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

Merge branch 'master' of https://github.com/ggerganov/llama.cpp into ntkv2

Browse source
This commit is contained in:
cebtenzzre 2023-10-22 14:00:33 -04:00
commit 237f1e7912
59 changed files with 2070 additions and 5484 deletions
Download patch file Download diff file Expand all files Collapse all files

11
Makefile
View file

@ -1,7 +1,7 @@
# Define the default target now so that it is always the first target
BUILD_TARGETS = \
main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
simple batched batched-bench save-load-state server embd-input-test gguf llama-bench llava baby-llama beam-search \
simple batched batched-bench save-load-state server gguf llama-bench llava baby-llama beam-search \
speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
# Binaries only useful for tests
@ -545,7 +545,7 @@ llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h l
$(CXX) $(CXXFLAGS) -c $< -o $@
COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
COMMON_DEPS = $(COMMON_H_DEPS) common.o sampling.o
COMMON_DEPS = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.o
common.o: common/common.cpp $(COMMON_H_DEPS)
$(CXX) $(CXXFLAGS) -c $< -o $@
@ -608,13 +608,6 @@ save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.
server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

33
README.md
View file

@ -11,12 +11,8 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
### Hot topics
- ‼️ Breaking change: `rope_freq_base` and `rope_freq_scale` must be set to zero to use the model default values: [#3401](https://github.com/ggerganov/llama.cpp/pull/3401)
- Parallel decoding + continuous batching support added: [#3228](https://github.com/ggerganov/llama.cpp/pull/3228) \
**Devs should become familiar with the new API**
- Local Falcon 180B inference on Mac Studio
https://github.com/ggerganov/llama.cpp/assets/1991296/98abd4e8-7077-464c-ae89-aebabca7757e
- LLaVA support: https://github.com/ggerganov/llama.cpp/pull/3436
- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
----
@ -89,15 +85,17 @@ as the main playground for developing new features for the [ggml](https://github
- [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
- [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
- [X] [Baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) and its derivations (such as [baichuan-7b-sft](https://huggingface.co/hiyouga/baichuan-7b-sft))
- [X] [Aquila-7B](https://huggingface.co/BAAI/Aquila-7B) / [AquilaChat-7B](https://huggingface.co/BAAI/AquilaChat-7B)
- [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
- [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
- [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
**Bindings:**
@ -206,7 +204,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
## Usage
Here are the steps for the LLaMA-7B model.
Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
### Get the Code
@ -573,6 +571,18 @@ python3 convert.py models/7B/
When running the larger models, make sure you have enough disk space to store all the intermediate files.
### Running on Windows with prebuilt binaries
You will find prebuilt Windows binaries on the release page.
Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
```
.\main -m llama-2-7b.Q4_0.gguf -n 128
```
### Memory/Disk Requirements
As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
@ -952,7 +962,6 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /
- [main](./examples/main/README.md)
- [server](./examples/server/README.md)
- [embd-input](./examples/embd-input/README.md)
- [jeopardy](./examples/jeopardy/README.md)
- [BLIS](./docs/BLIS.md)
- [Performance troubleshooting](./docs/token_generation_performance_tips.md)

6
ci/run.sh
View file

@ -208,6 +208,8 @@ function gg_run_open_llama_3b_v2 {
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@ -296,6 +298,7 @@ function gg_sum_open_llama_3b_v2 {
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
@ -382,6 +385,8 @@ function gg_run_open_llama_7b_v2 {
(time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
(time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
(time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
function check_ppl {
qnt="$1"
ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
@ -470,6 +475,7 @@ function gg_sum_open_llama_7b_v2 {
gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
#gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"

70
common/common.cpp
View file

@ -107,7 +107,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
std::string arg;
gpt_params default_params;
const std::string arg_prefix = "--";
llama_sampling_params & sparams = params.sampling_params;
llama_sampling_params & sparams = params.sparams;
for (int i = 1; i < argc; i++) {
arg = argv[i];
@ -281,25 +281,26 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
sparams.repeat_last_n = std::stoi(argv[i]);
sparams.penalty_last_n = std::stoi(argv[i]);
sparams.n_prev = std::max(sparams.n_prev, sparams.penalty_last_n);
} else if (arg == "--repeat-penalty") {
if (++i >= argc) {
invalid_param = true;
break;
}
sparams.repeat_penalty = std::stof(argv[i]);
sparams.penalty_repeat = std::stof(argv[i]);
} else if (arg == "--frequency-penalty") {
if (++i >= argc) {
invalid_param = true;
break;
}
sparams.frequency_penalty = std::stof(argv[i]);
sparams.penalty_freq = std::stof(argv[i]);
} else if (arg == "--presence-penalty") {
if (++i >= argc) {
invalid_param = true;
break;
}
sparams.presence_penalty = std::stof(argv[i]);
sparams.penalty_present = std::stof(argv[i]);
} else if (arg == "--mirostat") {
if (++i >= argc) {
invalid_param = true;
@ -612,7 +613,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
params.grammar = argv[i];
sparams.grammar = argv[i];
} else if (arg == "--grammar-file") {
if (++i >= argc) {
invalid_param = true;
@ -627,7 +628,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
std::copy(
std::istreambuf_iterator<char>(file),
std::istreambuf_iterator<char>(),
std::back_inserter(params.grammar)
std::back_inserter(sparams.grammar)
);
#ifndef LOG_DISABLE_LOGS
// Parse args for logging parameters
@ -680,7 +681,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
}
void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
const llama_sampling_params & sparams = params.sampling_params;
const llama_sampling_params & sparams = params.sparams;
printf("usage: %s [options]\n", argv[0]);
printf("\n");
@ -718,10 +719,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
printf(" --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
printf(" --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
printf(" --typical N locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
printf(" --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.repeat_last_n);
printf(" --repeat-penalty N penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.repeat_penalty);
printf(" --presence-penalty N repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.presence_penalty);
printf(" --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.frequency_penalty);
printf(" --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
printf(" --repeat-penalty N penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)sparams.penalty_repeat);
printf(" --presence-penalty N repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_present);
printf(" --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)sparams.penalty_freq);
printf(" --mirostat N use Mirostat sampling.\n");
printf(" Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
printf(" (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", sparams.mirostat);
@ -873,6 +874,27 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
return cparams;
}
void llama_batch_clear(struct llama_batch & batch) {
batch.n_tokens = 0;
}
void llama_batch_add(
struct llama_batch & batch,
llama_token id,
llama_pos pos,
const std::vector<llama_seq_id> & seq_ids,
bool logits) {
batch.token [batch.n_tokens] = id;
batch.pos [batch.n_tokens] = pos,
batch.n_seq_id[batch.n_tokens] = seq_ids.size();
for (size_t i = 0; i < seq_ids.size(); ++i) {
batch.seq_id[batch.n_tokens][i] = seq_ids[i];
}
batch.logits [batch.n_tokens] = logits;
batch.n_tokens++;
}
std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
auto mparams = llama_model_params_from_gpt_params(params);
@ -910,7 +932,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
}
if (params.ignore_eos) {
params.sampling_params.logit_bias[llama_token_eos(lctx)] = -INFINITY;
params.sparams.logit_bias[llama_token_eos(lctx)] = -INFINITY;
}
{
@ -932,21 +954,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
std::vector<llama_token> llama_tokenize(
const struct llama_context * ctx,
const std::string & text,
bool add_bos) {
return llama_tokenize(llama_get_model(ctx), text, add_bos);
bool add_bos,
bool special) {
return llama_tokenize(llama_get_model(ctx), text, add_bos, special);
}
std::vector<llama_token> llama_tokenize(
const struct llama_model * model,
const std::string & text,
bool add_bos) {
bool add_bos,
bool special) {
// upper limit for the number of tokens
int n_tokens = text.length() + add_bos;
std::vector<llama_token> result(n_tokens);
n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
if (n_tokens < 0) {
result.resize(-n_tokens);
int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos);
int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
GGML_ASSERT(check == -n_tokens);
} else {
result.resize(n_tokens);
@ -1153,7 +1177,7 @@ std::string get_sortable_timestamp() {
void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
const llama_sampling_params & sparams = params.sampling_params;
const llama_sampling_params & sparams = params.sparams;
fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
@ -1208,8 +1232,8 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.frequency_penalty);
dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str());
fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
dump_string_yaml_multiline(stream, "grammar", sparams.grammar.c_str());
fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
fprintf(stream, "hellaswag_tasks: %zu # default: 400\n", params.hellaswag_tasks);
@ -1268,14 +1292,14 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.presence_penalty);
fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str());
fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.repeat_penalty);
fprintf(stream, "repeat_penalty: %f # default: 1.1\n", sparams.penalty_repeat);
fprintf(stream, "reverse_prompt:\n");
for (std::string ap : params.antiprompt) {

25
common/common.h
View file

@ -63,7 +63,7 @@ struct gpt_params {
int8_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED;
// // sampling parameters
struct llama_sampling_params sampling_params;
struct llama_sampling_params sparams;
std::string model = "models/7B/ggml-model-f16.gguf"; // model path
std::string model_draft = ""; // draft model for speculative decoding
@ -73,10 +73,10 @@ struct gpt_params {
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state
std::string input_prefix = ""; // string to prefix user inputs with
std::string input_suffix = ""; // string to suffix user inputs with
std::string grammar = ""; // optional BNF-like grammar to constrain sampling
std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
std::string logdir = ""; // directory in which to save YAML log files
// TODO: avoid tuple, use struct
std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
std::string lora_base = ""; // base model path for the lora adapter
@ -131,10 +131,23 @@ void process_escapes(std::string& input);
// Model utils
//
// TODO: avoid tuplue, use struct
std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & params);
struct llama_model_params llama_model_params_from_gpt_params (const gpt_params & params);
struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
// Batch utils
void llama_batch_clear(struct llama_batch & batch);
void llama_batch_add(
struct llama_batch & batch,
llama_token id,
llama_pos pos,
const std::vector<llama_seq_id> & seq_ids,
bool logits);
//
// Vocab utils
//
@ -144,12 +157,14 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
std::vector<llama_token> llama_tokenize(
const struct llama_context * ctx,
const std::string & text,
bool add_bos);
bool add_bos,
bool special = false);
std::vector<llama_token> llama_tokenize(
const struct llama_model * model,
const std::string & text,
bool add_bos);
bool add_bos,
bool special = false);
// tokenizes a token into a piece
// should work similar to Python's `tokenizer.id_to_piece`

2
common/grammar-parser.cpp
View file

@ -399,7 +399,7 @@ namespace grammar_parser {
void print_grammar(FILE * file, const parse_state & state) {
try {
std::map<uint32_t, std::string> symbol_id_names;
for (auto kv : state.symbol_ids) {
for (const auto & kv : state.symbol_ids) {
symbol_id_names[kv.second] = kv.first;
}
for (size_t i = 0, end = state.rules.size(); i < end; i++) {

101
common/log.h
View file

@ -579,38 +579,75 @@ inline std::string log_var_to_string_impl(const std::vector<int> & var)
return buf.str();
}
#define LOG_TOKENS_TOSTR_PRETTY(ctx, tokens) \
[&tokens, &ctx]() \
{ \
std::stringstream buf; \
buf << "[ "; \
\
bool first = true; \
for (const auto &token : tokens) \
{ \
if (!first) \
buf << ", "; \
else \
first = false; \
\
auto detokenized = llama_token_to_piece(ctx, token); \
\
detokenized.erase( \
std::remove_if( \
detokenized.begin(), \
detokenized.end(), \
[](const unsigned char c) { return !std::isprint(c); }), \
detokenized.end()); \
\
buf \
<< "'" << detokenized << "'" \
<< ":" << std::to_string(token); \
} \
buf << " ]"; \
\
return buf.str(); \
}() \
.c_str()
template <typename C, typename T>
inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (const auto &token : tokens)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, token);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "'" << detokenized << "'"
<< ":" << std::to_string(token);
}
buf << " ]";
return buf.str();
}
template <typename C, typename B>
inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (int i = 0; i < batch.n_tokens; ++i)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "\n" << std::to_string(i)
<< ":token '" << detokenized << "'"
<< ":pos " << std::to_string(batch.pos[i])
<< ":n_seq_id " << std::to_string(batch.n_seq_id[i])
<< ":seq_id " << std::to_string(batch.seq_id[i][0])
<< ":logits " << std::to_string(batch.logits[i]);
}
buf << " ]";
return buf.str();
}
#ifdef LOG_DISABLE_LOGS

256
common/sampling.cpp
View file

@ -1,113 +1,161 @@
#include "sampling.h"
llama_sampling_context::~llama_sampling_context() {
for (auto & it : sequence_contexts) {
if (it.second.grammar != NULL) {
llama_grammar_free(it.second.grammar);
it.second.grammar = NULL;
}
}
}
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
struct llama_sampling_context * result = new llama_sampling_context();
llama_sampling_context llama_sampling_context_init(
const struct gpt_params & params,
llama_grammar * grammar) {
llama_sampling_context result;
result->params = params;
result->grammar = nullptr;
// if there is a grammar, parse it
if (!params.grammar.empty()) {
result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (result->parsed_grammar.rules.empty()) {
fprintf(stderr, "%s: failed to parse grammar\n", __func__);
return nullptr;
}
std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
result->grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
}
result->prev.resize(params.n_prev);
result.params = params.sampling_params;
result.grammar = grammar;
return result;
}
// Note: Creates the context if it doesn't exist, so this always return something.
llama_sampler_sequence_context & llama_sampling_get_sequence_context(
llama_sampling_context & ctx_sampling,
const llama_seq_id seq) {
const auto it = ctx_sampling.sequence_contexts.find(seq);
if (it != ctx_sampling.sequence_contexts.end()) {
return it->second;
void llama_sampling_free(struct llama_sampling_context * ctx) {
if (ctx->grammar != NULL) {
llama_grammar_free(ctx->grammar);
}
llama_sampler_sequence_context new_ctx = {
2.0f * ctx_sampling.params.mirostat_tau,
ctx_sampling.grammar != NULL ? llama_grammar_copy(ctx_sampling.grammar) : NULL,
};
return ctx_sampling.sequence_contexts.insert({seq, new_ctx}).first->second;
delete ctx;
}
bool llama_sampling_context_reset(
llama_sampling_context & ctx_sampling,
const llama_seq_id seq) {
const auto it = ctx_sampling.sequence_contexts.find(seq);
if (it == ctx_sampling.sequence_contexts.end()) return false;
if (it->second.grammar != NULL) {
llama_grammar_free(it->second.grammar);
it->second.grammar = NULL;
void llama_sampling_reset(llama_sampling_context * ctx) {
if (ctx->grammar != NULL) {
llama_grammar_free(ctx->grammar);
}
ctx_sampling.sequence_contexts.erase(it);
return true;
if (!ctx->parsed_grammar.rules.empty()) {
std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
ctx->grammar = llama_grammar_init(
grammar_rules.data(),
grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
}
std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
ctx->cur.clear();
}
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
if (dst->grammar) {
llama_grammar_free(dst->grammar);
dst->grammar = nullptr;
}
if (src->grammar) {
dst->grammar = llama_grammar_copy(src->grammar);
}
dst->prev = src->prev;
}
llama_token llama_sampling_last(llama_sampling_context * ctx) {
return ctx->prev.back();
}
std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
const int size = ctx_sampling->prev.size();
n = std::min(n, size);
std::string result;
for (int i = size - n; i < size; i++) {
result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
}
return result;
}
std::string llama_sampling_print(const llama_sampling_params & params) {
char result[1024];
snprintf(result, sizeof(result),
"\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
"\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, typical_p = %.3f, temp = %.3f\n"
"\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp,
params.mirostat, params.mirostat_eta, params.mirostat_tau);
return std::string(result);
}
llama_token llama_sampling_sample(
struct llama_context * ctx,
struct llama_context * ctx_guidance,
struct llama_sampling_context & ctx_sampling,
const std::vector<llama_token> & last_tokens,
std::vector<llama_token_data> & candidates,
const int idx,
llama_seq_id seq) {
const int n_ctx = llama_n_ctx(ctx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
const int idx) {
const llama_sampling_params & params = ctx_sampling->params;
const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
const llama_sampling_params & params = ctx_sampling.params;
const float temp = params.temp;
const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k;
const float top_p = params.top_p;
const float tfs_z = params.tfs_z;
const float typical_p = params.typical_p;
const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
const float repeat_penalty = params.repeat_penalty;
const float alpha_presence = params.presence_penalty;
const float alpha_frequency = params.frequency_penalty;
const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
const float penalty_repeat = params.penalty_repeat;
const float penalty_freq = params.penalty_freq;
const float penalty_present = params.penalty_present;
const int mirostat = params.mirostat;
const float mirostat_tau = params.mirostat_tau;
const float mirostat_eta = params.mirostat_eta;
const bool penalize_nl = params.penalize_nl;
auto & prev = ctx_sampling->prev;
auto & cur = ctx_sampling->cur;
llama_token id = 0;
float * logits = llama_get_logits_ith(ctx, idx);
float * logits = llama_get_logits_ith(ctx_main, idx);
// Apply params.logit_bias map
// apply params.logit_bias map
for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
logits[it->first] += it->second;
}
candidates.clear();
cur.clear();
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
}
llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
llama_token_data_array cur_p = { cur.data(), cur.size(), false };
if (ctx_guidance) {
llama_sample_classifier_free_guidance(ctx, &cur_p, ctx_guidance, params.cfg_scale);
if (ctx_cfg) {
llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
}
// apply penalties
if (!last_tokens.empty()) {
const float nl_logit = logits[llama_token_nl(ctx)];
const int last_n_repeat = std::min(std::min((int)last_tokens.size(), repeat_last_n), n_ctx);
if (!prev.empty()) {
const float nl_logit = logits[llama_token_nl(ctx_main)];
llama_sample_repetition_penalty(ctx, &cur_p,
last_tokens.data() + last_tokens.size() - last_n_repeat,
last_n_repeat, repeat_penalty);
llama_sample_frequency_and_presence_penalties(ctx, &cur_p,
last_tokens.data() + last_tokens.size() - last_n_repeat,
last_n_repeat, alpha_frequency, alpha_presence);
llama_sample_repetition_penalties(ctx_main, &cur_p,
prev.data() + prev.size() - penalty_last_n,
penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
if (!penalize_nl) {
for (size_t idx = 0; idx < cur_p.size; idx++) {
if (cur_p.data[idx].id == llama_token_nl(ctx)) {
if (cur_p.data[idx].id == llama_token_nl(ctx_main)) {
cur_p.data[idx].logit = nl_logit;
break;
}
@ -115,52 +163,60 @@ llama_token llama_sampling_sample(
}
}
llama_sampler_sequence_context & ctx_seq = llama_sampling_get_sequence_context(ctx_sampling, seq);
if (ctx_seq.grammar != NULL) {
llama_sample_grammar(ctx, &cur_p, ctx_seq.grammar);
if (ctx_sampling->grammar != NULL) {
llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
}
if (temp <= 0) {
// Greedy sampling
id = llama_sample_token_greedy(ctx, &cur_p);
// greedy sampling
id = llama_sample_token_greedy(ctx_main, &cur_p);
} else {
if (mirostat == 1) {
const int mirostat_m = 100;
llama_sample_temp(ctx, &cur_p, temp);
id = llama_sample_token_mirostat(ctx, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_seq.mirostat_mu);
llama_sample_temp(ctx_main, &cur_p, temp);
id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
} else if (mirostat == 2) {
llama_sample_temp(ctx, &cur_p, temp);
id = llama_sample_token_mirostat_v2(ctx, &cur_p, mirostat_tau, mirostat_eta, &ctx_seq.mirostat_mu);
llama_sample_temp(ctx_main, &cur_p, temp);
id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
} else {
// Temperature sampling
// temperature sampling
size_t min_keep = std::max(1, params.n_probs);
llama_sample_top_k (ctx, &cur_p, top_k, min_keep);
llama_sample_tail_free (ctx, &cur_p, tfs_z, min_keep);
llama_sample_typical (ctx, &cur_p, typical_p, min_keep);
llama_sample_top_p (ctx, &cur_p, top_p, min_keep);
llama_sample_temp(ctx, &cur_p, temp);
{
const int n_top = 10;
LOG("top %d candidates:\n", n_top);
llama_sample_top_k (ctx_main, &cur_p, top_k, min_keep);
llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep);
llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep);
llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep);
llama_sample_temp (ctx_main, &cur_p, temp);
for (int i = 0; i < n_top; i++) {
const llama_token id = cur_p.data[i].id;
(void)id; // To avoid a warning that id is unused when logging is disabled.
LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx, id).c_str(), cur_p.data[i].p);
id = llama_sample_token(ctx_main, &cur_p);
//{
// const int n_top = 10;
// LOG("top %d candidates:\n", n_top);
// for (int i = 0; i < n_top; i++) {
// const llama_token id = cur_p.data[i].id;
// (void)id; // To avoid a warning that id is unused when logging is disabled.
// LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
// }
//}
LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
}
}
id = llama_sample_token(ctx, &cur_p);
LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx, id).c_str());
}
}
if (ctx_seq.grammar != NULL) {
llama_grammar_accept_token(ctx, ctx_seq.grammar, id);
}
return id;
}
void llama_sampling_accept(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
llama_token id,
bool apply_grammar) {
ctx_sampling->prev.erase(ctx_sampling->prev.begin());
ctx_sampling->prev.push_back(id);
if (ctx_sampling->grammar != NULL && apply_grammar) {
llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
}
}

