_system, true);
const int32_t n_tokens_system = tokens_system.size();
llama_seq_id g_seq_id = 0;
@ -182,19 +183,19 @@ int main(int argc, char ** argv) {
const auto t_main_start = ggml_time_us();
LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_TEE("\n");
LOG_INF("%s: Simulating parallel requests from clients:\n", __func__);
LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_INF("\n");
{
LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
for (int32_t i = 0; i < n_tokens_system; ++i) {
llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
common_batch_add(batch, tokens_system[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -203,18 +204,18 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("\n");
LOG_INF("\n");
}
LOG_TEE("Processing requests ...\n\n");
LOG_INF("Processing requests ...\n\n");
while (true) {
if (dump_kv_cache) {
llama_kv_cache_view_update(ctx, &kvc_view);
llama_kv_cache_dump_view_seqs(kvc_view, 40);
common_kv_cache_dump_view_seqs(kvc_view, 40);
}
llama_batch_clear(batch);
common_batch_clear(batch);
// decode any currently ongoing sequences
for (auto & client : clients) {
@ -224,7 +225,7 @@ int main(int argc, char ** argv) {
client.i_batch = batch.n_tokens;
llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
common_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
client.n_decoded += 1;
}
@ -237,7 +238,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("%s: clearing the KV cache\n", __func__);
LOG_INF("%s: clearing the KV cache\n", __func__);
}
// insert new sequences for decoding
@ -253,14 +254,14 @@ int main(int argc, char ** argv) {
client.prompt = client.input + "\nAssistant:";
client.response = "";
gpt_sampler_reset(client.smpl);
common_sampler_reset(client.smpl);
// do not prepend BOS because we have a system prompt!
std::vector<llama_token> tokens_prompt;
tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
tokens_prompt = common_tokenize(ctx, client.prompt, false);
for (size_t i = 0; i < tokens_prompt.size(); ++i) {
llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
common_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
}
// extract the logits only for the last token
@ -272,7 +273,7 @@ int main(int argc, char ** argv) {
client.n_decoded = 0;
client.i_batch = batch.n_tokens - 1;
LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
LOG_INF("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
g_seq_id += 1;
@ -309,18 +310,17 @@ int main(int argc, char ** argv) {
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
};
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
if (n_batch == 1 || ret < 0) {
// if you get here, it means the KV cache is full - try increasing it via the context size
LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
LOG_ERR("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
return 1;
}
LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
LOG_ERR("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
n_cache_miss += 1;
@ -331,7 +331,7 @@ int main(int argc, char ** argv) {
continue;
}
LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
LOG_DBG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
for (auto & client : clients) {
if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
@ -341,9 +341,9 @@ int main(int argc, char ** argv) {
//printf("client %d, seq %d, token %d, pos %d, batch %d\n",
// client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
const llama_token id = gpt_sampler_sample(client.smpl, ctx, client.i_batch - i);
const llama_token id = common_sampler_sample(client.smpl, ctx, client.i_batch - i);
gpt_sampler_accept(client.smpl, id, true);
common_sampler_accept(client.smpl, id, true);
if (client.n_decoded == 1) {
// start measuring generation time after the first token to make sure all concurrent clients
@ -351,7 +351,7 @@ int main(int argc, char ** argv) {
client.t_start_gen = ggml_time_us();
}
const std::string token_str = llama_token_to_piece(ctx, id);
const std::string token_str = common_token_to_piece(ctx, id);
client.response += token_str;
client.sampled = id;
@ -360,7 +360,7 @@ int main(int argc, char ** argv) {
// client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
if (client.n_decoded > 2 &&
(llama_token_is_eog(model, id) ||
(llama_vocab_is_eog(vocab, id) ||
(params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
client.response.find("User:") != std::string::npos ||
client.response.find('\n') != std::string::npos)) {
@ -376,7 +376,7 @@ int main(int argc, char ** argv) {
const auto t_main_end = ggml_time_us();
LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput: %s\n\033[35mResponse: %s\033[0m\n\n",
LOG_INF("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput: %s\n\033[35mResponse: %s\033[0m\n\n",
client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
(t_main_end - client.t_start_prompt) / 1e6,
(double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
@ -399,31 +399,28 @@ int main(int argc, char ** argv) {
print_date_time();
LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
if (params.prompt_file.empty()) {
params.prompt_file = "used built-in defaults";
}
LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
LOG_TEE("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str());
LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str());
LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Total speed (AVG): %6s speed: %5.2f t/s\n", "", (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Cache misses: %6d\n", n_cache_miss);
LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Total speed (AVG): %6s speed: %5.2f t/s\n", "", (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Cache misses: %6d\n", n_cache_miss);
LOG_TEE("\n");
LOG_INF("\n");
// TODO: print sampling/grammar timings for all clients
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

