mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-23 23:18:26 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'ollama:main' into main

Browse Source
This commit is contained in:
likelovewant
2024-05-31 18:43:24 +08:00
committed by GitHub
parent cafde1f8ce 763bb65dbb
commit 16ce79eb3b
19 changed files with 909 additions and 487 deletions
Download Patch File Download Diff File Expand all files Collapse all files

150
llm/ext_server/server.cpp vendored
View File

@@ -140,7 +140,6 @@ struct server_slot {
std::vector<llama_token> cache_tokens;
std::vector<completion_token_output> generated_token_probs;
bool infill = false;
bool embedding = false;
bool has_next_token = true;
bool truncated = false;
@@ -187,7 +186,6 @@ struct server_slot {
n_past = 0;
n_sent_text = 0;
n_sent_token_probs = 0;
infill = false;
ga_i = 0;
n_past_se = 0;
@@ -600,16 +598,6 @@ struct llama_server_context
slot->params.n_predict = slot->n_predict;
}
// infill
if (data.count("input_prefix") != 0)
{
slot->params.input_prefix = data["input_prefix"];
}
else
{
slot->params.input_prefix = "";
}
if (data.count("input_suffix") != 0)
{
slot->params.input_suffix = data["input_suffix"];
@@ -897,15 +885,6 @@ struct llama_server_context
system_need_update = true;
}
void system_prompt_process(const json &sys_props) {
system_prompt = sys_props.value("prompt", "");
name_user = sys_props.value("anti_prompt", "");
name_assistant = sys_props.value("assistant_name", "");
system_prompt_notify();
}
static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
const stop_type type, server_slot &slot)
{
@@ -1263,13 +1242,12 @@ struct llama_server_context
queue_results.send(res);
}
void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id)
void request_completion(int task_id, json data, bool embedding, int multitask_id)
{
task_server task;
task.id = task_id;
task.target_id = 0;
task.data = std::move(data);
task.infill_mode = infill;
task.embedding_mode = embedding;
task.type = TASK_TYPE_COMPLETION;
task.multitask_id = multitask_id;
@@ -1415,8 +1393,8 @@ struct llama_server_context
json subtask_data = multiprompt_task.data;
subtask_data["prompt"] = subtask_data["prompt"][i];
// subtasks inherit everything else (infill mode, embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
// subtasks inherit everything else (embedding mode, etc.)
request_completion(subtask_ids[i], subtask_data, multiprompt_task.embedding_mode, multitask_id);
}
}
@@ -1434,26 +1412,8 @@ struct llama_server_context
break;
}
if (task.data.contains("system_prompt"))
{
if (!all_slots_are_idle) {
send_error(task, "system prompt can only be updated when all slots are idle");
break;
}
system_prompt_process(task.data["system_prompt"]);
// reset cache_tokens for all slots
for (server_slot &slot : slots)
{
slot.cache_tokens.clear();
slot.n_past = 0;
slot.n_past_se = 0;
}
}
slot->reset();
slot->infill = task.infill_mode;
slot->embedding = task.embedding_mode;
slot->task_id = task.id;
slot->multitask_id = task.multitask_id;
@@ -1679,8 +1639,7 @@ struct llama_server_context
const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
// empty prompt passed -> release the slot and send empty response
// note: infill mode allows empty prompt
if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
{
slot.release();
slot.print_timings();
@@ -1697,33 +1656,7 @@ struct llama_server_context
slot.t_start_process_prompt = ggml_time_us();
slot.t_start_genereration = 0;
if (slot.infill)
{
bool suff_rm_leading_spc = true;
if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1)
{
params.input_suffix.erase(0, 1);
suff_rm_leading_spc = false;
}
auto prefix_tokens = tokenize(slot.params.input_prefix, false);
auto suffix_tokens = tokenize(slot.params.input_suffix, false);