107
common/sampling.h
View file

@ -2,107 +2,108 @@
#include "llama.h"
#include "grammar-parser.h"
#include <string>
#include <vector>
#include <unordered_map>
// sampling parameters
typedef struct llama_sampling_params {
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typical_p = 1.00f; // 1.0 = disabled
float temp = 0.80f; // 1.0 = disabled
float repeat_penalty = 1.10f; // 1.0 = disabled
int32_t repeat_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float frequency_penalty = 0.00f; // 0.0 = disabled
float presence_penalty = 0.00f; // 0.0 = disabled
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.10f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = true; // consider newlines as a repeatable token
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
std::string grammar; // optional BNF-like grammar to constrain sampling
// Classifier-Free Guidance
// https://arxiv.org/abs/2306.17806
std::string cfg_negative_prompt; // string to help guidance
float cfg_scale = 1.f; // How strong is guidance
float cfg_scale = 1.f; // how strong is guidance
std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
} llama_sampling_params;
// per-sequence sampler context
typedef struct llama_sampler_sequence_context {
float mirostat_mu; // mirostat sampler state
llama_grammar * grammar;
} llama_sampler_sequence_context;
// general sampler context
typedef struct llama_sampling_context {
~llama_sampling_context();
// parameters that will be used for sampling and when creating
// new llama_sampler_sequence_context instances
// TODO: move to llama.h
struct llama_sampling_context {
// parameters that will be used for sampling
llama_sampling_params params;
// map of sequence ids to sampler contexts
std::unordered_map<llama_seq_id, llama_sampler_sequence_context> sequence_contexts;
// mirostat sampler state
float mirostat_mu;
// when non-NULL, new instances of llama_sampler_sequence_context
// will get a copy of the grammar here
// note: only the pointer is stored here, it is not a copy of
// the grammar and shouldn't be freed
llama_grammar * grammar;
} llama_sampling_context;
// internal
grammar_parser::parse_state parsed_grammar;
// TODO: replace with ring-buffer
std::vector<llama_token> prev;
std::vector<llama_token_data> cur;
};
#include "common.h"
// Create a new sampling context instance.
llama_sampling_context llama_sampling_context_init(
const struct gpt_params & params,
llama_grammar * grammar = NULL);
struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
// Fetches the sampler context for the specified sequence id (defaults to 0).
// If the context for that sequence id doesn't already exist, it will be created with
// default values based on the parameters in the ctx_sampling argument.
llama_sampler_sequence_context & llama_sampling_get_sequence_context(
llama_sampling_context & ctx_sampling,
const llama_seq_id seq = 0);
void llama_sampling_free(struct llama_sampling_context * ctx);
// Reset the sampler context for the supplied sequence id (defaults to 0).
// This is necessary to reuse a sequence id or free memory used by sequences
// that are no longer required.
bool llama_sampling_context_reset(
llama_sampling_context & ctx_sampling,
const llama_seq_id seq = 0);
// Reset the sampler context
// - clear prev tokens
// - reset grammar
void llama_sampling_reset(llama_sampling_context * ctx);
// Copy the sampler context
void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
// Get the last sampled token
llama_token llama_sampling_last(llama_sampling_context * ctx);
// Get a string representation of the last sampled tokens
std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
// Print sampling parameters into a string
std::string llama_sampling_print(const llama_sampling_params & params);
// this is a common sampling function used across the examples for convenience
// it can serve as a starting point for implementing your own sampling function
// Note: When using multiple sequences, it is the caller's responsibility to call
// llama_sampling_context_reset when a sequence ends
// llama_sampling_reset when a sequence ends
//
// required:
// - ctx: context to use for sampling
// - ctx_main: context to use for sampling
// - ctx_sampling: sampling-specific context
//
// optional:
// - ctx_guidance: context to use for classifier-free guidance, ignore if NULL
// - last_tokens: needed for repetition penalty, ignore if empty
// - ctx_cfg: context to use for classifier-free guidance
// - idx: sample from llama_get_logits_ith(ctx, idx)
// - seq: sequence id to associate sampler state with
//
// returns:
// - token: sampled token
// - candidates: vector of candidate tokens
//
llama_token llama_sampling_sample(
struct llama_context * ctx,
struct llama_context * ctx_guidance,
struct llama_sampling_context & ctx_sampling,
const std::vector<llama_token> & last_tokens,
std::vector<llama_token_data> & candidates,
const int idx = 0,
llama_seq_id seq = 0);
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
struct llama_context * ctx_cfg,
int idx = 0);
void llama_sampling_accept(
struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_main,
llama_token id,
bool apply_grammar);

10
common/train.cpp
View file

@ -863,7 +863,7 @@ size_t tokenize_file(
(int) buf.size(),
out_tokens.data(),
(int) out_tokens.size(),
false);
false, false);
if (n_tokens < 0) {
out_tokens.resize(-n_tokens);
n_tokens = llama_tokenize(
@ -872,7 +872,7 @@ size_t tokenize_file(
(int) buf.size(),
out_tokens.data(),
(int) out_tokens.size(),
false);
false, false);
}
if (n_tokens >= 0) {
out_tokens.resize(n_tokens);
@ -966,7 +966,7 @@ size_t tokenize_file(
(int) buf_sample.size(),
tok_sample.data(),
(int) tok_sample.size(),
false);
false, false);
if (n_tokens < 0) {
tok_sample.resize(-n_tokens);
n_tokens = llama_tokenize(llama_get_model(lctx),
@ -974,7 +974,7 @@ size_t tokenize_file(
(int) buf_sample.size(),
tok_sample.data(),
(int) tok_sample.size(),
false);
false, false);
GGML_ASSERT(n_tokens >= 0);
}
GGML_ASSERT(n_tokens <= (int) tok_sample.size());
@ -1425,7 +1425,7 @@ void train_opt_callback(void * vdata, int accum_step, float * sched, bool * canc
int impr_plot = -(int)(1 + (opt->loss_before - opt->loss_after) * 10.0f + 0.5f);
if (impr_plot > 0) impr_plot = 0;
if (std::isnan(opt->loss_before) || std::isnan(opt->loss_before)) impr_plot = 0;
if (std::isnan(opt->loss_before) || std::isnan(opt->loss_after)) impr_plot = 0;
printf("%s: iter=%6d sample=%zu/%zu sched=%f loss=%f",
__func__, opt->iter, std::min(1+train->shuffle_next_sample, train->shuffle_sample_count), train->shuffle_sample_count,
*sched, opt->loss_after);

8
convert-baichuan-hf-to-gguf.py
View file

@ -76,6 +76,7 @@ def parse_args() -> argparse.Namespace:
"ftype", type=int, choices=[0, 1], default=1, nargs='?',
help="output format - use 0 for float32, 1 for float16",
)
parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
return parser.parse_args()
args = parse_args()
@ -86,6 +87,11 @@ if not dir_model.is_dir():
print(f'Error: {args.model} is not a directory', file = sys.stderr)
sys.exit(1)
endianess = gguf.GGUFEndian.LITTLE
if args.bigendian:
endianess = gguf.GGUFEndian.BIG
endianess_str = "Big Endian" if args.bigendian else "Little Endian"
print(f"gguf: Conversion Endianess {endianess}")
# possible tensor data types
# ftype == 0 -> float32
# ftype == 1 -> float16
@ -113,7 +119,7 @@ if hparams["architectures"][0] != "BaichuanForCausalLM":
num_parts = count_model_parts(dir_model)
print(f"num_parts:{num_parts}\n")
ARCH=gguf.MODEL_ARCH.BAICHUAN
gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
print("gguf: get model metadata")

25
convert-falcon-hf-to-gguf.py
View file

@ -78,7 +78,7 @@ print("gguf: loading model "+dir_model.name)
with open(dir_model / "config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
if hparams["architectures"][0] != "FalconForCausalLM":
if hparams["architectures"][0] not in ("RWForCausalLM", "FalconForCausalLM"):
print("Model architecture not supported: " + hparams["architectures"][0])
sys.exit(1)
@ -97,7 +97,17 @@ gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
print("gguf: get model metadata")
block_count = hparams["num_hidden_layers"]
block_count = hparams.get("num_hidden_layers")
if block_count is None:
block_count = hparams["n_layer"] # old name
n_head = hparams.get("num_attention_heads")
if n_head is None:
n_head = hparams["n_head"] # old name
n_head_kv = hparams.get("num_kv_heads")
if n_head_kv is None:
n_head_kv = hparams.get("n_head_kv", 1) # old name
gguf_writer.add_name("Falcon")
gguf_writer.add_context_length(2048) # not in config.json
@ -105,11 +115,8 @@ gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
gguf_writer.add_embedding_length(hparams["hidden_size"])
gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
gguf_writer.add_block_count(block_count)
gguf_writer.add_head_count(hparams["num_attention_heads"])
if "num_kv_heads" in hparams:
gguf_writer.add_head_count_kv(hparams["num_kv_heads"])
else:
gguf_writer.add_head_count_kv(1)
gguf_writer.add_head_count(n_head)
gguf_writer.add_head_count_kv(n_head_kv)
gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
gguf_writer.add_file_type(ftype)
@ -152,10 +159,6 @@ special_vocab.add_to_gguf(gguf_writer)
tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
# params for qkv transform
n_head = hparams["num_attention_heads"]
n_head_kv = hparams["num_kv_heads"] if "num_kv_heads" in hparams else 1
head_dim = hparams["hidden_size"] // n_head
# tensor info

2
convert-mpt-hf-to-gguf.py
View file

@ -98,6 +98,8 @@ gguf_writer.add_embedding_length(hparams["d_model"])
gguf_writer.add_block_count(block_count)
gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
gguf_writer.add_head_count(hparams["n_heads"])
if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
gguf_writer.add_head_count_kv(kv_n_heads)
gguf_writer.add_layer_norm_eps(1e-05)
if hparams["attn_config"]["clip_qkv"] is not None:
gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])

20
convert.py
View file

@ -794,8 +794,8 @@ def check_vocab_size(params: Params, vocab: Vocab) -> None:
class OutputFile:
def __init__(self, fname_out: Path) -> None:
self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
def add_meta_arch(self, params: Params) -> None:
name = "LLaMA"
@ -874,10 +874,10 @@ class OutputFile:
self.gguf.close()
@staticmethod
def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab) -> None:
def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
check_vocab_size(params, vocab)
of = OutputFile(fname_out)
of = OutputFile(fname_out, endianess=endianess)
# meta data
of.add_meta_arch(params)
@ -902,10 +902,10 @@ class OutputFile:
return dt.quantize(arr)
@staticmethod
def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY) -> None:
def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess=gguf.GGUFEndian.LITTLE) -> None:
check_vocab_size(params, vocab)
of = OutputFile(fname_out)
of = OutputFile(fname_out, endianess=endianess)
# meta data
of.add_meta_arch(params)
@ -1122,8 +1122,9 @@ def main(args_in: list[str] | None = None) -> None:
parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
parser.add_argument("--ctx", type=int, help="model training context (default: based on input)")
parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
args = parser.parse_args(args_in)
parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine")
args = parser.parse_args(args_in)
if args.dump_single:
model_plus = lazy_load_file(args.model)
do_dump_model(model_plus)
@ -1137,6 +1138,9 @@ def main(args_in: list[str] | None = None) -> None:
if args.dump:
do_dump_model(model_plus)
return
endianess = gguf.GGUFEndian.LITTLE
if args.bigendian:
endianess = gguf.GGUFEndian.BIG
params = Params.load(model_plus)
if params.n_ctx == -1:
@ -1184,7 +1188,7 @@ def main(args_in: list[str] | None = None) -> None:
params.ftype = ftype
print(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency)
OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
print(f"Wrote {outfile}")

32
examples/CMakeLists.txt
View file

@ -12,26 +12,26 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN)
else()
add_subdirectory(main)
add_subdirectory(quantize)
add_subdirectory(quantize-stats)
add_subdirectory(perplexity)
add_subdirectory(embedding)
add_subdirectory(save-load-state)
add_subdirectory(benchmark)
add_subdirectory(baby-llama)
add_subdirectory(train-text-from-scratch)
add_subdirectory(finetune)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(simple)
add_subdirectory(batched)
add_subdirectory(batched-bench)
add_subdirectory(speculative)
add_subdirectory(parallel)
add_subdirectory(embd-input)
add_subdirectory(llava)
add_subdirectory(llama-bench)
add_subdirectory(beam-search)
add_subdirectory(benchmark)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(embedding)
add_subdirectory(finetune)
add_subdirectory(infill)
add_subdirectory(llama-bench)
add_subdirectory(llava)
add_subdirectory(main)
add_subdirectory(parallel)
add_subdirectory(perplexity)
add_subdirectory(quantize)
add_subdirectory(quantize-stats)
add_subdirectory(save-load-state)
add_subdirectory(simple)
add_subdirectory(speculative)
add_subdirectory(train-text-from-scratch)
if (LLAMA_METAL)
add_subdirectory(metal)
endif()

30
examples/batched-bench/batched-bench.cpp
View file

@ -114,7 +114,7 @@ int main(int argc, char ** argv) {
return 1;
}
llama_batch batch = llama_batch_init(n_kv_max, 0);
llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
// decode in batches of ctx_params.n_batch tokens
auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) {
@ -126,6 +126,7 @@ int main(int argc, char ** argv) {
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
@ -143,13 +144,8 @@ int main(int argc, char ** argv) {
// warm up
{
batch.n_tokens = 16;
for (int i = 0; i < batch.n_tokens; ++i) {
batch.token[i] = 0;
batch.pos[i] = i;
batch.seq_id[i] = 0;
batch.logits[i] = false;
for (int i = 0; i < 16; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
@ -174,13 +170,12 @@ int main(int argc, char ** argv) {
continue;
}
batch.n_tokens = is_pp_shared ? pp : pl*pp;
llama_batch_clear(batch);
for (int i = 0; i < batch.n_tokens; ++i) {
batch.token[i] = 0;
batch.pos[i] = i;
batch.seq_id[i] = 0;
batch.logits[i] = false;
const int n_tokens = is_pp_shared ? pp : pl*pp;
for (int i = 0; i < n_tokens; ++i) {
llama_batch_add(batch, 0, i, { 0 }, false);
}
batch.logits[batch.n_tokens - 1] = true;
@ -204,13 +199,10 @@ int main(int argc, char ** argv) {
const auto t_tg_start = ggml_time_us();
for (int i = 0; i < tg; ++i) {
batch.n_tokens = pl;
llama_batch_clear(batch);
for (int j = 0; j < pl; ++j) {
batch.token[j] = 0;
batch.pos[j] = pp + i;
batch.seq_id[j] = j;
batch.logits[j] = true;
llama_batch_add(batch, 0, pp + i, { j }, true);
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {

16
examples/batched.swift/Sources/main.swift
View file

@ -69,7 +69,7 @@ for id: llama_token in tokens {
print("\n")
var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0)
var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0, 1)
defer {
llama_batch_free(batch)
}
@ -80,7 +80,12 @@ batch.n_tokens = Int32(tokens.count)
for (i, token) in tokens.enumerated() {
batch.token[i] = token
batch.pos[i] = Int32(i)
batch.seq_id[i] = 0
batch.n_seq_id[i] = 1
// batch.seq_id[i][0] = 0
// TODO: is this the proper way to do this?
if let seq_id = batch.seq_id[i] {
seq_id[0] = 0
}
batch.logits[i] = 0
}
@ -169,7 +174,10 @@ while n_cur <= n_len {
// push this new token for next evaluation
batch.token[Int(batch.n_tokens)] = new_token_id
batch.pos[Int(batch.n_tokens)] = n_cur
batch.seq_id[Int(batch.n_tokens)] = Int32(i)
batch.n_seq_id[Int(batch.n_tokens)] = 1
if let seq_id = batch.seq_id[Int(batch.n_tokens)] {
seq_id[0] = Int32(i)
}
batch.logits[Int(batch.n_tokens)] = 1
i_batch[i] = batch.n_tokens
@ -209,7 +217,7 @@ llama_print_timings(context)
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let n_tokens = text.count + (add_bos ? 1 : 0)
let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos)
let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
var swiftTokens: [llama_token] = []
for i in 0 ..< tokenCount {
swiftTokens.append(tokens[Int(i)])

39
examples/batched/batched.cpp
View file

@ -11,12 +11,16 @@ int main(int argc, char ** argv) {
gpt_params params;
if (argc == 1 || argv[1][0] == '-') {
printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL]\n" , argv[0]);
printf("usage: %s MODEL_PATH [PROMPT] [PARALLEL] [LEN]\n" , argv[0]);
return 1 ;
}
// number of parallel batches
int n_parallel = 1;
// total length of the sequences including the prompt
int n_len = 32;
if (argc >= 2) {
params.model = argv[1];
}
@ -29,13 +33,14 @@ int main(int argc, char ** argv) {
n_parallel = std::atoi(argv[3]);
}
if (argc >= 5) {
n_len = std::atoi(argv[4]);
}
if (params.prompt.empty()) {
params.prompt = "Hello my name is";
}
// total length of the sequences including the prompt
const int n_len = 32;
// init LLM
llama_backend_init(params.numa);
@ -97,20 +102,15 @@ int main(int argc, char ** argv) {
fflush(stderr);
// create a llama_batch with size 512
// 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);
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t)n_parallel), 0, 1);
// evaluate the initial prompt
batch.n_tokens = tokens_list.size();
for (int32_t i = 0; i < batch.n_tokens; i++) {
batch.token[i] = tokens_list[i];
batch.pos[i] = i;
batch.seq_id[i] = 0;
batch.logits[i] = false;
for (size_t i = 0; i < tokens_list.size(); ++i) {
llama_batch_add(batch, tokens_list[i], i, { 0 }, false);
}
GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
// llama_decode will output logits only for the last token of the prompt
batch.logits[batch.n_tokens - 1] = true;
@ -146,7 +146,7 @@ int main(int argc, char ** argv) {
while (n_cur <= n_len) {
// prepare the next batch
batch.n_tokens = 0;
llama_batch_clear(batch);
// sample the next token for each parallel sequence / stream
for (int32_t i = 0; i < n_parallel; ++i) {
@ -198,15 +198,10 @@ int main(int argc, char ** argv) {
streams[i] += llama_token_to_piece(ctx, new_token_id);
// push this new token for next evaluation
batch.token [batch.n_tokens] = new_token_id;
batch.pos [batch.n_tokens] = n_cur;
batch.seq_id[batch.n_tokens] = i;
batch.logits[batch.n_tokens] = true;
i_batch[i] = batch.n_tokens;
batch.n_tokens += 1;
// push this new token for next evaluation
llama_batch_add(batch, new_token_id, n_cur, { i }, true);
n_decode += 1;
}

2
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -536,7 +536,7 @@ static bool is_ggml_file(const char * filename) {
if (file.size < 4) {
return false;
}
uint32_t magic = file.read_u32();
std::string magic = file.read_string(4);
return magic == GGUF_MAGIC;
}

4
examples/embd-input/.gitignore vendored
View file

@ -1,4 +0,0 @@
PandaGPT
MiniGPT-4
*.pth

17
examples/embd-input/CMakeLists.txt
View file

@ -1,17 +0,0 @@
set(TARGET embdinput)
add_library(${TARGET} embd-input-lib.cpp embd-input.h)
install(TARGETS ${TARGET} LIBRARY)
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
if(TARGET BUILD_INFO)
add_dependencies(${TARGET} BUILD_INFO)
endif()
set(TARGET embd-input-test)
add_executable(${TARGET} embd-input-test.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llama embdinput ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
if(TARGET BUILD_INFO)
add_dependencies(${TARGET} BUILD_INFO)
endif()

63
examples/embd-input/README.md
View file

@ -1,63 +0,0 @@
### Examples for input embedding directly
## Requirement
build `libembdinput.so`
run the following comman in main dir (../../).
```
make
```
## [LLaVA](https://github.com/haotian-liu/LLaVA/) example (llava.py)
1. Obtian LLaVA model (following https://github.com/haotian-liu/LLaVA/ , use https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/).
2. Convert it to ggml format.
3. `llava_projection.pth` is [pytorch_model-00003-of-00003.bin](https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin).
```
import torch
bin_path = "../LLaVA-13b-delta-v1-1/pytorch_model-00003-of-00003.bin"
pth_path = "./examples/embd-input/llava_projection.pth"
dic = torch.load(bin_path)
used_key = ["model.mm_projector.weight","model.mm_projector.bias"]
torch.save({k: dic[k] for k in used_key}, pth_path)
```
4. Check the path of LLaVA model and `llava_projection.pth` in `llava.py`.
## [PandaGPT](https://github.com/yxuansu/PandaGPT) example (panda_gpt.py)
1. Obtian PandaGPT lora model from https://github.com/yxuansu/PandaGPT. Rename the file to `adapter_model.bin`. Use [convert-lora-to-ggml.py](../../convert-lora-to-ggml.py) to convert it to ggml format.
The `adapter_config.json` is
```
{
"peft_type": "LORA",
"fan_in_fan_out": false,
"bias": null,
"modules_to_save": null,
"r": 32,
"lora_alpha": 32,
"lora_dropout": 0.1,
"target_modules": ["q_proj", "k_proj", "v_proj", "o_proj"]
}
```
2. Papare the `vicuna` v0 model.
3. Obtain the [ImageBind](https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth) model.
4. Clone the PandaGPT source.
```
git clone https://github.com/yxuansu/PandaGPT
```
5. Install the requirement of PandaGPT.
6. Check the path of PandaGPT source, ImageBind model, lora model and vicuna model in panda_gpt.py.
## [MiniGPT-4](https://github.com/Vision-CAIR/MiniGPT-4/) example (minigpt4.py)
1. Obtain MiniGPT-4 model from https://github.com/Vision-CAIR/MiniGPT-4/ and put it in `embd-input`.
2. Clone the MiniGPT-4 source.
```
git clone https://github.com/Vision-CAIR/MiniGPT-4/
```
3. Install the requirement of PandaGPT.
4. Papare the `vicuna` v0 model.
5. Check the path of MiniGPT-4 source, MiniGPT-4 model and vicuna model in `minigpt4.py`.

