Merge branch 'ggerganov:master' into sycl_cmake
This commit is contained in:
Abhilash Majumder
GitHub
commit
27899c1c79
11 changed files with 246 additions and 28 deletions
4
Makefile
4
Makefile
|
|
@ -867,3 +867,7 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
|
|||
tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
|
||||
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
|
||||
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
|
||||
|
||||
tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
|
||||
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
|
||||
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
|
||||
|
|
|
|||
|
|
@ -61,7 +61,7 @@ variety of hardware - locally and in the cloud.
|
|||
- Plain C/C++ implementation without any dependencies
|
||||
- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
|
||||
- AVX, AVX2 and AVX512 support for x86 architectures
|
||||
- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
|
||||
- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
|
||||
- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
|
||||
- Vulkan, SYCL, and (partial) OpenCL backend support
|
||||
- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
|
||||
|
|
|
|||
|
|
@ -1533,16 +1533,17 @@ int main(int argc, char ** argv) {
|
|||
|
||||
int n_past = 0;
|
||||
|
||||
ggml_cgraph gf = {};
|
||||
struct ggml_cgraph * gf = NULL;
|
||||
gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
|
||||
|
||||
get_example_targets_batch(ctx0, 64*ex+0, tokens_input, targets);
|
||||
|
||||
struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
|
||||
struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
|
||||
// struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
|
||||
struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
|
||||
|
||||
ggml_build_forward_expand(&gf, e);
|
||||
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
|
||||
ggml_build_forward_expand(gf, e);
|
||||
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
|
||||
|
||||
float error_before_opt = ggml_get_f32_1d(e, 0);
|
||||
|
||||
|
|
@ -1552,8 +1553,8 @@ int main(int argc, char ** argv) {
|
|||
opt_params_lbfgs.lbfgs.n_iter = 16;
|
||||
ggml_opt(ctx0, opt_params_lbfgs, e);
|
||||
//
|
||||
ggml_build_forward_expand(&gf, e);
|
||||
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
|
||||
ggml_build_forward_expand(gf, e);
|
||||
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
|
||||
|
||||
float error_after_opt = ggml_get_f32_1d(e, 0);
|
||||
|
||||
|
|
@ -1600,13 +1601,14 @@ int main(int argc, char ** argv) {
|
|||
};
|
||||
struct ggml_context * ctx0 = ggml_init(params);
|
||||
|
||||
ggml_cgraph gf = {};
|
||||
struct ggml_cgraph * gf = NULL;
|
||||
gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
|
||||
|
||||
int n_past = 0;
|
||||
struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
|
||||
struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
|
||||
|
||||
ggml_build_forward_expand(&gf, logits);
|
||||
ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
|
||||
ggml_build_forward_expand(gf, logits);
|
||||
ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
|
||||
|
||||
struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
|
||||
struct ggml_tensor * probs = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);
|
||||
|
|
|
|||
|
|
@ -25,9 +25,6 @@ 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"):
|
||||
|
|
@ -35,7 +32,6 @@ if len(clip_tensors) > 0:
|
|||
f.write("{}\n")
|
||||
|
||||
|
||||
torch.save(checkpoint, path)
|
||||
|
||||
print("Done!")