const int space_token = 29871; // TODO: this should not be hardcoded
if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
suffix_tokens.erase(suffix_tokens.begin());
}
prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
prefix_tokens.push_back(llama_token_middle(model));
prompt_tokens = prefix_tokens;
}
else
{
prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
}
prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt
slot.n_prompt_tokens = prompt_tokens.size();
@@ -2130,8 +2063,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf("\n");
}
static void server_params_parse(int argc, char **argv, server_params &sparams,
gpt_params &params, llama_server_context& llama)
static void server_params_parse(int argc, char **argv, server_params &sparams, gpt_params &params)
{
gpt_params default_params;
server_params default_sparams;
@@ -2546,27 +2478,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
}
params.n_predict = std::stoi(argv[i]);
}
else if (arg == "-spf" || arg == "--system-prompt-file")
{
if (++i >= argc)
{
invalid_param = true;
break;
}
std::ifstream file(argv[i]);
if (!file) {
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
invalid_param = true;
break;
}
std::string systm_content;
std::copy(
std::istreambuf_iterator<char>(file),
std::istreambuf_iterator<char>(),
std::back_inserter(systm_content)
);
llama.system_prompt_process(json::parse(systm_content));
}
else if (arg == "-ctk" || arg == "--cache-type-k") {
params.cache_type_k = argv[++i];
}
@@ -2714,21 +2625,6 @@ static json format_partial_response(
return res;
}
static json format_tokenizer_response(const std::vector<llama_token> &tokens)
{
return json {
{"tokens", tokens}
};
}
static json format_detokenized_response(std::string content)
{
return json {
{"content", content}
};
}
static void log_server_request(const httplib::Request &req, const httplib::Response &res)
{
// skip GH copilot requests when using default port
@@ -2818,7 +2714,7 @@ int main(int argc, char **argv) {
// struct that contains llama context and inference
llama_server_context llama;
server_params_parse(argc, argv, sparams, params, llama);
server_params_parse(argc, argv, sparams, params);
if (params.model_alias == "unknown")
{
@@ -3150,7 +3046,7 @@ int main(int argc, char **argv) {
json data = json::parse(req.body);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, data, false, false, -1);
llama.request_completion(task_id, data, false, -1);
if (!json_value(data, "stream", false)) {
std::string completion_text;
task_result result = llama.queue_results.recv(task_id);
@@ -3218,34 +3114,6 @@ int main(int argc, char **argv) {
}
});
svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
const json body = json::parse(req.body);
std::vector<llama_token> tokens;
if (body.count("content") != 0)
{
tokens = llama.tokenize(body["content"], false);
}
const json data = format_tokenizer_response(tokens);
return res.set_content(data.dump(), "application/json; charset=utf-8");
});
svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
const json body = json::parse(req.body);
std::string content;
if (body.count("tokens") != 0)
{
const std::vector<llama_token> tokens = body["tokens"];
content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend());
}
const json data = format_detokenized_response(content);
return res.set_content(data.dump(), "application/json; charset=utf-8");
});
svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
{
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
@@ -3272,7 +3140,7 @@ int main(int argc, char **argv) {
// create and queue the task
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1);
llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, true, -1);
// get the result
task_result result = llama.queue_results.recv(task_id);