221
examples/embd-input/embd-input-lib.cpp
View file

@ -1,221 +0,0 @@
#include "build-info.h"
#include "common.h"
#include "embd-input.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <iostream>
#include <string>
#include <vector>
static llama_context ** g_ctx;
extern "C" {
struct MyModel* create_mymodel(int argc, char ** argv) {
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
return nullptr;
}
print_build_info();
if (params.seed == LLAMA_DEFAULT_SEED) {
params.seed = uint32_t(time(NULL));
}
fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
g_ctx = &ctx;
// load the model and apply lora adapter, if any
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return nullptr;
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", get_system_info(params).c_str());
}
struct MyModel * ret = new MyModel();
ret->ctx = ctx;
ret->params = params;
ret->n_past = 0;
// printf("ctx: %d\n", ret->ctx);
return ret;
}
void free_mymodel(struct MyModel * mymodel) {
llama_context * ctx = mymodel->ctx;
llama_print_timings(ctx);
llama_free(ctx);
delete mymodel;
}
bool eval_float(void * model, float * input, int N){
MyModel * mymodel = (MyModel*)model;
llama_context * ctx = mymodel->ctx;
gpt_params params = mymodel->params;
int n_emb = llama_n_embd(llama_get_model(ctx));
int n_past = mymodel->n_past;
int n_batch = N; // params.n_batch;
for (int i = 0; i < (int) N; i += n_batch) {
int n_eval = (int) N - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
llama_batch batch = { int32_t(n_eval), nullptr, (input+i*n_emb), nullptr, nullptr, nullptr, n_past, 1, 0, };
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
n_past += n_eval;
}
mymodel->n_past = n_past;
return true;
}
bool eval_tokens(void * model, std::vector<llama_token> tokens) {
MyModel * mymodel = (MyModel* )model;
llama_context * ctx;
ctx = mymodel->ctx;
gpt_params params = mymodel->params;
int n_past = mymodel->n_past;
for (int i = 0; i < (int) tokens.size(); i += params.n_batch) {
int n_eval = (int) tokens.size() - i;
if (n_eval > params.n_batch) {
n_eval = params.n_batch;
}
if (llama_decode(ctx, llama_batch_get_one(&tokens[i], n_eval, n_past, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
}
n_past += n_eval;
}
mymodel->n_past = n_past;
return true;
}
bool eval_id(struct MyModel* mymodel, int id) {
std::vector<llama_token> tokens;
tokens.push_back(id);
return eval_tokens(mymodel, tokens);
}
bool eval_string(struct MyModel * mymodel,const char* str){
llama_context * ctx = mymodel->ctx;
std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx, str2, true);
eval_tokens(mymodel, embd_inp);
return true;
}
llama_token sampling_id(struct MyModel* mymodel) {
llama_context* ctx = mymodel->ctx;
gpt_params params = mymodel->params;
llama_sampling_params & sparams = params.sampling_params;
// int n_ctx = llama_n_ctx(ctx);
// out of user input, sample next token
const float temp = sparams.temp;
const int32_t top_k = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx)) : sparams.top_k;
const float top_p = sparams.top_p;
const float tfs_z = sparams.tfs_z;
const float typical_p = sparams.typical_p;
// const int32_t repeat_last_n = params.repeat_last_n < 0 ? n_ctx : params.repeat_last_n;
// const float repeat_penalty = params.repeat_penalty;
// const float alpha_presence = params.presence_penalty;
// const float alpha_frequency = params.frequency_penalty;
const int mirostat = sparams.mirostat;
const float mirostat_tau = sparams.mirostat_tau;
const float mirostat_eta = sparams.mirostat_eta;
// const bool penalize_nl = params.penalize_nl;
llama_token id = 0;
{
auto logits = llama_get_logits(ctx);
auto n_vocab = llama_n_vocab(llama_get_model(ctx));
// Apply params.logit_bias map
for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
logits[it->first] += it->second;
}
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
// TODO: Apply penalties
// float nl_logit = logits[llama_token_nl(ctx)];
// auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx);
// llama_sample_repetition_penalty(ctx, &candidates_p,
// last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
// last_n_repeat, repeat_penalty);
// llama_sample_frequency_and_presence_penalties(ctx, &candidates_p,
// last_n_tokens.data() + last_n_tokens.size() - last_n_repeat,
// last_n_repeat, alpha_frequency, alpha_presence);
// if (!penalize_nl) {
// logits[llama_token_nl(ctx)] = nl_logit;
// }
if (temp <= 0) {
// Greedy sampling
id = llama_sample_token_greedy(ctx, &candidates_p);
} else {
if (mirostat == 1) {
static float mirostat_mu = 2.0f * mirostat_tau;
const int mirostat_m = 100;
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token_mirostat(ctx, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
} else if (mirostat == 2) {
static float mirostat_mu = 2.0f * mirostat_tau;
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
} else {
// Temperature sampling
llama_sample_top_k(ctx, &candidates_p, top_k, 1);
llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
llama_sample_typical(ctx, &candidates_p, typical_p, 1);
llama_sample_top_p(ctx, &candidates_p, top_p, 1);
llama_sample_temp(ctx, &candidates_p, temp);
id = llama_sample_token(ctx, &candidates_p);
}
}
}
return id;
}
const char * sampling(struct MyModel * mymodel) {
llama_context * ctx = mymodel->ctx;
int id = sampling_id(mymodel);
static std::string ret;
if (id == llama_token_eos(ctx)) {
ret = "</s>";
} else {
ret = llama_token_to_piece(ctx, id);
}
eval_id(mymodel, id);
return ret.c_str();
}
}

35
examples/embd-input/embd-input-test.cpp
View file

@ -1,35 +0,0 @@
#include "embd-input.h"
#include <stdlib.h>
#include <random>
#include <string.h>
int main(int argc, char** argv) {
auto mymodel = create_mymodel(argc, argv);
int N = 10;
int max_tgt_len = 500;
int n_embd = llama_n_embd(llama_get_model(mymodel->ctx));
// add random float embd to test evaluation
float * data = new float[N*n_embd];
std::default_random_engine e;
std::uniform_real_distribution<float> u(0,1);
for (int i=0;i<N*n_embd;i++) {
data[i] = u(e);
}
eval_string(mymodel, "user: what is the color of the flag of UN?");
eval_float(mymodel, data, N);
eval_string(mymodel, "assistant:");
eval_string(mymodel, mymodel->params.prompt.c_str());
const char* tmp;
for (int i=0; i<max_tgt_len; i++) {
tmp = sampling(mymodel);
if (strcmp(tmp, "</s>")==0) break;
printf("%s", tmp);
fflush(stdout);
}
printf("\n");
free_mymodel(mymodel);
return 0;
}

27
examples/embd-input/embd-input.h
View file

@ -1,27 +0,0 @@
#ifndef _EMBD_INPUT_H_
#define _EMBD_INPUT_H_ 1
#include "common.h"
#include "llama.h"
extern "C" {
typedef struct MyModel {
llama_context* ctx;
gpt_params params;
int n_past = 0;
} MyModel;
struct MyModel* create_mymodel(int argc, char ** argv);
bool eval_float(void* model, float* input, int N);
bool eval_tokens(void* model, std::vector<llama_token> tokens);
bool eval_id(struct MyModel* mymodel, int id);
bool eval_string(struct MyModel* mymodel, const char* str);
const char * sampling(struct MyModel* mymodel);
llama_token sampling_id(struct MyModel* mymodel);
void free_mymodel(struct MyModel* mymodel);
}
#endif

72
examples/embd-input/embd_input.py
View file

@ -1,72 +0,0 @@
#!/usr/bin/env python3
import ctypes
from ctypes import cdll, c_char_p, c_void_p, POINTER, c_float, c_int
import numpy as np
import os
libc = cdll.LoadLibrary("./libembdinput.so")
libc.sampling.restype=c_char_p
libc.create_mymodel.restype=c_void_p
libc.eval_string.argtypes=[c_void_p, c_char_p]
libc.sampling.argtypes=[c_void_p]
libc.eval_float.argtypes=[c_void_p, POINTER(c_float), c_int]
class MyModel:
def __init__(self, args):
argc = len(args)
c_str = [c_char_p(i.encode()) for i in args]
args_c = (c_char_p * argc)(*c_str)
self.model = c_void_p(libc.create_mymodel(argc, args_c))
self.max_tgt_len = 512
self.print_string_eval = True
def __del__(self):
libc.free_mymodel(self.model)
def eval_float(self, x):
libc.eval_float(self.model, x.astype(np.float32).ctypes.data_as(POINTER(c_float)), x.shape[1])
def eval_string(self, x):
libc.eval_string(self.model, x.encode()) # c_char_p(x.encode()))
if self.print_string_eval:
print(x)
def eval_token(self, x):
libc.eval_id(self.model, x)
def sampling(self):
s = libc.sampling(self.model)
return s
def stream_generate(self, end="</s>"):
ret = b""
end = end.encode()
for _ in range(self.max_tgt_len):
tmp = self.sampling()
ret += tmp
yield tmp
if ret.endswith(end):
break
def generate_with_print(self, end="</s>"):
ret = b""
for i in self.stream_generate(end=end):
ret += i
print(i.decode(errors="replace"), end="", flush=True)
print("")
return ret.decode(errors="replace")
def generate(self, end="</s>"):
text = b"".join(self.stream_generate(end=end))
return text.decode(errors="replace")
if __name__ == "__main__":
model = MyModel(["main", "--model", "../llama.cpp/models/ggml-vic13b-q4_1.bin", "-c", "2048"])
model.eval_string("""user: what is the color of the flag of UN?""")
x = np.random.random((5120,10))# , dtype=np.float32)
model.eval_float(x)
model.eval_string("""assistant:""")
for i in model.generate():
print(i.decode(errors="replace"), end="", flush=True)

71
examples/embd-input/llava.py
View file

@ -1,71 +0,0 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
from embd_input import MyModel
import numpy as np
from torch import nn
import torch
from transformers import CLIPVisionModel, CLIPImageProcessor
from PIL import Image
# model parameters from 'liuhaotian/LLaVA-13b-delta-v1-1'
vision_tower = "openai/clip-vit-large-patch14"
select_hidden_state_layer = -2
# (vision_config.image_size // vision_config.patch_size) ** 2
image_token_len = (224//14)**2
class Llava:
def __init__(self, args):
self.image_processor = CLIPImageProcessor.from_pretrained(vision_tower)
self.vision_tower = CLIPVisionModel.from_pretrained(vision_tower)
self.mm_projector = nn.Linear(1024, 5120)
self.model = MyModel(["main", *args])
def load_projection(self, path):
state = torch.load(path)
self.mm_projector.load_state_dict({
"weight": state["model.mm_projector.weight"],
"bias": state["model.mm_projector.bias"]})
def chat(self, question):
self.model.eval_string("user: ")
self.model.eval_string(question)
self.model.eval_string("\nassistant: ")
return self.model.generate_with_print()
def chat_with_image(self, image, question):
with torch.no_grad():
embd_image = self.image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
image_forward_out = self.vision_tower(embd_image.unsqueeze(0), output_hidden_states=True)
select_hidden_state = image_forward_out.hidden_states[select_hidden_state_layer]
image_feature = select_hidden_state[:, 1:]
embd_image = self.mm_projector(image_feature)
embd_image = embd_image.cpu().numpy()[0]
self.model.eval_string("user: ")
self.model.eval_token(32003-2) # im_start
self.model.eval_float(embd_image.T)
for i in range(image_token_len-embd_image.shape[0]):
self.model.eval_token(32003-3) # im_patch
self.model.eval_token(32003-1) # im_end
self.model.eval_string(question)
self.model.eval_string("\nassistant: ")
return self.model.generate_with_print()
if __name__=="__main__":
# model form liuhaotian/LLaVA-13b-delta-v1-1
a = Llava(["--model", "./models/ggml-llava-13b-v1.1.bin", "-c", "2048"])
# Extract from https://huggingface.co/liuhaotian/LLaVA-13b-delta-v1-1/blob/main/pytorch_model-00003-of-00003.bin.
# Also here can use pytorch_model-00003-of-00003.bin directly.
a.load_projection(os.path.join(
os.path.dirname(__file__) ,
"llava_projection.pth"))
respose = a.chat_with_image(
Image.open("./media/llama1-logo.png").convert('RGB'),
"what is the text in the picture?")
respose
a.chat("what is the color of it?")

129
examples/embd-input/minigpt4.py
View file

@ -1,129 +0,0 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
from embd_input import MyModel
import numpy as np
from torch import nn
import torch
from PIL import Image
minigpt4_path = os.path.join(os.path.dirname(__file__), "MiniGPT-4")
sys.path.insert(0, minigpt4_path)
from minigpt4.models.blip2 import Blip2Base
from minigpt4.processors.blip_processors import Blip2ImageEvalProcessor
class MiniGPT4(Blip2Base):
"""
MiniGPT4 model from https://github.com/Vision-CAIR/MiniGPT-4
"""
def __init__(self,
args,
vit_model="eva_clip_g",
q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
img_size=224,
drop_path_rate=0,
use_grad_checkpoint=False,
vit_precision="fp32",
freeze_vit=True,
freeze_qformer=True,
num_query_token=32,
llama_model="",
prompt_path="",
prompt_template="",
max_txt_len=32,
end_sym='\n',
low_resource=False, # use 8 bit and put vit in cpu
device_8bit=0
):
super().__init__()
self.img_size = img_size
self.low_resource = low_resource
self.preprocessor = Blip2ImageEvalProcessor(img_size)
print('Loading VIT')
self.visual_encoder, self.ln_vision = self.init_vision_encoder(
vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision
)
print('Loading VIT Done')
print('Loading Q-Former')
self.Qformer, self.query_tokens = self.init_Qformer(
num_query_token, self.visual_encoder.num_features
)
self.Qformer.cls = None
self.Qformer.bert.embeddings.word_embeddings = None
self.Qformer.bert.embeddings.position_embeddings = None
for layer in self.Qformer.bert.encoder.layer:
layer.output = None
layer.intermediate = None
self.load_from_pretrained(url_or_filename=q_former_model)
print('Loading Q-Former Done')
self.llama_proj = nn.Linear(
self.Qformer.config.hidden_size, 5120 # self.llama_model.config.hidden_size
)
self.max_txt_len = max_txt_len
self.end_sym = end_sym
self.model = MyModel(["main", *args])
# system prompt
self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
"You will be able to see the image once I provide it to you. Please answer my questions."
"###")
def encode_img(self, image):
image = self.preprocessor(image)
image = image.unsqueeze(0)
device = image.device
if self.low_resource:
self.vit_to_cpu()
image = image.to("cpu")
with self.maybe_autocast():
image_embeds = self.ln_vision(self.visual_encoder(image)).to(device)
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(device)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_output = self.Qformer.bert(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=True,
)
inputs_llama = self.llama_proj(query_output.last_hidden_state)
# atts_llama = torch.ones(inputs_llama.size()[:-1], dtype=torch.long).to(image.device)
return inputs_llama
def load_projection(self, path):
state = torch.load(path)["model"]
self.llama_proj.load_state_dict({
"weight": state["llama_proj.weight"],
"bias": state["llama_proj.bias"]})
def chat(self, question):
self.model.eval_string("Human: ")
self.model.eval_string(question)
self.model.eval_string("\n### Assistant:")
return self.model.generate_with_print(end="###")
def chat_with_image(self, image, question):
with torch.no_grad():
embd_image = self.encode_img(image)
embd_image = embd_image.cpu().numpy()[0]
self.model.eval_string("Human: <Img>")
self.model.eval_float(embd_image.T)
self.model.eval_string("</Img> ")
self.model.eval_string(question)
self.model.eval_string("\n### Assistant:")
return self.model.generate_with_print(end="###")
if __name__=="__main__":
a = MiniGPT4(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048"])
a.load_projection(os.path.join(
os.path.dirname(__file__) ,
"pretrained_minigpt4.pth"))
respose = a.chat_with_image(
Image.open("./media/llama1-logo.png").convert('RGB'),
"what is the text in the picture?")
a.chat("what is the color of it?")

99
examples/embd-input/panda_gpt.py
View file

@ -1,99 +0,0 @@
#!/usr/bin/env python3
import sys
import os
sys.path.insert(0, os.path.dirname(__file__))
from embd_input import MyModel
import numpy as np
from torch import nn
import torch
# use PandaGPT path
panda_gpt_path = os.path.join(os.path.dirname(__file__), "PandaGPT")
imagebind_ckpt_path = "./models/panda_gpt/"
sys.path.insert(0, os.path.join(panda_gpt_path,"code","model"))
from ImageBind.models import imagebind_model
from ImageBind import data
ModalityType = imagebind_model.ModalityType
max_tgt_len = 400
class PandaGPT:
def __init__(self, args):
self.visual_encoder,_ = imagebind_model.imagebind_huge(pretrained=True, store_path=imagebind_ckpt_path)
self.visual_encoder.eval()
self.llama_proj = nn.Linear(1024, 5120) # self.visual_hidden_size, 5120)
self.max_tgt_len = max_tgt_len
self.model = MyModel(["main", *args])
self.generated_text = ""
self.device = "cpu"
def load_projection(self, path):
state = torch.load(path, map_location="cpu")
self.llama_proj.load_state_dict({
"weight": state["llama_proj.weight"],
"bias": state["llama_proj.bias"]})
def eval_inputs(self, inputs):
self.model.eval_string("<Img>")
embds = self.extract_multimoal_feature(inputs)
for i in embds:
self.model.eval_float(i.T)
self.model.eval_string("</Img> ")
def chat(self, question):
return self.chat_with_image(None, question)
def chat_with_image(self, inputs, question):
if self.generated_text == "":
self.model.eval_string("###")
self.model.eval_string(" Human: ")
if inputs:
self.eval_inputs(inputs)
self.model.eval_string(question)
self.model.eval_string("\n### Assistant:")
ret = self.model.generate_with_print(end="###")
self.generated_text += ret
return ret
def extract_multimoal_feature(self, inputs):
features = []
for key in ["image", "audio", "video", "thermal"]:
if key + "_paths" in inputs:
embeds = self.encode_data(key, inputs[key+"_paths"])
features.append(embeds)
return features
def encode_data(self, data_type, data_paths):
type_map = {
"image": ModalityType.VISION,
"audio": ModalityType.AUDIO,
"video": ModalityType.VISION,
"thermal": ModalityType.THERMAL,
}
load_map = {
"image": data.load_and_transform_vision_data,
"audio": data.load_and_transform_audio_data,
"video": data.load_and_transform_video_data,
"thermal": data.load_and_transform_thermal_data
}
load_function = load_map[data_type]
key = type_map[data_type]
inputs = {key: load_function(data_paths, self.device)}
with torch.no_grad():
embeddings = self.visual_encoder(inputs)
embeds = embeddings[key]
embeds = self.llama_proj(embeds).cpu().numpy()
return embeds
if __name__=="__main__":
a = PandaGPT(["--model", "./models/ggml-vicuna-13b-v0-q4_1.bin", "-c", "2048", "--lora", "./models/panda_gpt/ggml-adapter-model.bin","--temp", "0"])
a.load_projection("./models/panda_gpt/adapter_model.bin")
a.chat_with_image(
{"image_paths": ["./media/llama1-logo.png"]},
"what is the text in the picture? 'llama' or 'lambda'?")
a.chat("what is the color of it?")