2
examples/passkey/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-passkey)
add_executable(${TARGET} passkey.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

101
examples/passkey/passkey.cpp
View file

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <cmath>
@ -7,23 +9,24 @@
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
params.n_junk = 250;
params.n_keep = 32;
params.i_pos = -1;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_PASSKEY, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
return 1;
}
common_init();
int n_junk = params.n_junk;
int n_keep = params.n_keep;
int n_grp = params.grp_attn_n;
@ -58,26 +61,28 @@ int main(int argc, char ** argv) {
// initialize the model
llama_model_params model_params = llama_model_params_from_gpt_params(params);
llama_model_params model_params = common_model_params_to_llama(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return 1;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
// initialize the context
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
llama_context_params ctx_params = common_context_params_to_llama(params);
ctx_params.n_ctx = llama_n_ctx_train(model)*n_grp + n_keep;
ctx_params.n_ctx = llama_model_n_ctx_train(model)*n_grp + n_keep;
GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
llama_context * ctx = llama_init_from_model(model, ctx_params);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return 1;
}
@ -89,10 +94,10 @@ int main(int argc, char ** argv) {
// tokenize the prompt
std::vector<llama_token> tokens_list;
tokens_list = ::llama_tokenize(ctx, params.prompt, true);
tokens_list = common_tokenize(ctx, params.prompt, true);
// tokenize the prefix and use it as a sink
const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
const int n_tokens_prefix = common_tokenize(ctx, prompt_prefix, true).size();
const int n_tokens_all = tokens_list.size();
@ -107,14 +112,14 @@ int main(int argc, char ** argv) {
const int n_batch = ctx_params.n_batch;
const int n_batch_grp = ctx_params.n_batch/n_grp;
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
LOG_INF("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
// print the prompt token-by-token
LOG_TEE("\n");
LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
LOG_TEE("prompt tokens: %d\n", n_tokens_all);
//LOG_TEE("prompt: %s\n", params.prompt.c_str());
LOG_INF("\n");
LOG_INF("prefix tokens: %d\n", n_tokens_prefix);
LOG_INF("prompt tokens: %d\n", n_tokens_all);
//LOG_INF("prompt: %s\n", params.prompt.c_str());
llama_batch batch = llama_batch_init(params.n_batch, 0, 1);
@ -134,10 +139,10 @@ int main(int argc, char ** argv) {
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
}
llama_batch_clear(batch);
common_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
}
if (i + n_batch >= n_tokens_all) {
@ -145,11 +150,11 @@ int main(int argc, char ** argv) {
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_INF("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
LOG_INF("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
if (i + n_batch >= n_tokens_all) {
break;
@ -159,7 +164,7 @@ int main(int argc, char ** argv) {
for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
const int n_discard = n_batch;
LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
@ -168,10 +173,10 @@ int main(int argc, char ** argv) {
n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
llama_batch_clear(batch);
common_batch_clear(batch);
for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
common_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
}
if (i + n_batch >= n_tokens_all) {
@ -179,18 +184,18 @@ int main(int argc, char ** argv) {
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
LOG_INF("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
}
{
const int n_discard = n_past - n_ctx + n_predict;
if (n_discard > 0) {
LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
@ -201,17 +206,16 @@ int main(int argc, char ** argv) {
}
}
LOG_TEE("\n");
LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
LOG_TEE("\n");
LOG_INF("\n");
LOG_INF("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
LOG_INF("\n");
// main loop
int n_cur = n_tokens_all;
int n_decode = 0;
LOG_TEE("%s", prompt_suffix.c_str());
fflush(stdout);
LOG_INF("%s", prompt_suffix.c_str());
const auto t_main_start = ggml_time_us();
@ -220,54 +224,51 @@ int main(int argc, char ** argv) {
{
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
llama_sampler_accept(smpl, new_token_id);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
LOG_TEE("\n");
if (llama_vocab_is_eog(vocab, new_token_id) || n_cur == n_len) {
LOG("\n");
break;
}
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
n_decode += 1;
// prepare the next batch
llama_batch_clear(batch);
common_batch_clear(batch);
// push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
common_batch_add(batch, new_token_id, n_past++, { 0 }, true);
}
n_cur += 1;
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
LOG("\n");
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
LOG("\n");
llama_sampler_free(smpl);
llama_batch_free(batch);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
llama_backend_free();