85
llm/generate/gen_darwin.sh
View File

@@ -32,42 +32,43 @@ case "${GOARCH}" in
echo "Building static library"
build
if [ -z "$OLLAMA_SKIP_CPU_GENERATE" ]; then
#
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
echo "Building LCD CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
#
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
echo "Building LCD CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
#
# ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
# Approximately 400% faster than LCD on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
#
# ~2011 CPU Dynamic library with more capabilities turned on to optimize performance
# Approximately 400% faster than LCD on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
#
# ~2013 CPU Dynamic library
# Approximately 10% faster than AVX on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
#
# ~2013 CPU Dynamic library
# Approximately 10% faster than AVX on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
fi
;;
"arm64")
@@ -79,13 +80,15 @@ case "${GOARCH}" in
echo "Building static library"
build
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/metal"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/metal"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build
sign ${BUILD_DIR}/bin/ollama_llama_server
compress
fi
;;
*)
echo "GOARCH must be set"

2
llm/llama.cpp

Submodule llm/llama.cpp updated: 74f33adf5f...5921b8f089

60
llm/llm.go
View File

@@ -12,6 +12,7 @@ package llm
import "C"
import (
"fmt"
"strings"
"unsafe"
)
@@ -37,3 +38,62 @@ func Quantize(infile, outfile string, ftype fileType) error {
return nil
}
type llamaModel struct {
m *C.struct_llama_model
}
func newLlamaModel(p string) *llamaModel {
cs := C.CString(p)
defer C.free(unsafe.Pointer(cs))
params := C.llama_model_default_params()
params.vocab_only = true
return &llamaModel{
C.llama_load_model_from_file(cs, params),
}
}
func (llm *llamaModel) Close() {
C.llama_free_model(llm.m)
}
func (llm *llamaModel) Tokenize(s string) []int {
cs := C.CString(s)
defer C.free(unsafe.Pointer(cs))
ltokens := make([]C.llama_token, len(s)+2)
n := C.llama_tokenize(
llm.m,
cs,
C.int32_t(len(s)),
&ltokens[0],
C.int32_t(len(ltokens)),
false,
true,
)
if n < 0 {
return nil
}
tokens := make([]int, n)
for i := 0; i < int(n); i++ {
tokens[i] = int(ltokens[i])
}
return tokens
}
func (llm *llamaModel) Detokenize(i32s []int) string {
var sb strings.Builder
for _, i32 := range i32s {
c := make([]byte, 512)
if n := C.llama_token_to_piece(llm.m, C.llama_token(i32), (*C.char)(unsafe.Pointer(&c[0])), C.int(len(c)), false); n > 0 {
sb.WriteString(unsafe.String(&c[0], n))
}
}
return sb.String()
}

45
llm/patches/05-default-pretokenizer.diff
View File

@@ -1,35 +1,32 @@
From d02a06f3f45a09255ace8684a66590e06ce44605 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Thu, 23 May 2024 11:33:20 -0700
Subject: [PATCH] default pretokenizer on unrecognized type
---
llama.cpp | 5 +----
1 file changed, 1 insertion(+), 4 deletions(-)
diff --git a/llama.cpp b/llama.cpp
index 15c66077..af1aede3 100644
index 40d2ec2c..74f3ee9c 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -4504,9 +4504,6 @@ static void llm_load_vocab(
LLAMA_LOG_WARN("%s: ************************************ \n", __func__);
LLAMA_LOG_WARN("%s: \n", __func__);
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
- } else if (
- tokenizer_pre == "default") {
@@ -4642,16 +4642,7 @@ static void llm_load_vocab(
// for now, only BPE models have pre-tokenizers
if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
- if (tokenizer_pre.empty()) {
- LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__);
- LLAMA_LOG_WARN("%s: \n", __func__);
- LLAMA_LOG_WARN("%s: ************************************ \n", __func__);
- LLAMA_LOG_WARN("%s: GENERATION QUALITY WILL BE DEGRADED! \n", __func__);
- LLAMA_LOG_WARN("%s: CONSIDER REGENERATING THE MODEL \n", __func__);
- LLAMA_LOG_WARN("%s: ************************************ \n", __func__);
- LLAMA_LOG_WARN("%s: \n", __func__);
- vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
- } else if (
+ if (
tokenizer_pre == "default") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
} else if (
tokenizer_pre == "llama3" ||
tokenizer_pre == "llama-v3" ||
@@ -4553,7 +4550,7 @@ static void llm_load_vocab(
tokenizer_pre == "dbrx") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX;
@@ -4703,7 +4694,8 @@ static void llm_load_vocab(
tokenizer_pre == "smaug-bpe") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
} else {
- throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
+ LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
+ vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
}
} else {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--
2.45.1