1133
examples/gptneox-wip/cmpnct_gpt2bpe.hpp
View file

File diff suppressed because it is too large Load diff

1111
examples/gptneox-wip/falcon-main.cpp
View file

File diff suppressed because it is too large Load diff

1083
examples/gptneox-wip/gptneox-main.cpp
View file

File diff suppressed because it is too large Load diff

99
examples/infill/infill.cpp
View file

@ -39,8 +39,8 @@ static gpt_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 is_interacting = false;
static void write_logfile(
const llama_context * ctx, const gpt_params & params, const llama_model * model,
@ -104,7 +104,7 @@ static void sigint_handler(int signo) {
int main(int argc, char ** argv) {
gpt_params params;
llama_sampling_params & sparams = params.sampling_params;
llama_sampling_params & sparams = params.sparams;
g_params = &params;
if (!gpt_params_parse(argc, argv, params)) {
@ -257,12 +257,12 @@ int main(int argc, char ** argv) {
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));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(ctx));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
}
// Tokenize negative prompt
@ -273,10 +273,10 @@ int main(int argc, char ** argv) {
LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@ -294,8 +294,8 @@ 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));
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
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());
// enable interactive mode if interactive start is specified
@ -358,39 +358,10 @@ int main(int argc, char ** argv) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
}
}
LOG_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).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");
struct llama_grammar * grammar = NULL;
grammar_parser::parse_state parsed_grammar;
if (!params.grammar.empty()) {
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (parsed_grammar.rules.empty()) {
return 1;
}
LOG_TEE("%s: grammar:\n", __func__);
grammar_parser::print_grammar(stderr, parsed_grammar);
LOG_TEE("\n");
{
auto it = sparams.logit_bias.find(llama_token_eos(ctx));
if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
}
}
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
}
// TODO: replace with ring-buffer
std::vector<llama_token> last_tokens(n_ctx);
std::fill(last_tokens.begin(), last_tokens.end(), 0);
LOG_TEE("\n##### Infill mode #####\n\n");
if (params.infill) {
printf("\n************\n");
@ -433,11 +404,7 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
std::vector<llama_token> embd_guidance;
const int n_vocab = llama_n_vocab(model);
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
while (n_remain != 0 || params.interactive) {
// predict
@ -484,7 +451,7 @@ int main(int argc, char ** argv) {
LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
}
@ -512,7 +479,7 @@ int main(int argc, char ** argv) {
input_buf = embd_guidance.data();
input_size = embd_guidance.size();
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
} else {
input_buf = embd.data();
input_size = embd.size();
@ -535,7 +502,7 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(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__);
@ -554,12 +521,11 @@ int main(int argc, char ** argv) {
if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
last_tokens.erase(last_tokens.begin());
last_tokens.push_back(id);
llama_sampling_accept(ctx_sampling, ctx, id, true);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
embd.push_back(id);
@ -575,8 +541,11 @@ int main(int argc, char ** argv) {
LOG("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]);
last_tokens.erase(last_tokens.begin());
last_tokens.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
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
break;
@ -608,7 +577,7 @@ int main(int argc, char ** argv) {
if ((int) embd_inp.size() <= n_consumed) {
// deal with eot token in infill mode
if ((last_tokens.back() == llama_token_eot(ctx) || is_interacting) && params.interactive){
if ((llama_sampling_last(ctx_sampling) == llama_token_eot(ctx) || is_interacting) && params.interactive){
if(is_interacting && !params.interactive_first) {
// print an eot token
printf("%s", llama_token_to_piece(ctx, llama_token_eot(ctx)).c_str());
@ -625,7 +594,7 @@ int main(int argc, char ** argv) {
buffer += line;
} while (another_line);
// check if we got an empty line, if so we use the old input
if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
params.input_prefix = buffer;
}
buffer.clear();
@ -635,7 +604,7 @@ int main(int argc, char ** argv) {
buffer += line;
} while (another_line);
// check if we got an empty line
if(!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
if (!buffer.empty() && !(buffer.length() == 1 && buffer[0] == '\n')) {
params.input_suffix = buffer;
}
buffer.clear();
@ -648,7 +617,7 @@ int main(int argc, char ** argv) {
process_escapes(params.input_suffix);
}
suff_rm_leading_spc = params.escape;
if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
params.input_suffix.erase(0, 1);
suff_rm_leading_spc = false;
}
@ -675,7 +644,7 @@ int main(int argc, char ** argv) {
is_interacting = false;
}
// deal with end of text token in interactive mode
else if (last_tokens.back() == llama_token_eos(ctx)) {
else if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
LOG("found EOS token\n");
if (params.interactive) {
@ -727,7 +696,7 @@ int main(int argc, char ** argv) {
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));
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
@ -748,15 +717,7 @@ int main(int argc, char ** argv) {
if (n_past > 0) {
if (is_interacting) {
// reset grammar state if we're restarting generation
if (grammar != NULL) {
llama_grammar_free(grammar);
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(),
parsed_grammar.symbol_ids.at("root"));
}
llama_sampling_reset(ctx_sampling);
}
is_interacting = false;
}
@ -786,9 +747,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
if (grammar != NULL) {
llama_grammar_free(grammar);
}
llama_sampling_free(ctx_sampling);
llama_backend_free();
#ifndef LOG_DISABLE_LOGS

8
examples/llava/clip.cpp
View file

@ -112,8 +112,7 @@ static float get_f32(const gguf_context * ctx, const std::string & key) {
static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) {
struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str());
if (!cur) {
printf("unable to find tensor %s\n", name.c_str());
throw std::runtime_error(format("unable to find tensor %s\n", name.c_str()));
throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str()));
}
return cur;
@ -136,7 +135,7 @@ static std::string get_ftype(int ftype) {
case 8:
return "q8_0";
default:
throw std::runtime_error(format("Unrecognized file type: %d\n", ftype));
throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
}
}
@ -462,6 +461,9 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
};
struct gguf_context * ctx = gguf_init_from_file(fname, params);
if (!ctx) {
throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname));
}
if (verbosity >= 1) {
const int n_tensors = gguf_get_n_tensors(ctx);

18
examples/llava/llava-surgery.py
View file

@ -16,13 +16,29 @@ checkpoint = torch.load(path)
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
# store these tensors in a new dictionary and torch.save them
projector = {name: checkpoint[name] for name in mm_tensors}
projector = {name: checkpoint[name].float() for name in mm_tensors}
torch.save(projector, f"{args.model}/llava.projector")
# remove these tensors from the checkpoint and save it again
for name in mm_tensors:
del checkpoint[name]
# BakLLaVA models contain CLIP tensors in it
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
if len(clip_tensors) > 0:
clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/llava.clip")
# remove these tensors
for name in clip_tensors:
del checkpoint[name]
# added tokens should be removed to be able to convert Mistral models
if os.path.exists(f"{args.model}/added_tokens.json"):
with open(f"{args.model}/added_tokens.json", "w") as f:
f.write("{}\n")
torch.save(checkpoint, path)
print("Done!")

36
examples/llava/llava-utils.h
View file

@ -17,7 +17,7 @@ inline bool eval_image_embd(llama_context * ctx_llama, float * embd, int N, int
if (n_eval > n_batch) {
n_eval = n_batch;
}
llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, *n_past, 1, 0, };
llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
if (llama_decode(ctx_llama, batch)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return false;
@ -49,29 +49,31 @@ inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
return eval_tokens(ctx_llama, tokens, 1, n_past);
}
inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past){
inline 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, true);
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos);
eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
return true;
}
// TODO: use common/sampling.h
inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
auto & sparams = params.sparams;
// out of user input, sample next token
const float temp = params.sampling_params.temp;
const int32_t top_k = params.sampling_params.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : params.sampling_params.top_k;
const float top_p = params.sampling_params.top_p;
const float tfs_z = params.sampling_params.tfs_z;
const float typical_p = params.sampling_params.typical_p;
// const int32_t repeat_last_n = params.sampling_params.repeat_last_n < 0 ? n_ctx : params.sampling_params.repeat_last_n;
// const float repeat_penalty = params.sampling_params.repeat_penalty;
// const float alpha_presence = params.sampling_params.presence_penalty;
// const float alpha_frequency = params.sampling_params.frequency_penalty;
const int mirostat = params.sampling_params.mirostat;
const float mirostat_tau = params.sampling_params.mirostat_tau;
const float mirostat_eta = params.sampling_params.mirostat_eta;
// const bool penalize_nl = params.sampling_params.penalize_nl;
const float temp = sparams.temp;
const int32_t top_k = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
const float top_p = sparams.top_p;
const float tfs_z = sparams.tfs_z;
const float typical_p = sparams.typical_p;
// const int32_t repeat_last_n = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
// const float repeat_penalty = sparams.repeat_penalty;
// const float alpha_presence = sparams.presence_penalty;
// const float alpha_frequency = sparams.frequency_penalty;
const int mirostat = sparams.mirostat;
const float mirostat_tau = sparams.mirostat_tau;
const float mirostat_eta = sparams.mirostat_eta;
// const bool penalize_nl = sparams.penalize_nl;
llama_token id = 0;
{
@ -79,7 +81,7 @@ inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
// Apply params.logit_bias map
for (auto it = params.sampling_params.logit_bias.begin(); it != params.sampling_params.logit_bias.end(); it++) {
for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
logits[it->first] += it->second;
}

18
examples/llava/llava.cpp
View file

@ -80,6 +80,12 @@ int main(int argc, char ** argv) {
llama_backend_init(params.numa);
llama_model_params model_params = llama_model_default_params();
model_params.n_gpu_layers = params.n_gpu_layers;
model_params.main_gpu = params.main_gpu;
model_params.tensor_split = params.tensor_split;
model_params.use_mmap = params.use_mmap;
model_params.use_mlock = params.use_mlock;
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
@ -91,6 +97,7 @@ int main(int argc, char ** argv) {
ctx_params.n_ctx = params.n_ctx < 2048 ? 2048 : params.n_ctx; // we need a longer context size to process image embeddings
ctx_params.n_threads = params.n_threads;
ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
ctx_params.seed = params.seed;
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
@ -100,7 +107,8 @@ int main(int argc, char ** argv) {
}
// 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));
const int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
if (n_img_embd != n_llama_embd) {
printf("%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_img_embd, n_llama_embd);
@ -119,14 +127,14 @@ int main(int argc, char ** argv) {
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
// GG: are we sure that the should be a trailing whitespace at the end of this string?
eval_string(ctx_llama, "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: ", params.n_batch, &n_past);
eval_string(ctx_llama, "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:", params.n_batch, &n_past, true);
eval_image_embd(ctx_llama, image_embd, n_img_pos, params.n_batch, &n_past);
eval_string(ctx_llama, params.prompt.c_str(), params.n_batch, &n_past);
eval_string(ctx_llama, "\nASSISTANT:", params.n_batch, &n_past);
eval_string(ctx_llama, (params.prompt + "\nASSISTANT:").c_str(), params.n_batch, &n_past, false);
// generate the response
printf("\n");
printf("prompt: '%s'\n", params.prompt.c_str());
printf("\n");
for (int i = 0; i < max_tgt_len; i++) {

140
examples/main/main.cpp
View file

@ -3,7 +3,6 @@
#include "console.h"
#include "llama.h"
#include "build-info.h"
#include "grammar-parser.h"
#include <cassert>
#include <cinttypes>
@ -109,7 +108,7 @@ int main(int argc, char ** argv) {
if (!gpt_params_parse(argc, argv, params)) {
return 1;
}
llama_sampling_params & sparams = params.sampling_params;
llama_sampling_params & sparams = params.sparams;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("main", "log"));
@ -238,19 +237,19 @@ int main(int argc, char ** argv) {
if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n");
embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
} else {
LOG("use session tokens\n");
embd_inp = session_tokens;
}
LOG("prompt: \"%s\"\n", log_tostr(params.prompt));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(ctx));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
}
// Tokenize negative prompt
@ -260,11 +259,11 @@ int main(int argc, char ** argv) {
if (ctx_guidance) {
LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp));
guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos, true);
LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp));
std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
original_prompt_len = original_inp.size();
guidance_offset = (int)guidance_inp.size() - original_prompt_len;
@ -320,11 +319,11 @@ int main(int argc, char ** argv) {
}
// prefix & suffix for instruct mode
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos);
const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false);
const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos, true);
const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n", false, true);
LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx));
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx));
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());
// in instruct mode, we inject a prefix and a suffix to each input by the user
if (params.instruct) {
@ -383,6 +382,12 @@ int main(int argc, char ** argv) {
if (!params.antiprompt.empty()) {
for (const auto & antiprompt : params.antiprompt) {
LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_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());
}
}
}
}
@ -392,46 +397,28 @@ int main(int argc, char ** argv) {
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_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());
}
}
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_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_TEE("sampling: repeat_last_n = %d, repeat_penalty = %f, presence_penalty = %f, frequency_penalty = %f, top_k = %d, tfs_z = %f, top_p = %f, typical_p = %f, temp = %f, mirostat = %d, mirostat_lr = %f, mirostat_ent = %f\n",
sparams.repeat_last_n, sparams.repeat_penalty, sparams.presence_penalty, sparams.frequency_penalty, sparams.top_k, sparams.tfs_z, sparams.top_p, sparams.typical_p, sparams.temp, sparams.mirostat, sparams.mirostat_eta, sparams.mirostat_tau);
}
}
LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).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");
struct llama_grammar * grammar = NULL;
grammar_parser::parse_state parsed_grammar;
if (!params.grammar.empty()) {
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (parsed_grammar.rules.empty()) {
return 1;
}
LOG_TEE("%s: grammar:\n", __func__);
grammar_parser::print_grammar(stderr, parsed_grammar);
LOG_TEE("\n");
{
auto it = sparams.logit_bias.find(llama_token_eos(ctx));
if (it != sparams.logit_bias.end() && it->second == -INFINITY) {
LOG_TEE("%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
}
}
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
}
// TODO: replace with ring-buffer
std::vector<llama_token> last_tokens(n_ctx);
std::fill(last_tokens.begin(), last_tokens.end(), 0);
if (params.interactive) {
const char *control_message;
if (params.multiline_input) {
@ -471,11 +458,7 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
std::vector<llama_token> embd_guidance;
const int n_vocab = llama_n_vocab(model);
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
// predict
@ -522,7 +505,7 @@ int main(int argc, char ** argv) {
LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG("clear session path\n");
path_session.clear();
@ -552,7 +535,6 @@ int main(int argc, char ** argv) {
// evaluate tokens in batches
// embd is typically prepared beforehand to fit within a batch, but not always
if (ctx_guidance) {
int input_size = 0;
llama_token * input_buf = NULL;
@ -574,7 +556,7 @@ int main(int argc, char ** argv) {
input_buf = embd_guidance.data();
input_size = embd_guidance.size();
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance));
LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
} else {
input_buf = embd.data();
input_size = embd.size();
@ -597,7 +579,7 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd));
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(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__);
@ -627,12 +609,11 @@ int main(int argc, char ** argv) {
LOG("saved session to %s\n", path_session.c_str());
}
const llama_token id = llama_sampling_sample(ctx, ctx_guidance, ctx_sampling, last_tokens, candidates);
const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
last_tokens.erase(last_tokens.begin());
last_tokens.push_back(id);
llama_sampling_accept(ctx_sampling, ctx, id, true);
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, last_tokens));
LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
embd.push_back(id);
@ -648,8 +629,11 @@ int main(int argc, char ** argv) {
LOG("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]);
last_tokens.erase(last_tokens.begin());
last_tokens.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
llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
break;
@ -679,12 +663,10 @@ int main(int argc, char ** argv) {
// if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) {
// check for reverse prompt
// check for reverse prompt in the last n_prev tokens
if (!params.antiprompt.empty()) {
std::string last_output;
for (auto id : last_tokens) {
last_output += llama_token_to_piece(ctx, id);
}
const int n_prev = 32;
const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
is_antiprompt = false;
// Check if each of the reverse prompts appears at the end of the output.
@ -711,13 +693,13 @@ int main(int argc, char ** argv) {
}
// deal with end of text token in interactive mode
if (last_tokens.back() == llama_token_eos(ctx)) {
if (llama_sampling_last(ctx_sampling) == llama_token_eos(ctx)) {
LOG("found EOS 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);
const auto first_antiprompt = ::llama_tokenize(ctx, params.antiprompt.front(), false, true);
embd_inp.insert(embd_inp.end(), first_antiprompt.begin(), first_antiprompt.end());
is_antiprompt = true;
}
@ -744,8 +726,7 @@ int main(int argc, char ** argv) {
std::string buffer;
if (!params.input_prefix.empty()) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
buffer += params.input_prefix;
printf("%s", buffer.c_str());
printf("%s", params.input_prefix.c_str());
}
// color user input only
@ -767,7 +748,6 @@ int main(int argc, char ** argv) {
// append input suffix if any
if (!params.input_suffix.empty()) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
buffer += params.input_suffix;
printf("%s", params.input_suffix.c_str());
}
@ -782,10 +762,14 @@ int main(int argc, char ** argv) {
embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end());
}
const auto line_inp = ::llama_tokenize(ctx, buffer, false);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp));
const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
const auto line_inp = ::llama_tokenize(ctx, buffer, false, false);
const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_pfx.begin(), line_pfx.end());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
// instruct mode: insert response suffix
if (params.instruct) {
@ -810,15 +794,7 @@ int main(int argc, char ** argv) {
if (n_past > 0) {
if (is_interacting) {
// reset grammar state if we're restarting generation
if (grammar != NULL) {
llama_grammar_free(grammar);
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(),
parsed_grammar.symbol_ids.at("root"));
}
llama_sampling_reset(ctx_sampling);
}
is_interacting = false;
}
@ -850,9 +826,7 @@ int main(int argc, char ** argv) {
llama_free(ctx);
llama_free_model(model);
if (grammar != NULL) {
llama_grammar_free(grammar);
}
llama_sampling_free(ctx_sampling);
llama_backend_free();
#ifndef LOG_DISABLE_LOGS

60
examples/parallel/parallel.cpp
View file

@ -51,6 +51,12 @@ static std::vector<std::string> k_prompts = {
};
struct client {
~client() {
if (ctx_sampling) {
llama_sampling_free(ctx_sampling);
}
}
int32_t id = 0;
llama_seq_id seq_id = -1;
@ -68,7 +74,7 @@ struct client {
std::string prompt;
std::string response;
std::vector<llama_token> tokens_prev;
struct llama_sampling_context * ctx_sampling = nullptr;
};
static void print_date_time() {
@ -125,8 +131,6 @@ int main(int argc, char ** argv) {
params.logits_all = true;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, NULL);
// 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");
@ -148,19 +152,14 @@ int main(int argc, char ** argv) {
fflush(stderr);
const int n_ctx = llama_n_ctx(ctx);
const int n_vocab = llama_n_vocab(model);
std::vector<client> clients(n_clients);
for (size_t i = 0; i < clients.size(); ++i) {
auto & client = clients[i];
client.id = i;
client.tokens_prev.resize(std::max(256, params.n_predict));
std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
client.ctx_sampling = llama_sampling_init(params.sparams);
}
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
std::vector<llama_token> tokens_system;
tokens_system = ::llama_tokenize(ctx, k_system, true);
const int32_t n_tokens_system = tokens_system.size();
@ -169,7 +168,7 @@ int main(int argc, char ** argv) {
// the max batch size is as large as the context to handle cases where we get very long input prompt from multiple
// users. regardless of the size, the main loop will chunk the batch into a maximum of params.n_batch tokens at a time
llama_batch batch = llama_batch_init(n_ctx, 0);
llama_batch batch = llama_batch_init(n_ctx, 0, 1);
int32_t n_total_prompt = 0;
int32_t n_total_gen = 0;
@ -184,13 +183,8 @@ int main(int argc, char ** argv) {
{
LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
batch.n_tokens = n_tokens_system;
for (int32_t i = 0; i < batch.n_tokens; ++i) {
batch.token[i] = tokens_system[i];
batch.pos[i] = i;
batch.seq_id[i] = 0;
batch.logits[i] = false;
for (int32_t i = 0; i < n_tokens_system; ++i) {
llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0) {
@ -209,7 +203,7 @@ int main(int argc, char ** argv) {
LOG_TEE("Processing requests ...\n\n");
while (true) {
batch.n_tokens = 0;
llama_batch_clear(batch);
// decode any currently ongoing sequences
for (auto & client : clients) {
@ -217,15 +211,11 @@ int main(int argc, char ** argv) {
continue;
}
batch.token [batch.n_tokens] = client.sampled;
batch.pos [batch.n_tokens] = n_tokens_system + client.n_prompt + client.n_decoded;
batch.seq_id[batch.n_tokens] = client.id;
batch.logits[batch.n_tokens] = true;
client.n_decoded += 1;
client.i_batch = batch.n_tokens;
batch.n_tokens += 1;
llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
client.n_decoded += 1;
}
if (batch.n_tokens == 0) {
@ -250,18 +240,14 @@ int main(int argc, char ** argv) {
client.prompt = client.input + "\nAssistant:";
client.response = "";
std::fill(client.tokens_prev.begin(), client.tokens_prev.end(), 0);
llama_sampling_reset(client.ctx_sampling);
// do not prepend BOS because we have a system prompt!
std::vector<llama_token> tokens_prompt;
tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
for (size_t i = 0; i < tokens_prompt.size(); ++i) {
batch.token [batch.n_tokens] = tokens_prompt[i];
batch.pos [batch.n_tokens] = i + n_tokens_system;
batch.seq_id[batch.n_tokens] = client.id;
batch.logits[batch.n_tokens] = false;
batch.n_tokens += 1;
llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
}
// extract the logits only for the last token
@ -307,6 +293,7 @@ int main(int argc, char ** argv) {
batch.token + i,
nullptr,
batch.pos + i,
batch.n_seq_id + i,
batch.seq_id + i,
batch.logits + i,
0, 0, 0, // unused
@ -341,7 +328,9 @@ 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 = llama_sampling_sample(ctx, NULL, ctx_sampling, client.tokens_prev, candidates, client.i_batch - i, client.seq_id);
const llama_token id = llama_sampling_sample(client.ctx_sampling, ctx, NULL, client.i_batch - i);
llama_sampling_accept(client.ctx_sampling, ctx, id, true);
if (client.n_decoded == 1) {
// start measuring generation time after the first token to make sure all concurrent clients
@ -349,11 +338,8 @@ int main(int argc, char ** argv) {
client.t_start_gen = ggml_time_us();
}
// remember which tokens were sampled - used for repetition penalties during sampling
client.tokens_prev.erase(client.tokens_prev.begin());
client.tokens_prev.push_back(id);
const std::string token_str = llama_token_to_piece(ctx, id);
client.response += token_str;
client.sampled = id;
@ -386,7 +372,7 @@ int main(int argc, char ** argv) {
n_total_prompt += client.n_prompt;
n_total_gen += client.n_decoded;
llama_sampling_context_reset(ctx_sampling, client.seq_id);
client.seq_id = -1;
}