2
examples/perplexity/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-perplexity)
add_executable(${TARGET} perplexity.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

613
examples/perplexity/perplexity.cpp
View file

File diff suppressed because it is too large Load diff

2
examples/quantize-stats/CMakeLists.txt
View file

@ -3,4 +3,4 @@ add_executable(${TARGET} quantize-stats.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
target_include_directories(${TARGET} PRIVATE ../../common)
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

41
examples/quantize-stats/quantize-stats.cpp
View file

@ -1,7 +1,7 @@
#include "common.h"
#include "ggml.h"
#include "llama.h"
#include "llama-impl.h"
#include "llama-context.h"
#include "common.h"
#include <algorithm>
#include <cassert>
@ -9,11 +9,9 @@
#include <cmath>
#include <cstdio>
#include <cstring>
#include <map>
#include <numeric>
#include <regex>
#include <string>
#include <unordered_map>
#include <vector>
#include <thread>
#include <mutex>
@ -142,7 +140,7 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
}
static void test_roundtrip_on_chunk(
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits_t & qfns, bool use_reference,
const ggml_tensor * layer, int64_t offset, int64_t chunk_size, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
float * input_scratch, char * quantized_scratch, float * output_scratch, error_stats & stats
) {
if (layer->type == GGML_TYPE_F16) {
@ -156,7 +154,7 @@ static void test_roundtrip_on_chunk(
if (use_reference) {
qfns.from_float_ref(input_scratch, quantized_scratch, chunk_size);
} else {
qfns.from_float(input_scratch, quantized_scratch, chunk_size);
qfns_cpu.from_float(input_scratch, quantized_scratch, chunk_size);
}
qfns.to_float(quantized_scratch, output_scratch, chunk_size);
@ -166,7 +164,7 @@ static void test_roundtrip_on_chunk(
// Run quantization function for a single layer and update error stats
static void test_roundtrip_on_layer(
std::string & name, bool print_layer_stats, const ggml_type_traits_t & qfns, bool use_reference,
std::string & name, bool print_layer_stats, const ggml_type_traits & qfns, const ggml_type_traits_cpu & qfns_cpu, bool use_reference,
const ggml_tensor * layer, std::vector<float> & input_scratch, std::vector<char> & quantized_scratch,
std::vector<float> & output_scratch, error_stats & total_error, int max_thread = 0
) {
@ -187,13 +185,13 @@ static void test_roundtrip_on_layer(
int num_chunks = (nelements + chunk_size - 1)/chunk_size;
if (num_chunks < 2 || max_thread < 2) {
test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
test_roundtrip_on_chunk(layer, 0, nelements, qfns, qfns_cpu, use_reference, input_scratch_ptr, quantized_scratch.data(),
output_scratch.data(), print_layer_stats ? layer_error : total_error);
} else {
auto & stats = print_layer_stats ? layer_error : total_error;
std::mutex mutex;
uint64_t counter = 0;
auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
auto compute = [&mutex, &counter, &stats, &qfns, &qfns_cpu, nelements, layer, use_reference, input_scratch_ptr,
&quantized_scratch, &output_scratch, chunk_size] () {