|
||||
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
|
||||
|
|
|
|||
|
|
@ -6205,7 +6205,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
|
|||
const int ix = rowx*ncols + col;
|
||||
const int iy = rowy*ncols + col;
|
||||
|
||||
const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + slope*pos[col];
|
||||
const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
|
||||
|
||||
vals[col] = val;
|
||||
max_val = max(max_val, val);
|
||||
|
|
@ -9170,17 +9170,17 @@ static void ggml_cuda_op_soft_max(
|
|||
memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
|
||||
|
||||
// positions tensor
|
||||
float * src2_dd = dst_dd; // default to avoid null checks in the kernel
|
||||
float * src2_dd = nullptr;
|
||||
cuda_pool_alloc<float> src2_f;
|
||||
|
||||
ggml_tensor * src2 = dst->src[2];
|
||||
const bool use_src2 = src2 != nullptr;
|
||||
|
||||
if (use_src2) {
|
||||
const bool src2_on_device = use_src2 && src2->backend == GGML_BACKEND_GPU;
|
||||
ggml_tensor_extra_gpu * src2_extra = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr;
|
||||
const bool src2_on_device = src2->backend == GGML_BACKEND_GPU;
|
||||
|
||||
if (src2_on_device) {
|
||||
ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
|
||||
src2_dd = (float *) src2_extra->data_device[g_main_device];
|
||||
} else {
|
||||
src2_dd = src2_f.alloc(ggml_nelements(src2));
|
||||
|
|
|
|||
|
|
@ -392,7 +392,7 @@ kernel void kernel_soft_max(
|
|||
float lmax = -INFINITY;
|
||||
|
||||
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
|
||||
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
|
||||
lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
|
||||
}
|
||||
|
||||
// find the max value in the block
|
||||
|
|
@ -417,7 +417,7 @@ kernel void kernel_soft_max(
|
|||
// parallel sum
|
||||
float lsum = 0.0f;
|
||||
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
|
||||
const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
|
||||
const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
|
||||
lsum += exp_psrc0;
|
||||
pdst[i00] = exp_psrc0;
|
||||
}
|
||||
|
|
@ -495,7 +495,7 @@ kernel void kernel_soft_max_4(
|
|||
float4 lmax4 = -INFINITY;
|
||||
|
||||
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
|
||||
lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]);
|
||||
lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
|
||||
}
|
||||
|
||||
const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
|
||||
|
|
@ -521,7 +521,7 @@ kernel void kernel_soft_max_4(
|
|||
// parallel sum
|
||||
float4 lsum4 = 0.0f;
|
||||
for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
|
||||
const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + slope*ppos[i00]) - max_val);
|
||||
const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
|
||||
lsum4 += exp_psrc4;
|
||||
pdst4[i00] = exp_psrc4;
|
||||
}
|
||||
|
|
|
|||
19
ggml.c
19
ggml.c
|
|
@ -23,6 +23,9 @@
|
|||
#include <limits.h>
|
||||
#include <stdarg.h>
|
||||
#include <signal.h>
|
||||
#if defined(__gnu_linux__)
|
||||
#include <syscall.h>
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_METAL
|
||||
#include <unistd.h>
|
||||
|
|
@ -1971,7 +1974,7 @@ struct ggml_numa_nodes {
|
|||
uint32_t n_nodes;
|
||||
uint32_t total_cpus; // hardware threads on system
|
||||
uint32_t current_node; // node on which main process is execting
|
||||
#ifdef __linux__
|
||||
#if defined(__gnu_linux__)
|
||||
cpu_set_t cpuset; // cpuset from numactl
|
||||
#else
|
||||
uint32_t cpuset; // no NUMA support outside of Linux at this time. Use a portable datatype
|
||||
|
|
@ -2009,7 +2012,7 @@ inline static void ggml_critical_section_end(void) {
|
|||
atomic_fetch_sub(&g_state_barrier, 1);
|
||||
}
|
||||
|
||||
#ifdef __linux__
|
||||
#if defined(__gnu_linux__)
|
||||
static cpu_set_t ggml_get_numa_affinity(void) {