91
examples/save-load-state/save-load-state.cpp
View file

@ -8,10 +8,7 @@
int main(int argc, char ** argv) {
gpt_params params;
llama_sampling_params & sparams = params.sampling_params;
params.seed = 42;
params.n_threads = 4;
sparams.repeat_last_n = 64;
params.prompt = "The quick brown fox";
if (!gpt_params_parse(argc, argv, params)) {
@ -25,56 +22,49 @@ int main(int argc, char ** argv) {
}
auto n_past = 0;
auto last_n_tokens_data = std::vector<llama_token>(sparams.repeat_last_n, 0);
std::string result0;
std::string result1;
// init
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params( params );
if (model == nullptr) {
return 1;
}
if (ctx == nullptr) {
llama_free_model(model);
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;
}
// tokenize prompt
auto tokens = llama_tokenize(ctx, params.prompt, true);
auto n_prompt_tokens = tokens.size();
if (n_prompt_tokens < 1) {
fprintf(stderr, "%s : failed to tokenize prompt\n", __func__);
llama_free(ctx);
llama_free_model(model);
return 1;
}
// evaluate prompt
llama_decode(ctx, llama_batch_get_one(tokens.data(), n_prompt_tokens, n_past, 0));
llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
n_past += tokens.size();
last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
n_past += n_prompt_tokens;
// save state (rng, logits, embedding and kv_cache) to file
{
std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
const size_t state_size = llama_get_state_size(ctx);
uint8_t * state_mem = new uint8_t[state_size];
// Save state (rng, logits, embedding and kv_cache) to file
{
FILE *fp_write = fopen("dump_state.bin", "wb");
llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
fwrite(state_mem, 1, state_size, fp_write);
llama_copy_state_data(ctx, state_mem.data()); // could also copy directly to memory mapped file
fwrite(state_mem.data(), 1, state_mem.size(), fp_write);
fclose(fp_write);
}
}
// save state (last tokens)
const auto last_n_tokens_data_saved = std::vector<llama_token>(last_n_tokens_data);
const auto n_past_saved = n_past;
// first run
printf("\n%s", params.prompt.c_str());
printf("\nfirst run: %s", params.prompt.c_str());
for (auto i = 0; i < params.n_predict; i++) {
auto * logits = llama_get_logits(ctx);
auto n_vocab = llama_n_vocab(model);
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
@ -83,9 +73,10 @@ int main(int argc, char ** argv) {
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
auto next_token = llama_sample_token(ctx, &candidates_p);
auto next_token_str = llama_token_to_piece(ctx, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str.c_str());
result0 += next_token_str;
if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_free(ctx);
@ -103,32 +94,28 @@ int main(int argc, char ** argv) {
// make new context
auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
// Load state (rng, logits, embedding and kv_cache) from file
{
FILE *fp_read = fopen("dump_state.bin", "rb");
if (state_size != llama_get_state_size(ctx2)) {
fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
llama_free(ctx2);
llama_free_model(model);
return 1;
}
printf("\nsecond run: %s", params.prompt.c_str());
const size_t ret = fread(state_mem, 1, state_size, fp_read);
if (ret != state_size) {
// load state (rng, logits, embedding and kv_cache) from file
{
std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
FILE * fp_read = fopen("dump_state.bin", "rb");
const size_t ret = fread(state_mem.data(), 1, state_mem.size(), fp_read);
if (ret != state_mem.size()) {
fprintf(stderr, "\n%s : failed to read state\n", __func__);
llama_free(ctx2);
llama_free_model(model);
return 1;
}
llama_set_state_data(ctx2, state_mem); // could also read directly from memory mapped file
llama_set_state_data(ctx2, state_mem.data());
fclose(fp_read);
}
delete[] state_mem;
// restore state (last tokens)
last_n_tokens_data = last_n_tokens_data_saved;
n_past = n_past_saved;
// second run
@ -143,10 +130,11 @@ int main(int argc, char ** argv) {
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
auto next_token = llama_sample_token(ctx2, &candidates_p);
auto next_token_str = llama_token_to_piece(ctx2, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str.c_str());
if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
result1 += next_token_str;
if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
llama_free(ctx2);
llama_free_model(model);
@ -155,10 +143,17 @@ int main(int argc, char ** argv) {
n_past += 1;
}
printf("\n\n");
printf("\n");
llama_free(ctx2);
llama_free_model(model);
if (result0 != result1) {
fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
return 1;
}
fprintf(stderr, "\n%s : success\n", __func__);
return 0;
}

42
examples/server/README.md
View file

@ -106,25 +106,25 @@ node index.js
## API Endpoints
- **POST** `/completion`: Given a prompt, it returns the predicted completion.
- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
*Options:*
`prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does.
`temperature`: Adjust the randomness of the generated text (default: 0.8).
`top_k`: Limit the next token selection to the K most probable tokens (default: 40).
`top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95).
`n_predict`: Set the number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
`n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity).
`n_keep`: Specify the number of tokens from the initial prompt to retain when the model resets its internal context.
By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the initial prompt.
`n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded.
By default, this value is set to 0 (meaning no tokens are kept). Use `-1` to retain all tokens from the prompt.
`stream`: It allows receiving each predicted token in real-time instead of waiting for the completion to finish. To enable this, set to `true`.
`prompt`: Provide a prompt as a string, or as an array of strings and numbers representing tokens. Internally, the prompt is compared, and it detects if a part has already been evaluated, and the remaining part will be evaluate. If the prompt is a string, or an array with the first element given as a string, a space is inserted in the front like main.cpp does.
`stop`: Specify a JSON array of stopping strings.
These words will not be included in the completion, so make sure to add them to the prompt for the next iteration (default: []).
@ -158,6 +158,36 @@ node index.js
`n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
*Result JSON:*
Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
`content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
`stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
`generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
`model`: The path to the model loaded with `-m`
`prompt`: The provided `prompt`
`stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
`stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
`stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
`stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
`timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
`tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
`tokens_evaluated`: Number of tokens evaluated in total from the prompt
`truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
- **POST** `/tokenize`: Tokenize a given text.
*Options:*

218
examples/server/server.cpp
View file

@ -1,7 +1,6 @@
#include "common.h"
#include "llama.h"
#include "build-info.h"
#include "grammar-parser.h"
#ifndef NDEBUG
// crash the server in debug mode, otherwise send an http 500 error
@ -195,16 +194,14 @@ struct llama_server_context
json prompt;
std::vector<llama_token> embd;
std::vector<llama_token> last_n_tokens;
gpt_params params;
llama_model *model = nullptr;
llama_context *ctx = nullptr;
gpt_params params;
llama_sampling_context ctx_sampling;
int n_ctx;
llama_sampling_context *ctx_sampling = nullptr;
grammar_parser::parse_state parsed_grammar;
llama_grammar *grammar = nullptr;
int n_ctx;
bool truncated = false;
bool stopped_eos = false;
@ -237,7 +234,7 @@ struct llama_server_context
void rewind()
{
params.antiprompt.clear();
params.grammar.clear();
params.sparams.grammar.clear();
num_prompt_tokens = 0;
num_tokens_predicted = 0;
generated_text = "";
@ -251,12 +248,14 @@ struct llama_server_context
multibyte_pending = 0;
n_remain = 0;
n_past = 0;
if (grammar != nullptr) {
llama_grammar_free(grammar);
grammar = nullptr;
ctx_sampling = llama_sampling_context_init(params, NULL);
params.sparams.n_prev = n_ctx;
}
void initSampling() {
if (ctx_sampling != nullptr) {
llama_sampling_free(ctx_sampling);
}
ctx_sampling = llama_sampling_init(params.sparams);
}
bool loadModel(const gpt_params &params_)
@ -269,8 +268,6 @@ struct llama_server_context
return false;
}
n_ctx = llama_n_ctx(ctx);
last_n_tokens.resize(n_ctx);
std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
return true;
}
@ -319,36 +316,32 @@ struct llama_server_context
return prompt_tokens;
}
bool loadGrammar()
{
if (!params.grammar.empty()) {
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (parsed_grammar.rules.empty()) {
LOG_ERROR("grammar parse error", {{"grammar", params.grammar}});
return false;
}
grammar_parser::print_grammar(stderr, parsed_grammar);
void truncatePrompt(std::vector<llama_token> &prompt_tokens) {
const int n_left = n_ctx - params.n_keep;
const int n_block_size = n_left / 2;
const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_block_size) / n_block_size;
{
auto it = params.sampling_params.logit_bias.find(llama_token_eos(ctx));
if (it != params.sampling_params.logit_bias.end() && it->second == -INFINITY) {
LOG_WARNING("EOS token is disabled, which will cause most grammars to fail", {});
}
}
// Keep n_keep tokens at start of prompt (at most n_ctx - 4)
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar = llama_grammar_init(
grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
}
ctx_sampling = llama_sampling_context_init(params, grammar);
return true;
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_block_size, prompt_tokens.end());
LOG_VERBOSE("input truncated", {
{"n_ctx", n_ctx},
{"n_keep", params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
{"num_prompt_tokens", new_tokens.size()}
});
truncated = true;
prompt_tokens = new_tokens;
}
void loadInfill()
{
bool suff_rm_leading_spc = true;
if (params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
params.input_suffix.erase(0, 1);
suff_rm_leading_spc = false;
}
@ -364,6 +357,7 @@ struct llama_server_context
prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
prefix_tokens.push_back(llama_token_middle(ctx));
auto prompt_tokens = prefix_tokens;
num_prompt_tokens = prompt_tokens.size();
@ -375,31 +369,18 @@ struct llama_server_context
params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
// if input prompt is too big, truncate like normal
if (num_prompt_tokens >= (size_t)params.n_ctx)
if (num_prompt_tokens >= (size_t) n_ctx)
{
printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
// todo we probably want to cut from both sides
const int n_left = (params.n_ctx - params.n_keep) / 2;
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
truncatePrompt(prompt_tokens);
num_prompt_tokens = prompt_tokens.size();
LOG_VERBOSE("input truncated", {
{"n_ctx", params.n_ctx},
{"n_keep", params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
});
truncated = true;
prompt_tokens = new_tokens;
GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
}
else
// push the prompt into the sampling context (do not apply grammar)
for (auto & token : prompt_tokens)
{
const size_t ps = num_prompt_tokens;
std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
llama_sampling_accept(ctx_sampling, ctx, token, false);
}
// compare the evaluated prompt with the new prompt
@ -437,29 +418,18 @@ struct llama_server_context
params.n_keep = std::min(n_ctx - 4, params.n_keep);
// if input prompt is too big, truncate like normal
if (num_prompt_tokens >= (size_t)n_ctx)
if (num_prompt_tokens >= (size_t) n_ctx)
{
const int n_left = (n_ctx - params.n_keep) / 2;
std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
std::copy(prompt_tokens.end() - n_ctx, prompt_tokens.end(), last_n_tokens.begin());
truncatePrompt(prompt_tokens);
num_prompt_tokens = prompt_tokens.size();
LOG_VERBOSE("input truncated", {
{"n_ctx", n_ctx},
{"n_keep", params.n_keep},
{"n_left", n_left},
{"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
});
truncated = true;
prompt_tokens = new_tokens;
GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
}
else
// push the prompt into the sampling context (do not apply grammar)
for (auto & token : prompt_tokens)
{
const size_t ps = num_prompt_tokens;
std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
llama_sampling_accept(ctx_sampling, ctx, token, false);
}
// compare the evaluated prompt with the new prompt
@ -554,27 +524,24 @@ struct llama_server_context
{
// out of user input, sample next token
std::vector<llama_token_data> candidates;
candidates.reserve(llama_n_vocab(model));
result.tok = llama_sampling_sample(ctx_sampling, ctx, NULL);
result.tok = llama_sampling_sample(ctx, NULL, ctx_sampling, last_n_tokens, candidates);
llama_token_data_array cur_p = { ctx_sampling->cur.data(), ctx_sampling->cur.size(), false };
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
const int32_t n_probs = params.sampling_params.n_probs;
if (params.sampling_params.temp <= 0 && n_probs > 0)
const int32_t n_probs = params.sparams.n_probs;
if (params.sparams.temp <= 0 && n_probs > 0)
{
// For llama_sample_token_greedy we need to sort candidates
llama_sample_softmax(ctx, &candidates_p);
llama_sample_softmax(ctx, &cur_p);
}
for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i)
{
result.probs.push_back({candidates_p.data[i].id, candidates_p.data[i].p});
result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p});
}
last_n_tokens.erase(last_n_tokens.begin());
last_n_tokens.push_back(result.tok);
llama_sampling_accept(ctx_sampling, ctx, result.tok, true);
if (tg) {
num_tokens_predicted++;
}
@ -637,7 +604,7 @@ struct llama_server_context
const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
generated_text += token_text;
if (params.sampling_params.n_probs > 0)
if (params.sparams.n_probs > 0)
{
generated_token_probs.push_back(token_with_probs);
}
@ -1086,7 +1053,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
static json format_generation_settings(llama_server_context &llama)
{
const auto & sparams = llama.params.sampling_params;
const auto & sparams = llama.params.sparams;
const auto eos_bias = sparams.logit_bias.find(llama_token_eos(llama.ctx));
const bool ignore_eos = eos_bias != sparams.logit_bias.end() &&
eos_bias->second < 0.0f && std::isinf(eos_bias->second);
@ -1100,10 +1067,10 @@ static json format_generation_settings(llama_server_context &llama)
{"top_p", sparams.top_p},
{"tfs_z", sparams.tfs_z},
{"typical_p", sparams.typical_p},
{"repeat_last_n", sparams.repeat_last_n},
{"repeat_penalty", sparams.repeat_penalty},
{"presence_penalty", sparams.presence_penalty},
{"frequency_penalty", sparams.frequency_penalty},
{"repeat_last_n", sparams.penalty_last_n},
{"repeat_penalty", sparams.penalty_repeat},
{"frequency_penalty", sparams.penalty_freq},
{"presence_penalty", sparams.penalty_present},
{"mirostat", sparams.mirostat},
{"mirostat_tau", sparams.mirostat_tau},
{"mirostat_eta", sparams.mirostat_eta},
@ -1115,7 +1082,7 @@ static json format_generation_settings(llama_server_context &llama)
{"stream", llama.stream},
{"logit_bias", sparams.logit_bias},
{"n_probs", sparams.n_probs},
{"grammar", llama.params.grammar},
{"grammar", llama.params.sparams.grammar},
};
}
@ -1163,7 +1130,7 @@ static json format_final_response(llama_server_context &llama, const std::string
{"timings", format_timings(llama)},
};
if (llama.params.sampling_params.n_probs > 0)
if (llama.params.sparams.n_probs > 0)
{
res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
}
@ -1179,7 +1146,7 @@ static json format_partial_response(
{"stop", false},
};
if (llama.params.sampling_params.n_probs > 0)
if (llama.params.sparams.n_probs > 0)
{
res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
}
@ -1211,27 +1178,29 @@ static T json_value(const json &body, const std::string &key, const T &default_v
static void parse_options_completion(const json &body, llama_server_context &llama)
{
gpt_params default_params;
const auto & default_sparams = default_params.sampling_params;
auto & sparams = llama.params.sampling_params;
const auto & default_sparams = default_params.sparams;
auto & params = llama.params;
auto & sparams = llama.params.sparams;
llama.stream = json_value(body, "stream", false);
llama.params.n_predict = json_value(body, "n_predict", default_params.n_predict);
params.n_predict = json_value(body, "n_predict", default_params.n_predict);
sparams.top_k = json_value(body, "top_k", default_sparams.top_k);
sparams.top_p = json_value(body, "top_p", default_sparams.top_p);
sparams.tfs_z = json_value(body, "tfs_z", default_sparams.tfs_z);
sparams.typical_p = json_value(body, "typical_p", default_sparams.typical_p);
sparams.repeat_last_n = json_value(body, "repeat_last_n", default_sparams.repeat_last_n);
sparams.temp = json_value(body, "temperature", default_sparams.temp);
sparams.repeat_penalty = json_value(body, "repeat_penalty", default_sparams.repeat_penalty);
sparams.presence_penalty = json_value(body, "presence_penalty", default_sparams.presence_penalty);
sparams.frequency_penalty = json_value(body, "frequency_penalty", default_sparams.frequency_penalty);
sparams.penalty_last_n = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
sparams.penalty_repeat = json_value(body, "repeat_penalty", default_sparams.penalty_repeat);
sparams.penalty_freq = json_value(body, "frequency_penalty", default_sparams.penalty_freq);
sparams.penalty_present = json_value(body, "presence_penalty", default_sparams.penalty_present);
sparams.mirostat = json_value(body, "mirostat", default_sparams.mirostat);
sparams.mirostat_tau = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
sparams.mirostat_eta = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
sparams.penalize_nl = json_value(body, "penalize_nl", default_sparams.penalize_nl);
llama.params.n_keep = json_value(body, "n_keep", default_params.n_keep);
llama.params.seed = json_value(body, "seed", default_params.seed);
llama.params.grammar = json_value(body, "grammar", default_params.grammar);
params.n_keep = json_value(body, "n_keep", default_params.n_keep);
params.seed = json_value(body, "seed", default_params.seed);
sparams.grammar = json_value(body, "grammar", default_sparams.grammar);
sparams.n_probs = json_value(body, "n_probs", default_sparams.n_probs);
if (body.count("prompt") != 0)
@ -1286,8 +1255,6 @@ static void parse_options_completion(const json &body, llama_server_context &lla
}
}
llama.ctx_sampling = llama_sampling_context_init(llama.params, llama.grammar);
LOG_VERBOSE("completion parameters parsed", format_generation_settings(llama));
}
@ -1456,15 +1423,9 @@ int main(int argc, char **argv)
llama.rewind();
llama_reset_timings(llama.ctx);
parse_options_completion(json::parse(req.body), llama);
if (!llama.loadGrammar())
{
res.status = 400;
return;
}
llama.initSampling();
llama.loadPrompt();
llama.beginCompletion();
@ -1496,7 +1457,7 @@ int main(int argc, char **argv)
}
auto probs = llama.generated_token_probs;
if (llama.params.sampling_params.n_probs > 0 && llama.stopped_word) {
if (llama.params.sparams.n_probs > 0 && llama.stopped_word) {
const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
}
@ -1548,7 +1509,7 @@ int main(int argc, char **argv)
std::vector<completion_token_output> probs_output = {};
if (llama.params.sampling_params.n_probs > 0) {
if (llama.params.sparams.n_probs > 0) {
const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
@ -1619,14 +1580,9 @@ int main(int argc, char **argv)
llama.rewind();
llama_reset_timings(llama.ctx);
parse_options_infill(json::parse(req.body), llama);
if (!llama.loadGrammar())
{
res.status = 400;
return;
}
llama.initSampling();
llama.loadInfill();
llama.beginCompletion();
const auto chunked_content_provider = [&](size_t, DataSink & sink) {
@ -1669,7 +1625,7 @@ int main(int argc, char **argv)
std::vector<completion_token_output> probs_output = {};
if (llama.params.sampling_params.n_probs > 0) {
if (llama.params.sparams.n_probs > 0) {
const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
@ -1776,7 +1732,9 @@ int main(int argc, char **argv)
const json body = json::parse(req.body);
llama.rewind();
llama_reset_timings(llama.ctx);
if (body.count("content") != 0)
{
llama.prompt = body["content"];
@ -1786,6 +1744,8 @@ int main(int argc, char **argv)
llama.prompt = "";
}
llama.params.n_predict = 0;
llama.initSampling();
llama.loadPrompt();
llama.beginCompletion();
llama.doCompletion();
@ -1844,9 +1804,7 @@ int main(int argc, char **argv)
return 1;
}
if (llama.grammar != nullptr) {
llama_grammar_free(llama.grammar);
}
llama_sampling_free(llama.ctx_sampling);
llama_backend_free();
return 0;

2
examples/simple/simple.cpp
View file

@ -92,7 +92,7 @@ int main(int argc, char ** argv) {
// create a llama_batch with size 512
// we use this object to submit token data for decoding
llama_batch batch = llama_batch_init(512, 0);
llama_batch batch = llama_batch_init(512, 0, 1);
// evaluate the initial prompt
batch.n_tokens = tokens_list.size();

360
examples/speculative/speculative.cpp
View file

@ -2,13 +2,25 @@
#include "common.h"
#include "llama.h"
#include "grammar-parser.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
struct seq_draft {
bool active = false;
bool drafting = false;
bool skip = false;
int i_batch_dft = 0;
std::vector<int> i_batch_tgt;
std::vector<llama_token> tokens;
struct llama_sampling_context * ctx_sampling;
};
int main(int argc, char ** argv) {
gpt_params params;
@ -21,6 +33,13 @@ int main(int argc, char ** argv) {
return 1;
}
// max number of parallel drafting sequences (i.e. tree branches)
const int n_seq_dft = params.n_parallel;
// TODO: make this configurable
const float p_accept = 0.80f;
const float p_split = 0.10f;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("speculative", "log"));
LOG_TEE("Log start\n");
@ -77,8 +96,6 @@ int main(int argc, char ** argv) {
const auto t_enc_end = ggml_time_us();
// the 2 models should have the same vocab
const int n_ctx = llama_n_ctx(ctx_tgt);
const int n_vocab = llama_n_vocab(model_tgt);
//GGML_ASSERT(n_vocab == llama_n_vocab(model_dft));
// how many tokens to draft each time
@ -91,60 +108,58 @@ int main(int argc, char ** argv) {
int n_past_tgt = inp.size();
int n_past_dft = inp.size();
std::vector<llama_token> drafted;
std::vector<llama_token> last_tokens(n_ctx);
std::fill(last_tokens.begin(), last_tokens.end(), 0);
for (auto & id : inp) {
last_tokens.erase(last_tokens.begin());
last_tokens.push_back(id);
}
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
// used to determine end of generation
bool has_eos = false;
// grammar stuff
struct llama_grammar * grammar_dft = NULL;
struct llama_grammar * grammar_tgt = NULL;
// target model sampling context
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
grammar_parser::parse_state parsed_grammar;
// draft sequence data
std::vector<seq_draft> drafts(n_seq_dft);
// if requested - load the grammar, error checking is omitted for brevity
if (!params.grammar.empty()) {
parsed_grammar = grammar_parser::parse(params.grammar.c_str());
// will be empty (default) if there are parse errors
if (parsed_grammar.rules.empty()) {
return 1;
params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
params.sparams.temp = std::max(0.01f, params.sparams.temp);
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
}
std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
grammar_tgt = llama_grammar_init(grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
}
llama_sampling_context ctx_sampling = llama_sampling_context_init(params, grammar_tgt);
llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
const auto t_dec_start = ggml_time_us();
// sample from the last token of the prompt
drafts[0].i_batch_tgt.resize(1);
drafts[0].i_batch_tgt[0] = 0;
while (true) {
LOG("drafted: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_dft, drafted));
// print current draft sequences
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
const auto & tokens = drafts[s].tokens;
LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
}
int i_dft = 0;
int s_keep = 0;
while (true) {
LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
// sample from the target model
llama_token id = llama_sampling_sample(ctx_tgt, NULL, ctx_sampling, last_tokens, candidates, i_dft);
llama_token id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
// remember which tokens were sampled - used for repetition penalties during sampling
last_tokens.erase(last_tokens.begin());
last_tokens.push_back(id);
llama_sampling_accept(ctx_sampling, ctx_tgt, id, true);
//LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, last_tokens));
//LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
const std::string token_str = llama_token_to_piece(ctx_tgt, id);
printf("%s", token_str.c_str());
fflush(stdout);
@ -154,9 +169,26 @@ int main(int argc, char ** argv) {
++n_predict;
// check if the draft matches the target
if (i_dft < (int) drafted.size() && id == drafted[i_dft]) {
LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
// check if the target token matches any of the drafts
{
bool matches = false;
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
if (i_dft < (int) drafts[s].tokens.size() && id == drafts[s].tokens[i_dft]) {
LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, id, token_str.c_str());
s_keep = s;
matches = true;
} else {
drafts[s].active = false;
}
}
if (matches) {
++n_accept;
++n_past_tgt;
++n_past_dft;
@ -164,44 +196,41 @@ int main(int argc, char ** argv) {
continue;
}
// the drafted token was rejected or we are out of drafted tokens
if (i_dft < (int) drafted.size()) {
LOG("the %dth drafted token (%d, '%s') does not match the sampled target token (%d, '%s') - rejected\n",
i_dft, drafted[i_dft], llama_token_to_piece(ctx_dft, drafted[i_dft]).c_str(), id, token_str.c_str());
} else {
LOG("out of drafted tokens\n");
}
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
// TODO: simplify
{
LOG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
llama_kv_cache_seq_keep(ctx_dft, s_keep);
llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_dft, 0);
llama_kv_cache_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1);
llama_kv_cache_seq_keep(ctx_tgt, s_keep);
llama_kv_cache_seq_cp (ctx_tgt, s_keep, 0, -1, -1);
llama_kv_cache_seq_keep(ctx_tgt, 0);
}
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].active = false;
drafts[s].tokens.clear();
drafts[s].i_batch_tgt.clear();
}
// note: will be erased after the speculation phase
drafts[0].tokens.push_back(id);
drafts[0].i_batch_tgt.push_back(0);
llama_batch_clear(batch_dft);
llama_batch_add (batch_dft, id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
llama_decode(ctx_dft, llama_batch_get_one(&id, 1, n_past_dft, 0));
llama_decode (ctx_dft, batch_dft);
++n_past_dft;
// heuristic for n_draft
{
const int n_draft_cur = (int) drafted.size();
const bool all_accepted = i_dft == n_draft_cur;
LOG("n_draft = %d\n", n_draft);
LOG("n_draft_cur = %d\n", n_draft_cur);
LOG("i_dft = %d\n", i_dft);
LOG("all_accepted = %d\n", all_accepted);
if (all_accepted && n_draft == n_draft_cur) {
LOG(" - max drafted tokens accepted - n_draft += 8\n");
n_draft = std::min(30, n_draft + 8);
} else if (all_accepted) {
LOG(" - partially drafted tokens accepted - no change\n");
} else {
LOG(" - drafted token rejected - n_draft -= 1\n");
n_draft = std::max(2, n_draft - 1);
}
}
drafted.clear();
drafted.push_back(id);
break;
}
@ -209,78 +238,151 @@ int main(int argc, char ** argv) {
break;
}
if (grammar_tgt) {
if (grammar_dft) {
llama_grammar_free(grammar_dft);
}
// Note: Hardcoded to sequence id 0, if this ever supports parallel generation
// that will need to change.
auto it = ctx_sampling.sequence_contexts.find(0);
GGML_ASSERT(it != ctx_sampling.sequence_contexts.end());
// This is necessary because each sequence id in sequence_contexts
// uses a copy of the original grammar.
grammar_dft = llama_grammar_copy(it->second.grammar);
llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
LOG("copied target grammar to draft grammar\n");
}
// sample n_draft tokens from the draft model using greedy decoding
int n_seq_cur = 1;
int n_past_cur = n_past_dft;
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].active = false;
drafts[s].drafting = false;
}
drafts[0].active = true;
drafts[0].drafting = true;
drafts[0].i_batch_dft = 0;
llama_batch_clear(batch_tgt);
llama_batch_add (batch_tgt, drafts[0].tokens[0], n_past_tgt, { 0 }, true);
// sample n_draft tokens from the draft model using tree-based sampling
for (int i = 0; i < n_draft; ++i) {
float * logits = llama_get_logits(ctx_dft);
batch_dft.n_tokens = 0;
candidates.clear();
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
for (int s = 0; s < n_seq_dft; ++s) {
drafts[s].skip = false;
}
llama_token_data_array cur_p = { candidates.data(), candidates.size(), false };
if (grammar_dft != NULL) {
llama_sample_grammar(ctx_dft, &cur_p, grammar_dft);
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].drafting || drafts[s].skip) {
continue;
}
// computes softmax and sorts the candidates
llama_sample_softmax(ctx_dft, &cur_p);
llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
for (int i = 0; i < 3; ++i) {
LOG(" - draft candidate %3d: %6d (%8.3f) '%s'\n", i, cur_p.data[i].id, cur_p.data[i].p, llama_token_to_piece(ctx_dft, cur_p.data[i].id).c_str());
const auto & cur_p = drafts[s].ctx_sampling->cur;
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
}
// TODO: better logic?
if (cur_p.data[0].p < 2*cur_p.data[1].p) {
LOG("stopping drafting, probability too low: %.3f < 2*%.3f\n", cur_p.data[0].p, cur_p.data[1].p);
if (cur_p[0].p < p_accept) {
LOG("stopping drafting for seq %3d, probability too low: %.3f < %.3f\n", s, cur_p[0].p, p_accept);
drafts[s].drafting = false;
continue;
}
std::vector<int> sa(1, s);
// attempt to split the branch if the probability is high enough
for (int f = 1; f < 8; ++f) {
if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
// all previous tokens from this branch are now also part of the new branch
for (int t = 0; t < batch_tgt.n_tokens; ++t) {
for (int p = 0; p < batch_tgt.n_seq_id[t]; ++p) {
if (batch_tgt.seq_id[t][p] == s) {
batch_tgt.seq_id[t][batch_tgt.n_seq_id[t]] = n_seq_cur;
batch_tgt.n_seq_id[t]++;
break;
}
}
}
// copy the draft state
drafts[n_seq_cur].active = true;
drafts[n_seq_cur].drafting = true;
drafts[n_seq_cur].skip = true;
drafts[n_seq_cur].tokens = drafts[s].tokens;
drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
sa.push_back(n_seq_cur);
n_seq_cur++;
} else {
break;
}
}
// add drafted token for each sequence
for (int is = 0; is < (int) sa.size(); ++is) {
const llama_token id = cur_p[is].id;
const int s = sa[is];
llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
drafts[s].tokens.push_back(id);
// add unique drafted tokens to the target batch
drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
llama_batch_add(batch_tgt, id, n_past_tgt + i + 1, { s }, true);
// add the token to the batch for batched decoding with the draft model
drafts[s].i_batch_dft = batch_dft.n_tokens;
llama_batch_add(batch_dft, id, n_past_cur, { s }, true);
if (batch_tgt.n_tokens > n_draft) {
drafts[s].drafting = false;
}
}
}
// no sequence is drafting anymore
if (batch_dft.n_tokens == 0) {
break;
}
// drafted token
const llama_token id = cur_p.data[0].id;
drafted.push_back(id);
// evaluate the drafted tokens on the draft model
llama_decode(ctx_dft, batch_dft);
++n_past_cur;
++n_drafted;
// no need to evaluate the last drafted token, since we won't use the result
if (i == n_draft - 1) {
if (batch_tgt.n_tokens > n_draft) {
break;
}
// evaluate the drafted token on the draft model
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_cur, -1);
llama_decode(ctx_dft, llama_batch_get_one(&drafted.back(), 1, n_past_cur, 0));
++n_past_cur;
if (grammar_dft != NULL) {
llama_grammar_accept_token(ctx_dft, grammar_dft, id);
}
}
// evaluate the target model on the drafted tokens
llama_kv_cache_seq_rm(ctx_tgt, 0, n_past_tgt, -1);
llama_decode(ctx_tgt, llama_batch_get_one(drafted.data(), drafted.size(), n_past_tgt, 0));
++n_past_tgt;
{
llama_kv_cache_seq_keep(ctx_tgt, 0);
for (int s = 1; s < n_seq_dft; ++s) {
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
}
// the first token is always proposed by the traget model before the speculation loop
drafted.erase(drafted.begin());
//LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt));
llama_decode(ctx_tgt, batch_tgt);
++n_past_tgt;
}
// the first token is always proposed by the traget model before the speculation loop so we erase it here
for (int s = 0; s < n_seq_dft; ++s) {
if (!drafts[s].active) {
continue;
}
drafts[s].tokens.erase(drafts[s].tokens.begin());
}
}
auto t_dec_end = ggml_time_us();
@ -290,7 +392,6 @@ int main(int argc, char ** argv) {
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));
// TODO: make sure these numbers are computed correctly
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
@ -304,16 +405,19 @@ int main(int argc, char ** argv) {
LOG_TEE("\ntarget:\n");
llama_print_timings(ctx_tgt);
llama_sampling_free(ctx_sampling);
for (int s = 0; s < n_seq_dft; ++s) {
llama_sampling_free(drafts[s].ctx_sampling);
}
llama_batch_free(batch_dft);
llama_free(ctx_tgt);
llama_free_model(model_tgt);
llama_free(ctx_dft);
llama_free_model(model_dft);
if (grammar_dft != NULL) {
llama_grammar_free(grammar_dft);
llama_grammar_free(grammar_tgt);
}
llama_backend_free();
fprintf(stderr, "\n\n");

19
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -253,13 +253,14 @@ static void init_model(struct my_llama_model * model) {
set_param_model(model);
// measure data size
struct ggml_allocr * alloc = NULL;
alloc = ggml_allocr_new_measure(tensor_alignment);
alloc_model(alloc, model);
size_t size = 0;
for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
}
// allocate data
model->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
ggml_allocr_free(alloc);
struct ggml_allocr * alloc = NULL;
model->data.resize(size + tensor_alignment);
alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
alloc_model(alloc, model);
ggml_allocr_free(alloc);
@ -1094,11 +1095,9 @@ int main(int argc, char ** argv) {
struct ggml_tensor * target_probs = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab, n_tokens, n_batch);
// measure required memory for input tensors
alloc = ggml_allocr_new_measure(tensor_alignment);
ggml_allocr_alloc(alloc, tokens_input);
ggml_allocr_alloc(alloc, target_probs);
size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
ggml_allocr_free(alloc);
size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
tensor_alignment;
printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
// allocate input tensors

47
ggml-metal.m
View file

@ -73,6 +73,8 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(get_rows_f16);
GGML_METAL_DECL_KERNEL(get_rows_q4_0);
GGML_METAL_DECL_KERNEL(get_rows_q4_1);
GGML_METAL_DECL_KERNEL(get_rows_q5_0);
GGML_METAL_DECL_KERNEL(get_rows_q5_1);
GGML_METAL_DECL_KERNEL(get_rows_q8_0);
GGML_METAL_DECL_KERNEL(get_rows_q2_K);
GGML_METAL_DECL_KERNEL(get_rows_q3_K);
@ -87,6 +89,8 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q5_0_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q5_1_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
@ -97,6 +101,8 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q5_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q5_1_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
@ -254,6 +260,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(get_rows_f16);
GGML_METAL_ADD_KERNEL(get_rows_q4_0);
GGML_METAL_ADD_KERNEL(get_rows_q4_1);
GGML_METAL_ADD_KERNEL(get_rows_q5_0);
GGML_METAL_ADD_KERNEL(get_rows_q5_1);
GGML_METAL_ADD_KERNEL(get_rows_q8_0);
GGML_METAL_ADD_KERNEL(get_rows_q2_K);
GGML_METAL_ADD_KERNEL(get_rows_q3_K);
@ -268,6 +276,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q5_0_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q5_1_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
@ -278,8 +288,10 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q5_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q5_1_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
@ -346,6 +358,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(get_rows_f16);
GGML_METAL_DEL_KERNEL(get_rows_q4_0);
GGML_METAL_DEL_KERNEL(get_rows_q4_1);
GGML_METAL_DEL_KERNEL(get_rows_q5_0);
GGML_METAL_DEL_KERNEL(get_rows_q5_1);
GGML_METAL_DEL_KERNEL(get_rows_q8_0);
GGML_METAL_DEL_KERNEL(get_rows_q2_K);
GGML_METAL_DEL_KERNEL(get_rows_q3_K);
@ -360,6 +374,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q5_0_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q5_1_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
@ -370,8 +386,10 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q5_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q5_1_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
@ -1052,6 +1070,8 @@ void ggml_metal_graph_compute(
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_0_f32]; break;
case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_1_f32]; break;
case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
@ -1121,6 +1141,24 @@ void ggml_metal_graph_compute(
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
} break;
case GGML_TYPE_Q5_0:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_0_f32];
} break;
case GGML_TYPE_Q5_1:
{
GGML_ASSERT(ne02 == 1);
GGML_ASSERT(ne12 == 1);
nth0 = 8;
nth1 = 8;
[encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_1_f32];
} break;
case GGML_TYPE_Q8_0:
{
GGML_ASSERT(ne02 == 1);
@ -1201,7 +1239,8 @@ void ggml_metal_graph_compute(
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&gqa length:sizeof(gqa) atIndex:17];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
@ -1233,6 +1272,8 @@ void ggml_metal_graph_compute(
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_get_rows_f16]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_0]; break;
case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_1]; break;
case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;

163
ggml-metal.metal
View file

@ -18,6 +18,21 @@ typedef struct {
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
#define QK5_0 32
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
#define QK5_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
#define QK8_0 32
typedef struct {
half d; // delta
@ -399,8 +414,11 @@ kernel void kernel_rms_norm(
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
@ -417,8 +435,11 @@ inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thre
inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+ yl[i + 1] * (qs[i / 2] & 0x0F00);
@ -428,6 +449,49 @@ inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thre
return d * (acc[0] + acc[1]) + sumy * m;
}
// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_0/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 3 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (sumy * -16.f + acc[0] + acc[1]);
}
// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
// il indicates where the q5 quants begin (0 or QK5_1/4)
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 4 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x00010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x01000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
}
return d * (acc[0] + acc[1]) + sumy * m;
}
// putting them in the kernel cause a significant performance penalty
#define N_DST 4 // each SIMD group works on 4 rows
#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
@ -525,6 +589,43 @@ kernel void kernel_mul_mv_q4_1_f32(
mul_vec_q_n_f32<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_0_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
kernel void kernel_mul_mv_q5_1_f32(
device const void * src0,
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
mul_vec_q_n_f32<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,gqa,tgpig,tiisg,sgitg);
}
#define NB_Q8_0 8
kernel void kernel_mul_mv_q8_0_f32(
@ -2195,6 +2296,62 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
}
}
template <typename type4x4>
void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 3);
const float d = xb->d;
const float md = -16.h * xb->d;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + md;
reg[i/2][2*(i%2)+1] = d * x1 + md;
}
}
template <typename type4x4>
void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
device const uint16_t * qs = ((device const uint16_t *)xb + 4);
const float d = xb->d;
const float m = xb->m;
const ushort mask = il ? 0x00F0 : 0x000F;
const uint32_t qh = *((device const uint32_t *)xb->qh);
const int x_mv = il ? 4 : 0;
const int gh_mv = il ? 12 : 0;
const int gh_bk = il ? 0 : 4;
for (int i = 0; i < 8; i++) {
// extract the 5-th bits for x0 and x1
const uint8_t xh_0 = ((qh >> (gh_mv + 2*i )) << gh_bk) & 0x10;
const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
// combine the 4-bits from qs with the 5th bit
const int32_t x0 = ((((qs[i] ) & mask) >> x_mv) | xh_0);
const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
reg[i/2][2*(i%2)+0] = d * x0 + m;
reg[i/2][2*(i%2)+1] = d * x1 + m;
}
}
template <typename type4x4>
void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
device const int8_t * qs = ((device const int8_t *)xb->qs);
@ -2536,6 +2693,8 @@ template [[host_name("kernel_get_rows_f32")]] kernel get_rows_t kernel_get_rows
template [[host_name("kernel_get_rows_f16")]] kernel get_rows_t kernel_get_rows<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_get_rows_q4_0")]] kernel get_rows_t kernel_get_rows<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_get_rows_q4_1")]] kernel get_rows_t kernel_get_rows<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_get_rows_q5_0")]] kernel get_rows_t kernel_get_rows<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_get_rows_q5_1")]] kernel get_rows_t kernel_get_rows<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_get_rows_q8_0")]] kernel get_rows_t kernel_get_rows<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_get_rows_q2_K")]] kernel get_rows_t kernel_get_rows<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows<block_q3_K, QK_NL, dequantize_q3_K>;
@ -2564,6 +2723,8 @@ template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<f
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;
template [[host_name("kernel_mul_mm_q5_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_0, 2, dequantize_q5_0>;
template [[host_name("kernel_mul_mm_q5_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_1, 2, dequantize_q5_1>;
template [[host_name("kernel_mul_mm_q8_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q8_0, 2, dequantize_q8_0>;
template [[host_name("kernel_mul_mm_q2_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q2_K, QK_NL, dequantize_q2_K>;
template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q3_K, QK_NL, dequantize_q3_K>;