error_stats local_stats {};
while (true) {
@ -205,7 +203,7 @@ static void test_roundtrip_on_layer(
}
lock.unlock();
uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
test_roundtrip_on_chunk(layer, offset, chunk, qfns, qfns_cpu, use_reference, input_scratch_ptr + offset,
quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
}
};
@ -311,7 +309,7 @@ int main(int argc, char ** argv) {
auto mparams = llama_model_default_params();
mparams.use_mlock = false;
model = llama_load_model_from_file(params.model.c_str(), mparams);
model = llama_model_load_from_file(params.model.c_str(), mparams);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
@ -321,22 +319,22 @@ int main(int argc, char ** argv) {
auto cparams = llama_context_default_params();
cparams.n_ctx = 256;
ctx = llama_new_context_with_model(model, cparams);
ctx = llama_init_from_model(model, cparams);
if (ctx == NULL) {
fprintf(stderr, "%s: error: failed to create context with model '%s'\n", __func__, params.model.c_str());
llama_free_model(model);
llama_model_free(model);
return 1;
}
}
const auto &tensors = llama_internal_get_tensor_map(ctx);
const auto & tensors = llama_internal_get_tensor_map(ctx);
// check layer tensors
int included_layers = 0;
int64_t max_nelements = 0;
bool is_f16 = false;
for (const auto& kv_tensor : tensors) {
for (const auto & kv_tensor : tensors) {
if (!layer_included(params, kv_tensor.first)) {
continue;
}
@ -349,7 +347,7 @@ int main(int argc, char ** argv) {
fprintf(stderr, "%s: error: Quantization should be tested with a float model, "
"this model contains already quantized layers (%s is type %d)\n", __func__, kv_tensor.first.c_str(), kv_tensor.second->type);
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
return 1;
}
included_layers++;
@ -371,8 +369,9 @@ int main(int argc, char ** argv) {
if (!params.include_types.empty() && std::find(params.include_types.begin(), params.include_types.end(), i) == params.include_types.end()) {
continue;
}
ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
if (qfns.from_float && qfns.to_float) {
const auto * qfns = ggml_get_type_traits(type);
const auto * qfns_cpu = ggml_get_type_traits_cpu(type);
if (qfns_cpu->from_float && qfns->to_float) {
if (params.verbose) {
printf("testing %s ...\n", ggml_type_name(type));
}
@ -381,7 +380,7 @@ int main(int argc, char ** argv) {
error_stats global_stats {};
for (const auto& kv_tensor : tensors) {
for (const auto & kv_tensor : tensors) {
if (!layer_included(params, kv_tensor.first)) {
continue;
}
@ -393,7 +392,7 @@ int main(int argc, char ** argv) {
test_roundtrip_on_layer(
layer_name,
params.per_layer_stats,
qfns,
*qfns, *qfns_cpu,
params.reference,
kv_tensor.second,
input_scratch,
@ -410,7 +409,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
llama_model_free(model);
// report timing
{
const int64_t t_main_end_us = ggml_time_us();

4
examples/quantize/CMakeLists.txt
View file

@ -1,6 +1,6 @@
set(TARGET llama-quantize)
add_executable(${TARGET} quantize.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_include_directories(${TARGET} PRIVATE ../../common)
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