|
||||
cpu_set_t cpuset;
|
||||
pthread_t thread;
|
||||
|
|
@ -2031,7 +2034,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
|
|||
return;
|
||||
}
|
||||
|
||||
#ifdef __linux__
|
||||
#if defined(__gnu_linux__)
|
||||
struct stat st;
|
||||
char path[256];
|
||||
int rv;
|
||||
|
|
@ -2063,7 +2066,13 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
|
|||
|
||||
// figure out which node we're on
|
||||
uint current_cpu;
|
||||
int getcpu_ret = getcpu(¤t_cpu, &g_state.numa.current_node);
|
||||
int getcpu_ret = 0;
|
||||
#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28)
|
||||
getcpu_ret = getcpu(¤t_cpu, &g_state.numa.current_node);
|
||||
#else
|
||||
// old glibc doesn't have a wrapper for this call. Fall back on direct syscall
|
||||
getcpu_ret = syscall(SYS_getcpu,¤t_cpu,&g_state.numa.current_node);
|
||||
#endif
|
||||
|
||||
if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) {
|
||||
g_state.numa.n_nodes = 0;
|
||||
|
|
@ -16734,7 +16743,7 @@ typedef pthread_t ggml_thread_t;
|
|||
#endif
|
||||
|
||||
// Android's libc implementation "bionic" does not support setting affinity
|
||||
#if defined(__linux__) && !defined(__BIONIC__)
|
||||
#if defined(__gnu_linux__)
|
||||
static void set_numa_thread_affinity(int thread_n) {
|
||||
if (!ggml_is_numa()) {
|
||||
return;
|
||||
|
|
|
|||
117
llama.cpp
117
llama.cpp
|
|
@ -12508,6 +12508,123 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
|
|||
return 0;
|
||||
}
|
||||
|
||||
// trim whitespace from the beginning and end of a string
|
||||
static std::string trim(const std::string & str) {
|
||||
size_t start = 0;
|
||||
size_t end = str.size();
|
||||
while (start < end && isspace(str[start])) {
|
||||
start += 1;
|
||||
}
|
||||
while (end > start && isspace(str[end - 1])) {
|
||||
end -= 1;
|
||||
}
|
||||
return str.substr(start, end - start);
|
||||
}
|
||||
|
||||
// Simple version of "llama_apply_chat_template" that only works with strings
|
||||
// This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
|
||||
static int32_t llama_chat_apply_template_internal(
|
||||
const std::string & tmpl,
|
||||
const std::vector<const llama_chat_message *> & chat,
|
||||
std::string & dest, bool add_ass) {
|
||||
// Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
|
||||
std::stringstream ss;
|
||||
if (tmpl.find("<|im_start|>") != std::string::npos) {
|
||||
// chatml template
|
||||
for (auto message : chat) {
|
||||
ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
|
||||
}
|
||||
if (add_ass) {
|
||||
ss << "<|im_start|>assistant\n";
|
||||
}
|
||||
} else if (tmpl.find("[INST]") != std::string::npos) {
|
||||
// llama2 template and its variants
|
||||
// [variant] support system message
|
||||
bool support_system_message = tmpl.find("<<SYS>>") != std::string::npos;
|
||||
// [variant] space before + after response
|
||||
bool space_around_response = tmpl.find("' ' + eos_token") != std::string::npos;
|
||||
// [variant] add BOS inside history
|
||||
bool add_bos_inside_history = tmpl.find("bos_token + '[INST]") != std::string::npos;
|
||||
// [variant] trim spaces from the input message
|
||||
bool strip_message = tmpl.find("content.strip()") != std::string::npos;
|
||||
// construct the prompt
|
||||
bool is_inside_turn = true; // skip BOS at the beginning
|
||||
ss << "[INST] ";
|
||||
for (auto message : chat) {
|
||||
std::string content = strip_message ? trim(message->content) : message->content;
|
||||
std::string role(message->role);
|
||||
if (!is_inside_turn) {
|
||||
is_inside_turn = true;
|
||||
ss << (add_bos_inside_history ? "<s>[INST] " : "[INST] ");
|
||||
}
|
||||
if (role == "system") {
|
||||
if (support_system_message) {
|
||||
ss << "<<SYS>>\n" << content << "\n<</SYS>>\n\n";
|
||||
} else {
|
||||