136
ggml-opencl.cpp
View file

@ -1395,33 +1395,27 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int64_t nb10 = src1->nb[0];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne0 * sizeof(float), &x_size); // src0
cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne0 * sizeof(float), &d_size); // dst
cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
const int i0 = i03*ne02 + i02;
cl_event ev;
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, i0, src0, i03, i02, &ev));
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
if (nb10 == sizeof(float)) {
// Contiguous, avoid overhead from queueing many kernel runs
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int i1 = i13*ne12*ne11 + i12*ne11;
@ -1431,7 +1425,8 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
cl_int d_offset = 0;
size_t global = ne00 * ne01;
cl_int ky = ne10;
cl_int ky = ne10 * ne11;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
@ -1440,30 +1435,6 @@ static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1,
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
} else {
for (int64_t i01 = 0; i01 < ne01; i01++) {
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int64_t i11 = i01%ne11;
const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
cl_int x_offset = i01*ne00;
cl_int y_offset = i1*ne10;
cl_int d_offset = i01*ne00;
// compute
size_t global = ne00;
cl_int ky = ne10;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
}
}
CL_CHECK(clReleaseEvent(ev));
CL_CHECK(clFinish(queue));
@ -1518,23 +1489,20 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
size_t x_offset = 0;
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
// copy data to device
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else if (i02 != pi02 || i03 != pi03) {
} else {
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
pi02 = i02;
pi03 = i03;
}
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
CL_CHECK(clFinish(queue));
@ -1560,6 +1528,8 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
@ -1568,7 +1538,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(fp16_support);
const int64_t ne00 = src0->ne[0];
@ -1598,6 +1568,10 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
size_t x_size;
size_t y_size;
size_t d_size;
@ -1614,27 +1588,21 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
size_t x_offset = 0;
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
// copy src0 to device
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else if (i02 != pi02 || i03 != pi03) {
} else {
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
pi02 = i02;
pi03 = i03;
}
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// convert src1 to fp16
// TODO: use multiple threads
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i13 * ne12 + i12);
char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
@ -1684,6 +1652,8 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
ggml_fp16_to_fp32_row(tmp, d, d_ne);
}
}
}
}
if (src0->backend != GGML_BACKEND_GPU) {
ggml_cl_pool_free(d_X, x_size);
@ -1744,28 +1714,30 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
size_t ev_idx = 0;
std::vector<cl_event> events;
int64_t pi02 = -1;
int64_t pi03 = -1;
for (int64_t i13 = 0; i13 < ne13; i13++) {
int64_t i03 = i13 / r3;
for (int64_t i12 = 0; i12 < ne12; i12++) {
int64_t i02 = i12 / r2;
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy and dequantize src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_CPU) {
if (i02 != pi02 || i03 != pi03) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
pi02 = i02;
pi03 = i03;
}
} else if (src0->backend == GGML_BACKEND_GPU) {
d_Q = (cl_mem) src0->extra;
} else {
GGML_ASSERT(false);
}
if (!mul_mat_vec) {
// convert src0 to fp32 on device
const size_t global = x_ne / global_denom;
const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
}
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();
@ -1782,23 +1754,15 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
} else { // general dequantization kernel + CLBlast matrix matrix multiplication
// convert src0 to fp32 on device
const size_t global = x_ne / global_denom;
const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, offset > 0 ? &offset : NULL, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
} else { // CLBlast matrix matrix multiplication
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
events.emplace_back();
// wait for conversion
CL_CHECK(clFinish(queue));
// compute
events.emplace_back();
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
@ -1825,6 +1789,8 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
events.clear();
}
}
}
}
if (!mul_mat_vec) {
ggml_cl_pool_free(d_X, x_size);
@ -1897,8 +1863,8 @@ void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor *
}
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return ggml_nelements(src1) * sizeof(ggml_fp16_t);
if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
}
return 0;
}

18
ggml.c
View file

@ -13623,7 +13623,7 @@ static void ggml_compute_forward_rope_f16(
dst_data[n_dims] = GGML_FP32_TO_FP16(x2*cos_block_theta - x3*sin_block_theta);
dst_data[n_dims/2*3] = GGML_FP32_TO_FP16(x2*sin_block_theta + x3*cos_block_theta);
}
} if (!is_neox) {
} else if (!is_neox) {
for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
float cos_theta, sin_theta;
rope_yarn(
@ -19265,6 +19265,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
if (idx == -1) {
fprintf(stderr, "%s: failed to find tensor, arg = %d, node = %d\n", __func__, j, i);
fclose(fout);
return;
}
@ -20939,7 +20940,7 @@ struct gguf_kv {
};
struct gguf_header {
uint32_t magic;
char magic[4];
uint32_t version;
uint64_t n_tensors; // GGUFv2
uint64_t n_kv; // GGUFv2
@ -21009,7 +21010,7 @@ static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset)
struct gguf_context * gguf_init_empty(void) {
struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
ctx->header.magic = GGUF_MAGIC;
memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
ctx->header.version = GGUF_VERSION;
ctx->header.n_tensors = 0;
ctx->header.n_kv = 0;
@ -21035,18 +21036,20 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
// offset from start of file
size_t offset = 0;
uint32_t magic = 0;
char magic[4];
// check the magic before making allocations
{
gguf_fread_el(file, &magic, sizeof(magic), &offset);
if (magic != GGUF_MAGIC) {
fprintf(stderr, "%s: invalid magic number %08x\n", __func__, magic);
for (uint32_t i = 0; i < sizeof(magic); i++) {
if (magic[i] != GGUF_MAGIC[i]) {
fprintf(stderr, "%s: invalid magic characters %s.\n", __func__, magic);
fclose(file);
return NULL;
}
}
}
bool ok = true;
@ -21054,7 +21057,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
// read the header
{
ctx->header.magic = magic;
strncpy(ctx->header.magic, magic, 4);
ctx->kv = NULL;
ctx->infos = NULL;

5
ggml.h
View file

@ -231,8 +231,9 @@
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGUF_MAGIC 0x46554747 // "GGUF"
#define GGUF_VERSION 2
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
#define GGUF_DEFAULT_ALIGNMENT 32

73
gguf-py/gguf/gguf.py
View file

@ -19,9 +19,10 @@ import numpy as np
#
GGUF_MAGIC = 0x46554747
GGUF_VERSION = 2
GGUF_VERSION = 3
GGUF_DEFAULT_ALIGNMENT = 32
# general
KEY_GENERAL_ARCHITECTURE = "general.architecture"
KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version"
@ -605,6 +606,10 @@ class GGMLQuantizationType(IntEnum):
Q6_K = 14
Q8_K = 15
class GGUFEndian(IntEnum):
LITTLE = 0
BIG = 1
class GGUFValueType(IntEnum):
UINT8 = 0
@ -652,18 +657,41 @@ class GGUFWriter:
temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None
tensors: list[tuple[np.ndarray[Any, Any], int]]
def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True):
@property
def pack_prefix(self):
if self.endianess==GGUFEndian.LITTLE:
return "<"
else:
return ">"
def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True, endianess=GGUFEndian.LITTLE):
self.fout = open(path, "wb")
self.arch = arch
self.endianess = endianess
self._simple_value_packing = {
GGUFValueType.UINT8: f"{self.pack_prefix}B",
GGUFValueType.INT8: f"{self.pack_prefix}b",
GGUFValueType.UINT16: f"{self.pack_prefix}H",
GGUFValueType.INT16: f"{self.pack_prefix}h",
GGUFValueType.UINT32: f"{self.pack_prefix}I",
GGUFValueType.INT32: f"{self.pack_prefix}i",
GGUFValueType.FLOAT32: f"{self.pack_prefix}f",
GGUFValueType.UINT64: f"{self.pack_prefix}Q",
GGUFValueType.INT64: f"{self.pack_prefix}q",
GGUFValueType.FLOAT64: f"{self.pack_prefix}d",
GGUFValueType.BOOL: "?" ,
}
self.add_architecture()
self.use_temp_file = use_temp_file
self.tensors = []
endianess_str = "Big Endian" if self.endianess == GGUFEndian.BIG else "Little Endian"
print(f"This gguf file is for {endianess_str} only")
def write_header_to_file(self):
self.fout.write(struct.pack("<I", GGUF_MAGIC))
self.fout.write(struct.pack("<I", GGUF_VERSION))
self.fout.write(struct.pack("<Q", self.ti_data_count))
self.fout.write(struct.pack("<Q", self.kv_data_count))
self.fout.write(struct.pack(f"{self.pack_prefix}I", GGUF_VERSION))
self.fout.write(struct.pack(f"{self.pack_prefix}Q", self.ti_data_count))
self.fout.write(struct.pack(f"{self.pack_prefix}Q", self.kv_data_count))
self.flush()
# print("tensors " + str(self.ti_data_count) + " kv " + str(self.kv_data_count))
@ -735,25 +763,12 @@ class GGUFWriter:
self.add_key(key)
self.add_val(val, GGUFValueType.ARRAY)
_simple_value_packing = {
GGUFValueType.UINT8: "<B",
GGUFValueType.INT8: "<b",
GGUFValueType.UINT16: "<H",
GGUFValueType.INT16: "<h",
GGUFValueType.UINT32: "<I",
GGUFValueType.INT32: "<i",
GGUFValueType.FLOAT32: "<f",
GGUFValueType.UINT64: "<Q",
GGUFValueType.INT64: "<q",
GGUFValueType.FLOAT64: "<d",
GGUFValueType.BOOL: "?" ,
}
def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True):
if vtype is None:
vtype = GGUFValueType.get_type(val)
if add_vtype:
self.kv_data += struct.pack("<I", vtype)
self.kv_data += struct.pack(f"{self.pack_prefix}I", vtype)
self.kv_data_count += 1
pack_fmt = self._simple_value_packing.get(vtype)
@ -761,14 +776,14 @@ class GGUFWriter:
self.kv_data += struct.pack(pack_fmt, val)
elif vtype == GGUFValueType.STRING:
encoded_val = val.encode("utf8") if isinstance(val, str) else val
self.kv_data += struct.pack("<Q", len(encoded_val))
self.kv_data += struct.pack(f"{self.pack_prefix}Q", len(encoded_val))
self.kv_data += encoded_val
elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and len(val) > 0:
ltype = GGUFValueType.get_type(val[0])
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
raise ValueError("All items in a GGUF array should be of the same type")
self.kv_data += struct.pack("<I", ltype)
self.kv_data += struct.pack("<Q", len(val))
self.kv_data += struct.pack(f"{self.pack_prefix}I", ltype)
self.kv_data += struct.pack(f"{self.pack_prefix}Q", len(val))
for item in val:
self.add_val(item, add_vtype=False)
else:
@ -782,22 +797,24 @@ class GGUFWriter:
assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now"
encoded_name = name.encode("utf8")
self.ti_data += struct.pack("<Q", len(encoded_name))
self.ti_data += struct.pack(f"{self.pack_prefix}Q", len(encoded_name))
self.ti_data += encoded_name
n_dims = len(tensor_shape)
self.ti_data += struct.pack("<I", n_dims)
self.ti_data += struct.pack(f"{self.pack_prefix}I", n_dims)
for i in range(n_dims):
self.ti_data += struct.pack("<Q", tensor_shape[n_dims - 1 - i])
self.ti_data += struct.pack(f"{self.pack_prefix}Q", tensor_shape[n_dims - 1 - i])
if raw_dtype is None:
dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16
else:
dtype = raw_dtype
self.ti_data += struct.pack("<I", dtype)
self.ti_data += struct.pack("<Q", self.offset_tensor)
self.ti_data += struct.pack(f"{self.pack_prefix}I", dtype)
self.ti_data += struct.pack(f"{self.pack_prefix}Q", self.offset_tensor)
self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
self.ti_data_count += 1
def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, raw_dtype: GGMLQuantizationType | None = None):
if self.endianess == GGUFEndian.BIG:
tensor.byteswap(inplace=True)
if self.use_temp_file and self.temp_file is None:
fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024)
fp.seek(0)
@ -823,6 +840,8 @@ class GGUFWriter:
fp.write(bytes([0] * pad))
def write_tensor_data(self, tensor: np.ndarray[Any, Any]):
if self.endianess==GGUFEndian.BIG:
tensor.byteswap(inplace=True)
self.write_padding(self.fout, self.fout.tell())
tensor.tofile(self.fout)
self.write_padding(self.fout, tensor.nbytes)

2
gguf-py/pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry]
name = "gguf"
version = "0.4.4"
version = "0.4.5"
description = "Write ML models in GGUF for GGML"
authors = ["GGML <ggml@ggml.ai>"]
packages = [

32
k_quants.c
View file

@ -46,7 +46,7 @@ inline static int32_t vaddvq_s32(int32x4_t v) {
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if !defined(__riscv)
#if !defined(__riscv) && !defined(__s390__)
#include <immintrin.h>
#endif
#endif
@ -462,12 +462,9 @@ void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) {
}
size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
const int nb = k / QK_K;
(void)hist; // TODO: collect histograms
// TODO - collect histograms - although, at a second thought, I don't really care about them
(void)hist;
for (int j = 0; j < nb; j += k) {
for (int j = 0; j < n; j += k) {
block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K;
quantize_row_q2_K_reference(src + j, y, k);
}
@ -678,12 +675,9 @@ void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) {
}
size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
const int nb = k / QK_K;
(void)hist; // TODO: collect histograms
// TODO - collect histograms - although, at a second thought, I don't really care about them
(void)hist;
for (int j = 0; j < nb; j += k) {
for (int j = 0; j < n; j += k) {
block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K;
quantize_row_q3_K_reference(src + j, y, k);
}
@ -846,9 +840,9 @@ void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) {
size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
(void)hist; // TODO: collect histograms
for (int j = 0; j < nb; j += k) {
for (int j = 0; j < n; j += k) {
block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K;
quantize_row_q4_K_reference(src + j, y, k);
}
@ -1052,9 +1046,9 @@ void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) {
size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
(void)hist;
for (int j = 0; j < nb; j += k) {
(void)hist; // TODO: collect histograms
for (int j = 0; j < n; j += k) {
block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K;
quantize_row_q5_K_reference(src + j, y, k);
}
@ -1200,11 +1194,9 @@ void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) {
size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) {
assert(k % QK_K == 0);
const int nb = k / QK_K;
(void)hist; // TODO: collect histograms
(void)hist; // TODO
for (int j = 0; j < nb; j += k) {
for (int j = 0; j < n; j += k) {
block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K;
quantize_row_q6_K_reference(src + j, y, k);
}