2
examples/quantize/README.md
View file

@ -81,7 +81,7 @@ Several quantization methods are supported. They differ in the resulting model d
- [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
- [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
- [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
- [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
- [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
- [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
- [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
- [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)

34
examples/quantize/quantize.cpp
View file

@ -1,5 +1,6 @@
#include "common.h"
#include "llama.h"
#include "gguf.h"
#include <cstdio>
#include <cstring>
@ -47,9 +48,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 5.33G, +0.0569 ppl @ Llama-3-8B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 6.14G, +0.0217 ppl @ Llama-3-8B", },
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 7.96G, +0.0026 ppl @ Llama-3-8B", },
{ "Q4_0_4_4", LLAMA_FTYPE_MOSTLY_Q4_0_4_4, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_4_8", LLAMA_FTYPE_MOSTLY_Q4_0_4_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "Q4_0_8_8", LLAMA_FTYPE_MOSTLY_Q4_0_8_8, " 4.34G, +0.4685 ppl @ Llama-3-8B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "14.00G, +0.0020 ppl @ Mistral-7B", },
{ "BF16", LLAMA_FTYPE_MOSTLY_BF16, "14.00G, -0.0050 ppl @ Mistral-7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
@ -67,6 +65,16 @@ static const char * const LLM_KV_IMATRIX_DATASET = "imatrix.dataset";
static const char * const LLM_KV_IMATRIX_CHUNK_COUNT = "imatrix.chunk_count";
static const char * const LLM_KV_IMATRIX_CHUNK_SIZE = "imatrix.chunk_size";
static bool striequals(const char * a, const char * b) {
while (*a && *b) {
if (std::tolower(*a) != std::tolower(*b)) {
return false;
}
a++; b++;
}
return *a == *b;
}
static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
std::string ftype_str;
@ -74,7 +82,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
ftype_str.push_back(std::toupper(ch));
}
for (auto & it : QUANT_OPTIONS) {
if (it.name == ftype_str) {
if (striequals(it.name.c_str(), ftype_str.c_str())) {
ftype = it.ftype;
ftype_str_out = it.name;
return true;
@ -276,15 +284,15 @@ static int prepare_imatrix(const std::string & imatrix_file,
}
static ggml_type parse_ggml_type(const char * arg) {
ggml_type result = GGML_TYPE_COUNT;
for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
auto type = ggml_type(j);
for (int i = 0; i < GGML_TYPE_COUNT; ++i) {
auto type = (ggml_type)i;
const auto * name = ggml_type_name(type);
if (name && strcmp(arg, name) == 0) {
result = type; break;
if (name && striequals(name, arg)) {
return type;
}
}
return result;
fprintf(stderr, "%s: invalid ggml_type '%s'\n", __func__, arg);
return GGML_TYPE_COUNT;
}
int main(int argc, char ** argv) {
@ -305,12 +313,18 @@ int main(int argc, char ** argv) {
} else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
if (arg_idx < argc-1) {
params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
if (params.output_tensor_type == GGML_TYPE_COUNT) {
usage(argv[0]);
}
} else {
usage(argv[0]);
}
} else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
if (arg_idx < argc-1) {
params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
if (params.token_embedding_type == GGML_TYPE_COUNT) {
usage(argv[0]);
}
} else {
usage(argv[0]);
}

4
examples/quantize/tests.sh
View file

@ -47,7 +47,7 @@ echo PASS
echo
# 3a. Test the requanted model is loading properly
$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --n-predict 32
echo PASS
echo
@ -57,7 +57,7 @@ echo PASS
echo
# 4b. Test the requanted model is loading properly
$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --n-predict 32
$MAIN -no-cnv --model $WORK_PATH/ggml-model-requant-merge.gguf --n-predict 32
echo PASS
echo

2
examples/retrieval/CMakeLists.txt
View file

@ -2,4 +2,4 @@ set(TARGET llama-retrieval)
add_executable(${TARGET} retrieval.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