// if the model does not support system message, we still include it in the first message, but without <<SYS>>
|
||||
ss << content << "\n";
|
||||
}
|
||||
} else if (role == "user") {
|
||||
ss << content << " [/INST]";
|
||||
} else {
|
||||
ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << "</s>";
|
||||
is_inside_turn = false;
|
||||
}
|
||||
}
|
||||
// llama2 templates seem to not care about "add_generation_prompt"
|
||||
} else if (tmpl.find("<|user|>") != std::string::npos) {
|
||||
// zephyr template
|
||||
for (auto message : chat) {
|
||||
ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
|
||||
}
|
||||
if (add_ass) {
|
||||
ss << "<|assistant|>\n";
|
||||
}
|
||||
} else {
|
||||
// template not supported
|
||||
return -1;
|
||||
}
|
||||
dest = ss.str();
|
||||
return dest.size();
|
||||
}
|
||||
|
||||
LLAMA_API int32_t llama_chat_apply_template(
|
||||
const struct llama_model * model,
|
||||
const char * tmpl,
|
||||
const struct llama_chat_message * chat,
|
||||
size_t n_msg,
|
||||
bool add_ass,
|
||||
char * buf,
|
||||
int32_t length) {
|
||||
std::string curr_tmpl(tmpl == nullptr ? "" : tmpl);
|
||||
if (tmpl == nullptr) {
|
||||
GGML_ASSERT(model != nullptr);
|
||||
// load template from model
|
||||
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
|
||||
std::string template_key = "tokenizer.chat_template";
|
||||
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), curr_tmpl.size());
|
||||
if (res < 0) {
|
||||
// worst case: there is no information about template, we will use chatml by default
|
||||
curr_tmpl = "<|im_start|>"; // see llama_chat_apply_template_internal
|
||||
} else {
|
||||
curr_tmpl = std::string(model_template.data(), model_template.size());
|
||||
}
|
||||
}
|
||||
// format the chat to string
|
||||
std::vector<const llama_chat_message *> chat_vec;
|
||||
chat_vec.resize(n_msg);
|
||||
for (size_t i = 0; i < n_msg; i++) {
|
||||
chat_vec[i] = &chat[i];
|
||||
}
|
||||
std::string formatted_chat;
|
||||
int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass);
|
||||
if (res < 0) {
|
||||
return res;
|
||||
}
|
||||
strncpy(buf, formatted_chat.c_str(), length);
|
||||
return res;
|
||||
}
|
||||
|
||||
struct llama_timings llama_get_timings(struct llama_context * ctx) {
|
||||
struct llama_timings result = {
|
||||
/*.t_start_ms =*/ 1e-3 * ctx->t_start_us,
|
||||
|
|
|
|||
25
llama.h
25
llama.h
|
|
@ -305,6 +305,12 @@ extern "C" {
|
|||
int32_t n_eval;
|
||||
};
|
||||
|
||||
// used in chat template
|
||||
typedef struct llama_chat_message {
|
||||
const char * role;
|
||||
const char * content;
|
||||
} llama_chat_message;
|
||||
|
||||
// Helpers for getting default parameters
|
||||
LLAMA_API struct llama_model_params llama_model_default_params(void);
|
||||
LLAMA_API struct llama_context_params llama_context_default_params(void);
|
||||
|
|
@ -699,6 +705,25 @@ extern "C" {
|
|||
char * buf,
|
||||
int32_t length);
|
||||
|
||||
/// Apply chat template. Inspired by hf apply_chat_template() on python.
|
||||
/// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
|
||||
/// NOTE: This function only support some known jinja templates. It is not a jinja parser.
|
||||
/// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead.
|
||||
/// @param chat Pointer to a list of multiple llama_chat_message
|
||||
/// @param n_msg Number of llama_chat_message in this chat
|
||||
/// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message.
|
||||
/// @param buf A buffer to hold the output formatted prompt. The recommended alloc size is 2 * (total number of characters of all messages)
|
||||
/// @param length The size of the allocated buffer
|
||||
/// @return The total number of bytes of the formatted prompt. If is it larger than the size of buffer, you may need to re-alloc it and then re-apply the template.