154
llm/server.go
View File

@@ -57,6 +57,8 @@ type llmServer struct {
loadDuration time.Duration // Record how long it took the model to load
loadProgress float32
*llamaModel
sem *semaphore.Weighted
}
@@ -189,35 +191,38 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
params = append(params, "--memory-f32")
}
if opts.UseMLock {
params = append(params, "--mlock")
flashAttnEnabled := envconfig.FlashAttention
for _, g := range gpus {
// only cuda (compute capability 7+) and metal support flash attention
if g.Library != "metal" && (g.Library != "cuda" || g.DriverMajor < 7) {
flashAttnEnabled = false
}
// mmap has issues with partial offloading on metal
if g.Library == "metal" &&
uint64(opts.NumGPU) > 0 &&
uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
opts.UseMMap = false
}
}
if flashAttnEnabled {
params = append(params, "--flash-attn")
}
if !opts.UseMMap {
params = append(params, "--no-mmap")
}
if opts.UseMLock {
params = append(params, "--mlock")
}
if opts.UseNUMA {
params = append(params, "--numa")
}
flashAttnEnabled := envconfig.FlashAttention
// partial offloading does not support flash attention
if uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
flashAttnEnabled = false
}
// only cuda (compute capability 7+) and metal support flash attention
for _, g := range gpus {
if g.Library != "metal" && (g.Library != "cuda" || g.DriverMajor < 7) {
flashAttnEnabled = false
}
}
if flashAttnEnabled {
params = append(params, "--flash-attn")
}
numParallel := envconfig.NumParallel
// TODO (jmorganca): multimodal models don't support parallel yet
@@ -306,6 +311,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
totalLayers: ggml.KV().BlockCount() + 1,
gpuCount: gpuCount,
done: make(chan error, 1),
llamaModel: newLlamaModel(model),
}
s.cmd.Env = os.Environ()
@@ -843,12 +849,12 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(TokenizeRequest{Content: prompt})
if err != nil {
var b bytes.Buffer
if err := json.NewEncoder(&b).Encode(EmbeddingRequest{Content: prompt}); err != nil {
return nil, fmt.Errorf("error marshaling embed data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), bytes.NewBuffer(data))
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/embedding", s.port), &b)
if err != nil {
return nil, fmt.Errorf("error creating embed request: %w", err)
}
@@ -878,108 +884,12 @@ func (s *llmServer) Embedding(ctx context.Context, prompt string) ([]float64, er
return embedding.Embedding, nil
}
type TokenizeRequest struct {
Content string `json:"content"`
}
type TokenizeResponse struct {
Tokens []int `json:"tokens"`
}
func (s *llmServer) Tokenize(ctx context.Context, content string) ([]int, error) {
// Make sure the server is ready
status, err := s.getServerStatus(ctx)
if err != nil {
return nil, err
} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
return nil, fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(TokenizeRequest{Content: content})
if err != nil {
return nil, fmt.Errorf("marshaling encode data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/tokenize", s.port), bytes.NewBuffer(data))
if err != nil {
return nil, fmt.Errorf("encode request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
resp, err := http.DefaultClient.Do(req)
if err != nil {
return nil, fmt.Errorf("do encode request: %w", err)
}
defer resp.Body.Close()
body, err := io.ReadAll(resp.Body)
if err != nil {
return nil, fmt.Errorf("read encode request: %w", err)
}
if resp.StatusCode >= 400 {
log.Printf("llm encode error: %s", body)
return nil, fmt.Errorf("%s", body)
}
var encoded TokenizeResponse
if err := json.Unmarshal(body, &encoded); err != nil {
return nil, fmt.Errorf("unmarshal encode response: %w", err)
}
return encoded.Tokens, nil
}
type DetokenizeRequest struct {
Tokens []int `json:"tokens"`
}
type DetokenizeResponse struct {
Content string `json:"content"`
return s.llamaModel.Tokenize(content), nil
}
func (s *llmServer) Detokenize(ctx context.Context, tokens []int) (string, error) {
// Make sure the server is ready
status, err := s.getServerStatus(ctx)
if err != nil {
return "", err
} else if status != ServerStatusReady && status != ServerStatusNoSlotsAvailable {
return "", fmt.Errorf("unexpected server status: %s", status.ToString())
}
data, err := json.Marshal(DetokenizeRequest{Tokens: tokens})
if err != nil {
return "", fmt.Errorf("marshaling decode data: %w", err)
}
req, err := http.NewRequestWithContext(ctx, http.MethodPost, fmt.Sprintf("http://127.0.0.1:%d/detokenize", s.port), bytes.NewBuffer(data))
if err != nil {
return "", fmt.Errorf("decode request: %w", err)
}
req.Header.Set("Content-Type", "application/json")
resp, err := http.DefaultClient.Do(req)
if err != nil {
return "", fmt.Errorf("do decode request: %w", err)
}
defer resp.Body.Close()
body, err := io.ReadAll(resp.Body)
if err != nil {
return "", fmt.Errorf("read decode request: %w", err)
}
if resp.StatusCode >= 400 {
log.Printf("llm decode error: %s", body)
return "", fmt.Errorf("%s", body)
}
var decoded DetokenizeResponse
if err := json.Unmarshal(body, &decoded); err != nil {
return "", fmt.Errorf("unmarshal encode response: %w", err)
}
return decoded.Content, nil
return s.llamaModel.Detokenize(tokens), nil
}
func (s *llmServer) Close() error {
@@ -997,6 +907,10 @@ func (s *llmServer) Close() error {
slog.Debug("llama server stopped")
}
if s.llamaModel != nil {
s.llamaModel.Close()
}
return nil
}