112
examples/retrieval/retrieval.cpp
View file

@ -1,13 +1,16 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <fstream>
#include <iostream> // TODO: remove me
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
LOG("\n");
}
struct chunk {
@ -16,7 +19,7 @@ struct chunk {
// original file position
size_t filepos;
// original text data
std::string textdata = "";
std::string textdata;
// tokenized text data
std::vector<llama_token> tokens;
// embedding
@ -30,14 +33,14 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
std::ifstream f(filename.c_str());
if (!f.is_open()) {
fprintf(stderr, "Error: could not open file %s\n", filename.c_str());
LOG_ERR("could not open file %s\n", filename.c_str());
return chunks;
}
chunk current_chunk;
char buffer[1024];
int64_t filepos = 0;
std::string current = "";
std::string current;
while (f.read(buffer, 1024)) {
current += std::string(buffer, f.gcount());
size_t pos;
@ -74,7 +77,7 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
size_t n_tokens = tokens.size();
for (size_t i = 0; i < n_tokens; i++) {
llama_batch_add(batch, tokens[i], i, { seq_id }, true);
common_batch_add(batch, tokens[i], i, { seq_id }, true);
}
}
@ -83,9 +86,9 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
LOG_ERR("%s : failed to decode\n", __func__);
}
for (int i = 0; i < batch.n_tokens; i++) {
@ -98,42 +101,41 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
if (embd == NULL) {
embd = llama_get_embeddings_ith(ctx, i);
if (embd == NULL) {
fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
LOG_ERR("%s: failed to get embeddings for token %d\n", __func__, i);
continue;
}
}
float * out = output + batch.seq_id[i][0] * n_embd;
llama_embd_normalize(embd, out, n_embd);
common_embd_normalize(embd, out, n_embd, 2);
}
}
int main(int argc, char ** argv) {
gpt_params params;
common_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_RETRIEVAL, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
return 1;
}
common_init();
// For BERT models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch;
params.embedding = true;
if (params.chunk_size <= 0) {
fprintf(stderr, "chunk_size must be positive\n");
LOG_ERR("chunk_size must be positive\n");
return 1;
}
if (params.context_files.empty()) {
fprintf(stderr, "context_files must be specified\n");
LOG_ERR("context_files must be specified\n");
return 1;
}
print_build_info();
printf("processing files:\n");
LOG_INF("processing files:\n");
for (auto & context_file : params.context_files) {
printf("%s\n", context_file.c_str());
LOG_INF("%s\n", context_file.c_str());
}
std::vector<chunk> chunks;
@ -141,40 +143,42 @@ int main(int argc, char ** argv) {
std::vector<chunk> file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator);
chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
}
printf("Number of chunks: %ld\n", chunks.size());
LOG_INF("Number of chunks: %zu\n", chunks.size());
llama_backend_init();
llama_numa_init(params.numa);
// load the model
llama_init_result llama_init = llama_init_from_gpt_params(params);
common_init_result llama_init = common_init_from_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
llama_model * model = llama_init.model.get();
llama_context * ctx = llama_init.context.get();
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_ctx_train = llama_model_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
LOG_ERR("%s: pooling type NONE not supported\n", __func__);
return 1;
}
if (n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", common_params_get_system_info(params).c_str());
}
// max batch size
@ -183,15 +187,15 @@ int main(int argc, char ** argv) {
// tokenize the prompts and trim
for (auto & chunk : chunks) {
auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false);
auto inp = common_tokenize(ctx, chunk.textdata, true, false);
if (inp.size() > n_batch) {
fprintf(stderr, "%s: error: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
LOG_ERR("%s: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch);
return 1;
}
// add eos if not present
if (llama_token_eos(model) >= 0 && (inp.empty() || inp.back() != llama_token_eos(model))) {
inp.push_back(llama_token_eos(model));
if (llama_vocab_eos(vocab) >= 0 && (inp.empty() || inp.back() != llama_vocab_eos(vocab))) {
inp.push_back(llama_vocab_eos(vocab));
}
chunk.tokens = inp;
}
@ -199,12 +203,12 @@ int main(int argc, char ** argv) {
// tokenization stats
if (params.verbose_prompt) {
for (int i = 0; i < (int) chunks.size(); i++) {
fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
fprintf(stderr, "%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], common_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
}
fprintf(stderr, "\n\n");
LOG_INF("\n\n");
}
}
@ -213,7 +217,7 @@ int main(int argc, char ** argv) {
struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
// allocate output
const int n_embd = llama_n_embd(model);
const int n_embd = llama_model_n_embd(model);
std::vector<float> embeddings(n_chunks * n_embd, 0);
float * emb = embeddings.data();
@ -230,7 +234,7 @@ int main(int argc, char ** argv) {
if (batch.n_tokens + n_toks > n_batch) {
float * out = emb + p * n_embd;
batch_decode(ctx, batch, out, s, n_embd);
llama_batch_clear(batch);
common_batch_clear(batch);
p += s;
s = 0;
}
@ -256,22 +260,22 @@ int main(int argc, char ** argv) {
// start loop, receive query and return top k similar chunks based on cosine similarity
std::string query;
while (true) {
printf("Enter query: ");
LOG("Enter query: ");
std::getline(std::cin, query);
std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
std::vector<int32_t> query_tokens = common_tokenize(ctx, query, true);
batch_add_seq(query_batch, query_tokens, 0);
std::vector<float> query_emb(n_embd, 0);
batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
llama_batch_clear(query_batch);
common_batch_clear(query_batch);
// compute cosine similarities
{
std::vector<std::pair<int, float>> similarities;
for (int i = 0; i < n_chunks; i++) {
float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
float sim = common_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
similarities.push_back(std::make_pair(i, sim));
}
@ -280,23 +284,21 @@ int main(int argc, char ** argv) {
return a.second > b.second;
});
printf("Top %d similar chunks:\n", params.sparams.top_k);
for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
printf("filename: %s\n", chunks[similarities[i].first].filename.c_str());
printf("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
printf("similarity: %f\n", similarities[i].second);
printf("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
printf("--------------------\n");
LOG("Top %d similar chunks:\n", params.sampling.top_k);
for (int i = 0; i < std::min(params.sampling.top_k, (int) chunks.size()); i++) {
LOG("filename: %s\n", chunks[similarities[i].first].filename.c_str());
LOG("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
LOG("similarity: %f\n", similarities[i].second);
LOG("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
LOG("--------------------\n");
}
}
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
// clean up
llama_batch_free(query_batch);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
}

30
examples/rpc/README.md
View file

@ -10,20 +10,21 @@ This can be used for distributed LLM inference with `llama.cpp` in the following
```mermaid
flowchart TD
rpcb---|TCP|srva
rpcb---|TCP|srvb
rpcb-.-|TCP|srvn
rpcb<-->|TCP|srva
rpcb<-->|TCP|srvb
rpcb<-.->|TCP|srvn
subgraph hostn[Host N]
srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"]
end
subgraph hostb[Host B]
srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"]
end
subgraph hosta[Host A]
srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"]
end
subgraph host[Main Host]
ggml[llama.cpp]---rpcb[RPC backend]
local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli]
ggml[llama-cli]<-->rpcb[RPC backend]
end
style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
```
@ -62,17 +63,12 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
On the main host build `llama.cpp` only with `-DGGML_RPC=ON`:
```bash
mkdir build-rpc
cd build-rpc
cmake .. -DGGML_RPC=ON
cmake --build . --config Release
```
Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options.
Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`:
```bash
$ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
```
This way you can offload model layers to both local and remote devices.

28
examples/rpc/rpc-server.cpp
View file

@ -1,3 +1,5 @@
#include "ggml-cpu.h"
#ifdef GGML_USE_CUDA
#include "ggml-cuda.h"
#endif
@ -6,6 +8,14 @@
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_VULKAN
#include "ggml-vulkan.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#include "ggml-rpc.h"
#ifdef _WIN32
# include <windows.h>
@ -79,6 +89,18 @@ static ggml_backend_t create_backend() {
if (!backend) {
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
}
#elif GGML_USE_VULKAN
fprintf(stderr, "%s: using Vulkan backend\n", __func__);
backend = ggml_backend_vk_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_vulkan_init() failed\n", __func__);
}
#elif GGML_USE_SYCL
fprintf(stderr, "%s: using SYCL backend\n", __func__);
backend = ggml_backend_sycl_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
}
#endif
// if there aren't GPU Backends fallback to CPU backend
@ -92,6 +114,10 @@ static ggml_backend_t create_backend() {
static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
#ifdef GGML_USE_CUDA
ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
#elif GGML_USE_VULKAN
ggml_backend_vk_get_device_memory(0, free_mem, total_mem);
#elif GGML_USE_SYCL
ggml_backend_sycl_get_device_memory(0, free_mem, total_mem);
#else
#ifdef _WIN32
MEMORYSTATUSEX status;
@ -139,7 +165,7 @@ int main(int argc, char * argv[]) {
get_backend_memory(&free_mem, &total_mem);
}
printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem);
ggml_backend_free(backend);
return 0;
}

5
examples/run/CMakeLists.txt Normal file
View file

@ -0,0 +1,5 @@
set(TARGET llama-run)
add_executable(${TARGET} run.cpp linenoise.cpp/linenoise.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)

50
examples/run/README.md Normal file
View file

@ -0,0 +1,50 @@
# llama.cpp/example/run
The purpose of this example is to demonstrate a minimal usage of llama.cpp for running models.
```bash
llama-run granite3-moe
```
```bash
Description:
Runs a llm
Usage:
llama-run [options] model [prompt]
Options:
-c, --context-size <value>
Context size (default: 2048)
-n, -ngl, --ngl <value>
Number of GPU layers (default: 0)
--temp <value>
Temperature (default: 0.8)
-v, --verbose, --log-verbose
Set verbosity level to infinity (i.e. log all messages, useful for debugging)
-h, --help
Show help message
Commands:
model
Model is a string with an optional prefix of
huggingface:// (hf://), ollama://, https:// or file://.
If no protocol is specified and a file exists in the specified
path, file:// is assumed, otherwise if a file does not exist in
the specified path, ollama:// is assumed. Models that are being
pulled are downloaded with .partial extension while being
downloaded and then renamed as the file without the .partial
extension when complete.
Examples:
llama-run llama3
llama-run ollama://granite-code
llama-run ollama://smollm:135m
llama-run hf://QuantFactory/SmolLM-135M-GGUF/SmolLM-135M.Q2_K.gguf
llama-run huggingface://bartowski/SmolLM-1.7B-Instruct-v0.2-GGUF/SmolLM-1.7B-Instruct-v0.2-IQ3_M.gguf
llama-run https://example.com/some-file1.gguf
llama-run some-file2.gguf
llama-run file://some-file3.gguf
llama-run --ngl 999 some-file4.gguf
llama-run --ngl 999 some-file5.gguf Hello World
```

26
examples/run/linenoise.cpp/LICENSE Normal file
View file

@ -0,0 +1,26 @@
Copyright (c) 2010-2014, Salvatore Sanfilippo <antirez at gmail dot com>
Copyright (c) 2010-2013, Pieter Noordhuis <pcnoordhuis at gmail dot com>
Copyright (c) 2025, Eric Curtin <ericcurtin17 at gmail dot com>
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
(INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

1350
examples/run/linenoise.cpp/linenoise.cpp Normal file
View file

File diff suppressed because it is too large Load diff

Some files were not shown because too many files have changed in this diff Show more