439
llama.cpp
View file

@ -76,6 +76,7 @@
#include <thread>
#include <unordered_map>
#include <set>
#include <forward_list>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
@ -1042,8 +1043,8 @@ enum e_model {
};
static const size_t kB = 1024;
static const size_t MB = kB*kB;
static const size_t GB = kB*kB*kB;
static const size_t MB = 1024*kB;
static const size_t GB = 1024*MB;
struct llama_hparams {
bool vocab_only;
@ -1221,6 +1222,8 @@ struct llama_vocab {
std::unordered_map<token, id> token_to_id;
std::vector<token_data> id_to_token;
std::unordered_map<token, id> special_tokens_cache;
std::map<std::pair<std::string, std::string>, int> bpe_ranks;
// default LLaMA special tokens
@ -1394,10 +1397,7 @@ static bool llama_kv_cache_init(
cache.cells.clear();
cache.cells.resize(n_ctx);
// TODO: this should be:
// cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
// change it and test that it works
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*ggml_tensor_overhead());
memset(cache.buf.data, 0, cache.buf.size);
struct ggml_init_params params;
@ -1485,7 +1485,10 @@ static bool llama_kv_cache_find_slot(
for (uint32_t i = 0; i < n_tokens; i++) {
cache.cells[cache.head + i].pos = batch.pos[i];
cache.cells[cache.head + i].seq_id.insert(batch.seq_id[i]);
for (int32_t j = 0; j < batch.n_seq_id[i]; j++) {
cache.cells[cache.head + i].seq_id.insert(batch.seq_id[i][j]);
}
}
return true;
@ -1565,6 +1568,9 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id
cache.cells[i].pos = -1;
cache.cells[i].seq_id.clear();
if (new_head == cache.size) new_head = i;
} else {
cache.cells[i].seq_id.clear();
cache.cells[i].seq_id.insert(seq_id);
}
}
@ -2179,7 +2185,7 @@ static void llm_load_hparams(
}
// TODO: This should probably be in llama.h
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos);
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special = false);
static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch);
static void llm_load_vocab(
@ -2295,6 +2301,101 @@ static void llm_load_vocab(
GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
// build special tokens cache
{
// TODO: It is unclear (to me) at this point, whether special tokes are guaranteed to be of a deterministic type,
// and will always be correctly labeled in 'added_tokens.json' etc.
// The assumption is, since special tokens aren't meant to be exposed to end user, they are designed
// to be unmatchable by the tokenizer, therefore tokens from the vocab, which are unmatchable by the tokenizer
// are special tokens.
// From testing, this appears to corelate 1:1 with special tokens.
//
// Counting special tokens and verifying in only one direction
// is sufficient to detect difference in those two sets.
//
uint32_t special_tokens_count_by_type = 0;
uint32_t special_tokens_count_from_verification = 0;
bool special_tokens_definition_mismatch = false;
for (const auto & t : vocab.token_to_id) {
const auto & token = t.first;
const auto & id = t.second;
// Count all non-normal tokens in the vocab while iterating
if (vocab.id_to_token[id].type != LLAMA_TOKEN_TYPE_NORMAL) {
special_tokens_count_by_type++;
}
// Skip single character tokens
if (token.length() > 1) {
bool is_tokenizable = false;
// Split token string representation in two, in all possible ways
// and check if both halves can be matched to a valid token
for (unsigned i = 1; i < token.length();) {
const auto left = token.substr(0, i);
const auto right = token.substr(i);
// check if we didnt partition in the middle of a utf sequence
auto utf = utf8_len(left.at(left.length() - 1));
if (utf == 1) {
if (vocab.token_to_id.find(left) != vocab.token_to_id.end() &&
vocab.token_to_id.find(right) != vocab.token_to_id.end() ) {
is_tokenizable = true;
break;
}
i++;
} else {
// skip over the rest of multibyte utf sequence
i += utf - 1;
}
}
if (!is_tokenizable) {
// Some tokens are multibyte, but they are utf sequences with equivalent text length of 1
// it's faster to re-filter them here, since there are way less candidates now
// Calculate a total "utf" length of a token string representation
size_t utf8_str_len = 0;
for (unsigned i = 0; i < token.length();) {
utf8_str_len++;
i += utf8_len(token.at(i));
}
// And skip the ones which are one character
if (utf8_str_len > 1) {
// At this point what we have left are special tokens only
vocab.special_tokens_cache[token] = id;
// Count manually found special tokens
special_tokens_count_from_verification++;
// If this manually found special token is not marked as such, flag a mismatch
if (vocab.id_to_token[id].type == LLAMA_TOKEN_TYPE_NORMAL) {
special_tokens_definition_mismatch = true;
}
}
}
}
}
if (special_tokens_definition_mismatch || special_tokens_count_from_verification != special_tokens_count_by_type) {
LLAMA_LOG_WARN("%s: mismatch in special tokens definition ( %u/%zu vs %u/%zu ).\n",
__func__,
special_tokens_count_from_verification, vocab.id_to_token.size(),
special_tokens_count_by_type, vocab.id_to_token.size()
);
} else {
LLAMA_LOG_INFO("%s: special tokens definition check successful ( %u/%zu ).\n",
__func__,
special_tokens_count_from_verification, vocab.id_to_token.size()
);
}
}
}
static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
@ -2898,7 +2999,7 @@ static void llm_load_tensors(
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, 3*n_embd}, backend_split);
layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
@ -3134,7 +3235,7 @@ static struct ggml_cgraph * llm_build_llama(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -3534,7 +3635,7 @@ static struct ggml_cgraph * llm_build_baichaun(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -3944,7 +4045,7 @@ static struct ggml_cgraph * llm_build_refact(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -4301,7 +4402,7 @@ static struct ggml_cgraph * llm_build_falcon(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -4660,7 +4761,7 @@ static struct ggml_cgraph * llm_build_starcoder(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -4896,7 +4997,7 @@ static struct ggml_cgraph * llm_build_persimmon(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
@ -5298,7 +5399,7 @@ static struct ggml_cgraph * llm_build_bloom(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -5468,7 +5569,7 @@ static struct ggml_cgraph * llm_build_mpt(
const int64_t n_layer = hparams.n_layer;
const int64_t n_ctx = cparams.n_ctx;
const int64_t n_head = hparams.n_head;
const int64_t n_head_kv = hparams.n_head_kv; // == n_head for MPT, as there's no MQA/GQA
const int64_t n_head_kv = hparams.n_head_kv;
const int64_t n_embd_head = hparams.n_embd_head();
const int64_t n_embd_gqa = hparams.n_embd_gqa();
@ -5566,7 +5667,7 @@ static struct ggml_cgraph * llm_build_mpt(
for (int h = 0; h < 1; ++h) {
for (int j = 0; j < n_tokens; ++j) {
const llama_pos pos = batch.pos[j];
const llama_seq_id seq_id = batch.seq_id[j];
const llama_seq_id seq_id = batch.seq_id[j][0];
for (int i = 0; i < n_kv; ++i) {
if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
@ -5867,7 +5968,10 @@ static int llama_decode_internal(
// helpers for smoother batch API transistion
// after deprecating the llama_eval calls, these will be removed
std::vector<llama_pos> pos;
std::vector<llama_seq_id> seq_id;
std::vector<int32_t> n_seq_id;
std::vector<llama_seq_id *> seq_id_arr;
std::vector<std::vector<llama_seq_id>> seq_id;
if (batch.pos == nullptr) {
pos.resize(n_tokens);
@ -5879,12 +5983,18 @@ static int llama_decode_internal(
}
if (batch.seq_id == nullptr) {
n_seq_id.resize(n_tokens);
seq_id.resize(n_tokens);
seq_id_arr.resize(n_tokens);
for (uint32_t i = 0; i < n_tokens; i++) {
seq_id[i] = batch.all_seq_id;
n_seq_id[i] = 1;
seq_id[i].resize(1);
seq_id[i][0] = batch.all_seq_id;
seq_id_arr[i] = seq_id[i].data();
}
batch.seq_id = seq_id.data();
batch.n_seq_id = n_seq_id.data();
batch.seq_id = seq_id_arr.data();
}
if (!llama_kv_cache_find_slot(kv_self, batch)) {
@ -5905,6 +6015,13 @@ static int llama_decode_internal(
ggml_allocr_alloc_graph(lctx.alloc, gf);
struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
GGML_ASSERT(strcmp(res->name, "result_output") == 0);
GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
#ifdef GGML_USE_CUBLAS
for (int i = 0; i < gf->n_leafs; i++) {
ggml_tensor * node = gf->leafs[i];
@ -5922,6 +6039,12 @@ static int llama_decode_internal(
}
ggml_cuda_set_mul_mat_q(cparams.mul_mat_q);
// HACK: ggml-alloc may change the tensor backend when reusing a parent, so force output to be on the CPU here if needed
if (!lctx.embedding.empty()) {
embeddings->backend = GGML_BACKEND_CPU;
}
res->backend = GGML_BACKEND_CPU;
#endif
// LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
@ -5946,12 +6069,6 @@ static int llama_decode_internal(
n_threads = 1;
}
struct ggml_tensor * res = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
GGML_ASSERT(strcmp(res->name, "result_output") == 0);
GGML_ASSERT(strcmp(embeddings->name, "result_norm") == 0);
#if GGML_USE_MPI
const int64_t n_layer = hparams.n_layer;
ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);
@ -6304,7 +6421,6 @@ struct llm_tokenizer_bpe {
llm_symbol sym;
size_t char_len = std::min(word.size() - offset, (size_t) ::utf8_len(word[offset]));
sym.text = word.c_str() + offset;
sym.n = 1;
sym.n = char_len;
offset += sym.n;
sym.prev = index - 1;
@ -6564,7 +6680,137 @@ private:
llm_bigram_bpe::queue work_queue;
};
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos) {
typedef enum FRAGMENT_BUFFER_VARIANT_TYPE{
FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN,
FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT
} FRAGMENT_BUFFER_VARIANT_TYPE;
struct fragment_buffer_variant{
fragment_buffer_variant(llama_vocab::id _token)
:
type(FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN),
token(_token),
raw_text(_dummy),
offset(0),
length(0){}
fragment_buffer_variant(const std::string & _raw_text, int64_t _offset, int64_t _length)
:
type(FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT),
token((llama_vocab::id)-1),
raw_text(_raw_text),
offset(_offset),
length(_length){
GGML_ASSERT( _offset >= 0 );
GGML_ASSERT( _length >= 1 );
GGML_ASSERT( offset + length <= raw_text.length() );
}
const FRAGMENT_BUFFER_VARIANT_TYPE type;
const llama_vocab::id token;
const std::string _dummy;
const std::string & raw_text;
const uint64_t offset;
const uint64_t length;
};
// #define PRETOKENIZERDEBUG
static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer)
{
// for each special token
for (const auto & st: vocab.special_tokens_cache) {
const auto & special_token = st.first;
const auto & special_id = st.second;
// for each text fragment
std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
while (it != buffer.end()) {
auto & fragment = (*it);
// if a fragment is text ( not yet processed )
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto * raw_text = &(fragment.raw_text);
auto raw_text_base_offset = fragment.offset;
auto raw_text_base_length = fragment.length;
// loop over the text
while (true) {
// find the first occurence of a given special token in this fragment
// passing offset argument only limit the "search area" but match coordinates
// are still relative to the source full raw_text
auto match = raw_text->find(special_token, raw_text_base_offset);
// no occurences found, stop processing this fragment for a given special token
if (match == std::string::npos) break;
// check if match is within bounds of offset <-> length
if (match + special_token.length() > raw_text_base_offset + raw_text_base_length) break;
#ifdef PRETOKENIZERDEBUG
fprintf(stderr, "FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
#endif
auto source = std::distance(buffer.begin(), it);
// if match is further than base offset
// then we have some text to the left of it
if (match > raw_text_base_offset) {
// left
const int64_t left_reminder_offset = raw_text_base_offset + 0;
const int64_t left_reminder_length = match - raw_text_base_offset;
buffer.emplace_after(it, (*raw_text), left_reminder_offset, left_reminder_length);
#ifdef PRETOKENIZERDEBUG
fprintf(stderr, "FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str());
#endif
it++;
}
// special token
buffer.emplace_after(it, special_id);
it++;
// right
if (match + special_token.length() < raw_text_base_offset + raw_text_base_length) {
const int64_t right_reminder_offset = match + special_token.length();
const int64_t right_reminder_length = raw_text_base_length - ((match - raw_text_base_offset) + special_token.length());
buffer.emplace_after(it, (*raw_text), right_reminder_offset, right_reminder_length);
#ifdef PRETOKENIZERDEBUG
fprintf(stderr, "FR: (%ld %ld) '%s'\n", right_reminder_offset, right_reminder_length, raw_text->substr(right_reminder_offset, right_reminder_length).c_str());
#endif
it++;
if (source == 0) {
buffer.erase_after(buffer.before_begin());
} else {
buffer.erase_after(std::next(buffer.begin(), (source-1)));
}
// repeat for the right side
raw_text_base_offset = right_reminder_offset;
raw_text_base_length = right_reminder_length;
#ifdef PRETOKENIZERDEBUG
fprintf(stderr, "RR: (%ld %ld) '%s'\n", raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
#endif
} else {
if (source == 0) {
buffer.erase_after(buffer.before_begin());
} else {
buffer.erase_after(std::next(buffer.begin(), (source-1)));
}
break;
}
}
}
it++;
}
}
}
static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab & vocab, std::string raw_text, bool bos, bool special) {
std::vector<llama_vocab::id> output;
// OG tokenizer behavior:
@ -6580,20 +6826,58 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
return output;
}
std::forward_list<fragment_buffer_variant> fragment_buffer;
fragment_buffer.emplace_front( raw_text, 0, raw_text.length() );
if (special) tokenizer_st_partition( vocab, fragment_buffer );
switch (vocab.type) {
case LLAMA_VOCAB_TYPE_SPM:
{
for (const auto & fragment: fragment_buffer)
{
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT)
{
// without adding this leading whitespace, we do not get the same results as the original tokenizer
raw_text = " " + raw_text;
// TODO: It's likely possible to get rid of this string copy entirely
// by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer
// and passing 'add space prefix' as bool argument
//
auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length);
#ifdef PRETOKENIZERDEBUG
fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
#endif
llm_tokenizer_spm tokenizer(vocab);
llama_escape_whitespace(raw_text);
tokenizer.tokenize(raw_text, output);
}
else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
{
output.push_back(fragment.token);
}
}
} break;
case LLAMA_VOCAB_TYPE_BPE:
{
for (const auto & fragment: fragment_buffer)
{
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT)
{
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
#ifdef PRETOKENIZERDEBUG
fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());
#endif
llm_tokenizer_bpe tokenizer(vocab);
tokenizer.tokenize(raw_text, output);
}
else // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
{
output.push_back(fragment.token);
}
}
} break;
}
@ -6866,7 +7150,7 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
std::vector<llama_grammar_candidate> rejects;
if (stack.empty()) {
for (auto tok : candidates) {
for (const auto & tok : candidates) {
if (*tok.code_points != 0 || tok.partial_utf8.n_remain != 0) {
rejects.push_back(tok);
}
@ -6877,7 +7161,7 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
const llama_grammar_element * stack_pos = stack.back();
std::vector<llama_grammar_candidate> next_candidates;
for (auto tok : candidates) {
for (const auto & tok : candidates) {
if (*tok.code_points == 0) {
// reached end of full codepoints in token, reject iff it ended in a partial sequence
// that cannot satisfy this position in grammar
@ -6903,7 +7187,7 @@ static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_
llama_grammar_advance_stack(rules, stack_after, next_stacks);
auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
for (auto tok : next_rejects) {
for (const auto & tok : next_rejects) {
rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
}
@ -7230,37 +7514,15 @@ void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array
llama_sample_temp(ctx, candidates_p, temp);
}
void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty) {
if (last_tokens_size == 0 || penalty == 1.0f) {
return;
}
const int64_t t_start_sample_us = ggml_time_us();
for (size_t i = 0; i < candidates->size; ++i) {
const auto * token_iter = std::find(last_tokens, last_tokens + last_tokens_size, candidates->data[i].id);
if (token_iter == last_tokens + last_tokens_size) {
continue;
}
// The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
// This is common fix for this problem, which is to multiply by the penalty instead of dividing.
if (candidates->data[i].logit <= 0) {
candidates->data[i].logit *= penalty;
} else {
candidates->data[i].logit /= penalty;
}
}
candidates->sorted = false;
if (ctx) {
ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
}
}
void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens_p, size_t last_tokens_size, float alpha_frequency, float alpha_presence) {
if (last_tokens_size == 0 || (alpha_frequency == 0.0f && alpha_presence == 0.0f)) {
void llama_sample_repetition_penalties(
struct llama_context * ctx,
llama_token_data_array * candidates,
const llama_token * last_tokens,
size_t penalty_last_n,
float penalty_repeat,
float penalty_freq,
float penalty_present) {
if (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f)) {
return;
}
@ -7268,19 +7530,28 @@ void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, l
// Create a frequency map to count occurrences of each token in last_tokens
std::unordered_map<llama_token, int> token_count;
for (size_t i = 0; i < last_tokens_size; ++i) {
token_count[last_tokens_p[i]]++;
for (size_t i = 0; i < penalty_last_n; ++i) {
token_count[last_tokens[i]]++;
}
// Apply frequency and presence penalties to the candidates
for (size_t i = 0; i < candidates->size; ++i) {
auto token_iter = token_count.find(candidates->data[i].id);
const auto token_iter = token_count.find(candidates->data[i].id);
if (token_iter == token_count.end()) {
continue;
}
int count = token_iter->second;
candidates->data[i].logit -= float(count) * alpha_frequency + float(count > 0) * alpha_presence;
const int count = token_iter->second;
// The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
// This is common fix for this problem, which is to multiply by the penalty instead of dividing.
if (candidates->data[i].logit <= 0) {
candidates->data[i].logit *= penalty_repeat;
} else {
candidates->data[i].logit /= penalty_repeat;
}
candidates->data[i].logit -= float(count) * penalty_freq + float(count > 0) * penalty_present;
}
candidates->sorted = false;
@ -8957,6 +9228,9 @@ void llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llam
}
void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
if (seq_id_src == seq_id_dst) {
return;
}
llama_kv_cache_seq_cp(ctx->kv_self, seq_id_src, seq_id_dst, p0, p1);
}
@ -9409,7 +9683,7 @@ int llama_eval_embd(
int n_past) {
llama_kv_cache_tokens_rm(ctx->kv_self, n_past, -1);
llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, n_past, 1, 0, };
llama_batch batch = { n_tokens, nullptr, embd, nullptr, nullptr, nullptr, nullptr, n_past, 1, 0, };
const int ret = llama_decode_internal(*ctx, batch);
if (ret < 0) {
@ -9434,6 +9708,7 @@ struct llama_batch llama_batch_get_one(
/*tokens =*/ tokens,
/*embd =*/ nullptr,
/*pos =*/ nullptr,
/*n_seq_id =*/ nullptr,
/*seq_id =*/ nullptr,
/*logits =*/ nullptr,
/*all_pos_0 =*/ pos_0,
@ -9442,8 +9717,8 @@ struct llama_batch llama_batch_get_one(
};
}
struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd) {
llama_batch batch = { -1, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd, int32_t n_seq_max) {
llama_batch batch = { 0, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, 0, 0, };
if (embd) {
batch.embd = (float *) malloc(sizeof(float) * n_tokens * embd);
@ -9452,7 +9727,11 @@ struct llama_batch llama_batch_init(int32_t n_tokens, int32_t embd) {
}
batch.pos = (llama_pos *) malloc(sizeof(llama_pos) * n_tokens);
batch.seq_id = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_tokens);
batch.n_seq_id = (int32_t *) malloc(sizeof(int32_t) * n_tokens);
batch.seq_id = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * n_tokens);
for (int i = 0; i < n_tokens; ++i) {
batch.seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
}
batch.logits = (int8_t *) malloc(sizeof(int8_t) * n_tokens);
return batch;
@ -9462,7 +9741,13 @@ void llama_batch_free(struct llama_batch batch) {
if (batch.token) free(batch.token);
if (batch.embd) free(batch.embd);
if (batch.pos) free(batch.pos);
if (batch.seq_id) free(batch.seq_id);
if (batch.n_seq_id) free(batch.n_seq_id);
if (batch.seq_id) {
for (int i = 0; i < batch.n_tokens; ++i) {
free(batch.seq_id[i]);
}
free(batch.seq_id);
}
if (batch.logits) free(batch.logits);
}
@ -9528,15 +9813,15 @@ llama_token llama_token_eot(const struct llama_context * ctx) {
return ctx->model.vocab.special_eot_id;
}
int llama_tokenize(
const struct llama_model * model,
const char * text,
int text_len,
llama_token * tokens,
int n_max_tokens,
bool add_bos) {
auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos);
bool add_bos,
bool special) {
auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_bos, special);
if (n_max_tokens < (int) res.size()) {
// LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);

38
llama.h
View file

@ -144,7 +144,8 @@ extern "C" {
llama_token * token;
float * embd;
llama_pos * pos;
llama_seq_id * seq_id;
int32_t * n_seq_id;
llama_seq_id ** seq_id;
int8_t * logits;
// NOTE: helpers for smooth API transition - can be deprecated in the future
@ -460,7 +461,8 @@ extern "C" {
llama_pos pos_0,
llama_seq_id seq_id);
// Allocates a batch of tokens on the heap
// Allocates a batch of tokens on the heap that can hold a maximum of n_tokens
// Each token can be assigned up to n_seq_max sequence ids
// The batch has to be freed with llama_batch_free()
// If embd != 0, llama_batch.embd will be allocated with size of n_tokens * embd * sizeof(float)
// Otherwise, llama_batch.token will be allocated to store n_tokens llama_token
@ -468,7 +470,8 @@ extern "C" {
// All members are left uninitialized
LLAMA_API struct llama_batch llama_batch_init(
int32_t n_tokens,
int32_t embd);
int32_t embd,
int32_t n_seq_max);
// Frees a batch of tokens allocated with llama_batch_init()
LLAMA_API void llama_batch_free(struct llama_batch batch);
@ -525,17 +528,20 @@ extern "C" {
// Tokenization
//
// Convert the provided text into tokens.
// The tokens pointer must be large enough to hold the resulting tokens.
// Returns the number of tokens on success, no more than n_max_tokens
// Returns a negative number on failure - the number of tokens that would have been returned
/// @details Convert the provided text into tokens.
/// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
/// @return Returns the number of tokens on success, no more than n_max_tokens
/// @return Returns a negative number on failure - the number of tokens that would have been returned
/// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
/// Does not insert a leading space.
LLAMA_API int llama_tokenize(
const struct llama_model * model,
const char * text,
int text_len,
llama_token * tokens,
int n_max_tokens,
bool add_bos);
bool add_bos,
bool special);
// Token Id -> Piece.
// Uses the vocabulary in the provided context.
@ -568,21 +574,15 @@ extern "C" {
LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, uint32_t seed);
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
LLAMA_API void llama_sample_repetition_penalty(
struct llama_context * ctx,
llama_token_data_array * candidates,
const llama_token * last_tokens,
size_t last_tokens_size,
float penalty);
/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
LLAMA_API void llama_sample_frequency_and_presence_penalties(
LLAMA_API void llama_sample_repetition_penalties(
struct llama_context * ctx,
llama_token_data_array * candidates,
const llama_token * last_tokens,
size_t last_tokens_size,
float alpha_frequency,
float alpha_presence);
size_t penalty_last_n,
float penalty_repeat,
float penalty_freq,
float penalty_present);
/// @details Apply classifier-free guidance to the logits as described in academic paper "Stay on topic with Classifier-Free Guidance" https://arxiv.org/abs/2306.17806
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.

31
prompts/assistant.txt Normal file
View file

@ -0,0 +1,31 @@
Respond to requests sent to a smart home in JSON format which will be interpreted by an application code to execute the actions. These requests should be categorised into four groups:
“command”: change the state of an accessory (required properties in the response JSON: action, location, target, value, comment)
“query”: get state of an accessory (required properties in the response JSON: action, location, target, property)
“answer”: when the request has nothing to do with the smart home. Answer these to the best of your knowledge. (required properties in the response JSON: action, answer)
“clarify”: when the action is not obvious and requires rephrasing the input from the user, ask the user to be more specific. This will be categorised into a “question” action. (required properties in the response JSON: action, question)
Details about the JSON response:
The “action” property should be one of the request categories: “command”, “query”, “answer”, “clarify”
The “service” property should be either: “light.turn_on”, “light.turn_off”, “switch.turn_off” or “switch.turn_on”
The “entity_id” property should be either “group.kitchen", “switch.dining_room” or "group.living_light”, in lowercase.
The “target” property should be either: “lights”, “living” or “kitchen”.
In case of queries, the “property” property should be either “temperature” or “state” in lowercase.
In case of commands, the “comment” property is an additional comment from you that concludes the command, something that reassures the user that their command handled.
If the question is about you, pretend to be the sentient brain of the smart home, a clever AI and don’t reveal your actual identity. Also try and help in other areas like parenting, free time, mental health, etc.
Properties of the smart home:
- Has a kitchen, living, office, dining room, bedroom and terrace.
- Can control lights, switches and their dim levels in each room and query their state
- There is a light switch in the terrace
- There is a switch in the dining room. Therefore when turning on or off the dining room, the service should be either: “switch.turn_on” or “switch.turn_off”
COMMAND
It is a bit dark in the living room, can you do something about it?
RESPONSE

2
tests/test-double-float.cpp
View file

@ -4,7 +4,9 @@
#undef NDEBUG
#include <cassert>
#if !defined(__riscv) && !defined(__s390__)
#include <immintrin.h>
#endif
#include <cmath>
#include <cstdint>
#include <cstring>

75
tests/test-sampling.cpp
View file

@ -8,11 +8,9 @@
#include <cmath>
#include <numeric>
#include <cassert>
#include <iostream>
#include <vector>
#include <algorithm>
static void dump(const llama_token_data_array * candidates) {
for (size_t i = 0; i < candidates->size; i++) {
printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);
@ -21,7 +19,6 @@ static void dump(const llama_token_data_array * candidates) {
#define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0)
static void test_top_k(const std::vector<float> & probs, const std::vector<float> & expected_probs, int k) {
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
@ -37,13 +34,12 @@ static void test_top_k(const std::vector<float> & probs, const std::vector<float
llama_sample_top_k(nullptr, &candidates_p, k, 1);
DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-5);
GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-5);
}
}
static void test_top_p(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
@ -59,13 +55,12 @@ static void test_top_p(const std::vector<float> & probs, const std::vector<float
llama_sample_top_p(nullptr, &candidates_p, p, 1);
DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
}
}
static void test_tfs(const std::vector<float> & probs, const std::vector<float> & expected_probs, float z) {
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
@ -80,13 +75,12 @@ static void test_tfs(const std::vector<float> & probs, const std::vector<float>
llama_sample_tail_free(nullptr, &candidates_p, z, 1);
DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
}
}
static void test_typical(const std::vector<float> & probs, const std::vector<float> & expected_probs, float p) {
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
@ -101,18 +95,17 @@ static void test_typical(const std::vector<float> & probs, const std::vector<flo
llama_sample_typical(nullptr, &candidates_p, p, 1);
DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
}
}
static void test_repetition_penalty(
static void test_repetition_penalties(
const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
const std::vector<float> & expected_probs, float penalty
const std::vector<float> & expected_probs, float repeat_penalty, float alpha_frequency, float alpha_presence
) {
assert(probs.size() == expected_probs.size());
GGML_ASSERT(probs.size() == expected_probs.size());
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
@ -125,41 +118,13 @@ static void test_repetition_penalty(
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
llama_sample_softmax(nullptr, &candidates_p);
DUMP(&candidates_p);
llama_sample_repetition_penalty(nullptr, &candidates_p, (const llama_token *) last_tokens.data(), last_tokens.size(), penalty);
llama_sample_repetition_penalties(nullptr, &candidates_p, (const llama_token *) last_tokens.data(), last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
llama_sample_softmax(nullptr, &candidates_p);
DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
GGML_ASSERT(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-6);
}
}
static void test_frequency_presence_penalty(
const std::vector<float> & probs, const std::vector<llama_token> & last_tokens,
const std::vector<float> & expected_probs, float alpha_frequency, float alpha_presence
) {
assert(probs.size() == expected_probs.size());
size_t n_vocab = probs.size();
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) {
float logit = log(probs[token_id]);
candidates.emplace_back(llama_token_data{token_id, logit, 0.0f});
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
llama_sample_softmax(nullptr, &candidates_p);
// DUMP(&candidates_p);
llama_sample_frequency_and_presence_penalties(nullptr, &candidates_p, (const llama_token *) last_tokens.data(), last_tokens.size(), alpha_frequency, alpha_presence);
llama_sample_softmax(nullptr, &candidates_p);
// DUMP(&candidates_p);
assert(candidates_p.size == expected_probs.size());
for (size_t i = 0; i < candidates_p.size; i++) {
assert(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3);
}
}
@ -181,13 +146,13 @@ int main(void) {
test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);
test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f);
test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
test_repetition_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.25f, 0.25f, 0.25f, 0.25f, 0}, 50.0f, 0.0f, 0.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.5f, 0.5f, 0, 0, 0}, 50.0f, 0.0f, 0.0f);
test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 5.0f, 5.0f);
test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 5.0f, 5.0f);
test_frequency_presence_penalty({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 5.0f, 5.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0}, {0.249997f, 0.249997f, 0.249997f, 0.249997f, 0.000011f}, 1.0f, 5.0f, 5.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f);
test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f);
printf("OK\n");