|
||||
LLAMA_API int32_t llama_chat_apply_template(
|
||||
const struct llama_model * model,
|
||||
const char * tmpl,
|
||||
const struct llama_chat_message * chat,
|
||||
size_t n_msg,
|
||||
bool add_ass,
|
||||
char * buf,
|
||||
int32_t length);
|
||||
|
||||
//
|
||||
// Grammar
|
||||
//
|
||||
|
|
|
|||
|
|
@ -28,6 +28,7 @@ endfunction()
|
|||
llama_build_and_test_executable(test-quantize-fns.cpp)
|
||||
llama_build_and_test_executable(test-quantize-perf.cpp)
|
||||
llama_build_and_test_executable(test-sampling.cpp)
|
||||
llama_build_and_test_executable(test-chat-template.cpp)
|
||||
|
||||
llama_build_executable(test-tokenizer-0-llama.cpp)
|
||||
llama_test_executable (test-tokenizer-0-llama test-tokenizer-0-llama.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama.gguf)
|
||||
|
|
|
|||
64
tests/test-chat-template.cpp
Normal file
64
tests/test-chat-template.cpp
Normal file
|
|
@ -0,0 +1,64 @@
|
|||
#include <iostream>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
#include <sstream>
|
||||
|
||||
#undef NDEBUG
|
||||
#include <cassert>
|
||||
|
||||
#include "llama.h"
|
||||
|
||||
int main(void) {
|
||||
llama_chat_message conversation[] = {
|
||||
{"system", "You are a helpful assistant"},
|
||||
{"user", "Hello"},
|
||||
{"assistant", "Hi there"},
|
||||
{"user", "Who are you"},
|
||||
{"assistant", " I am an assistant "},
|
||||
{"user", "Another question"},
|
||||
};
|
||||
size_t message_count = 6;
|
||||
std::vector<std::string> templates = {
|
||||
// teknium/OpenHermes-2.5-Mistral-7B
|
||||
"{% for message in messages %}{{'<|im_start|>' + message['role'] + '\\n' + message['content'] + '<|im_end|>' + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\\n' }}{% endif %}",
|
||||
// mistralai/Mistral-7B-Instruct-v0.2
|
||||
"{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
|
||||
// TheBloke/FusionNet_34Bx2_MoE-AWQ
|
||||
"{%- for idx in range(0, messages|length) -%}\\n{%- if messages[idx]['role'] == 'user' -%}\\n{%- if idx > 1 -%}\\n{{- bos_token + '[INST] ' + messages[idx]['content'] + ' [/INST]' -}}\\n{%- else -%}\\n{{- messages[idx]['content'] + ' [/INST]' -}}\\n{%- endif -%}\\n{% elif messages[idx]['role'] == 'system' %}\\n{{- '[INST] <<SYS>>\\\\n' + messages[idx]['content'] + '\\\\n<</SYS>>\\\\n\\\\n' -}}\\n{%- elif messages[idx]['role'] == 'assistant' -%}\\n{{- ' ' + messages[idx]['content'] + ' ' + eos_token -}}\\n{% endif %}\\n{% endfor %}",
|
||||
// bofenghuang/vigogne-2-70b-chat
|
||||
"{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif true == true and not '<<SYS>>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'Vous êtes Vigogne, un assistant IA créé par Zaion Lab. Vous suivez extrêmement bien les instructions. Aidez autant que vous le pouvez.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<<SYS>>\\\\n' + system_message + '\\\\n<</SYS>>\\\\n\\\\n' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<<SYS>>\\\\n' + content.strip() + '\\\\n<</SYS>>\\\\n\\\\n' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}",
|
||||
};
|
||||
std::vector<std::string> expected_substr = {
|
||||
"<|im_start|>assistant\n I am an assistant <|im_end|>\n<|im_start|>user\nAnother question<|im_end|>\n<|im_start|>assistant",
|
||||
"[/INST]Hi there</s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
|
||||
"</s><s>[INST] Who are you [/INST] I am an assistant </s><s>[INST] Another question [/INST]",
|
||||
"[/INST] Hi there </s>[INST] Who are you [/INST] I am an assistant </s>[INST] Another question [/INST]",
|
||||
};
|
||||
std::vector<char> formatted_chat(1024);
|
||||
int32_t res;
|
||||
|
||||
// test invalid chat template
|
||||
res = llama_chat_apply_template(nullptr, "INVALID TEMPLATE", conversation, message_count, true, formatted_chat.data(), formatted_chat.size());
|
||||
assert(res < 0);
|
||||
|
||||
for (size_t i = 0; i < templates.size(); i++) {
|
||||
std::string custom_template = templates[i];
|
||||
std::string substr = expected_substr[i];
|
||||
formatted_chat.resize(1024);
|
||||
res = llama_chat_apply_template(
|
||||
nullptr,
|
||||
custom_template.c_str(),
|
||||
conversation,
|
||||
message_count,
|
||||
true,
|
||||
formatted_chat.data(),
|
||||
formatted_chat.size()
|
||||
);
|
||||
formatted_chat.resize(res);
|
||||
std::string output(formatted_chat.data(), formatted_chat.size());
|
||||
std::cout << output << "\n-------------------------\n";
|
||||
// expect the "formatted_chat" to contain pre-defined strings
|
||||
assert(output.find(substr) != std::string::npos);
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue