id: cmake_test
run: |
cd build
- ctest -L main --verbose --timeout 900
+ ctest -L 'main|curl' --verbose --timeout 900
- name: Determine tag name
id: tag
id: depends
run: |
sudo apt-get update
- sudo apt-get install build-essential
+ sudo apt-get install build-essential libcurl4-openssl-dev
- name: Build
id: cmake_build
run: |
mkdir build
cd build
- cmake .. -DLLAMA_FATAL_WARNINGS=ON
+ cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON
cmake --build . --config Release -j $(nproc)
- name: Test
id: cmake_test
run: |
cd build
- ctest -L main --verbose --timeout 900
+ ctest -L 'main|curl' --verbose --timeout 900
- name: Test llama2c conversion
id: llama2c_test
/convert-llama2c-to-ggml
/embd-input-test
/embedding
+/eval-callback
/gguf
/gguf-llama-simple
/gguf-split
# Define the default target now so that it is always the first target
BUILD_TARGETS = \
main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
- simple batched batched-bench save-load-state server gguf gguf-split llama-bench libllava.a llava-cli baby-llama beam-search \
+ simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search \
retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
# Binaries only useful for tests
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+eval-callback: examples/eval-callback/eval-callback.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)
+
train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
cparams.yarn_orig_ctx = params.yarn_orig_ctx;
cparams.pooling_type = params.pooling_type;
cparams.defrag_thold = params.defrag_thold;
+ cparams.cb_eval = params.cb_eval;
+ cparams.cb_eval_user_data = params.cb_eval_user_data;
cparams.offload_kqv = !params.no_kv_offload;
cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
}
- {
+ if (params.warmup) {
LOG("warming up the model with an empty run\n");
std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
int32_t yarn_orig_ctx = 0; // YaRN original context length
float defrag_thold = -1.0f; // KV cache defragmentation threshold
+ ggml_backend_sched_eval_callback cb_eval = nullptr;
+ void * cb_eval_user_data = nullptr;
+
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
bool infill = false; // use infill mode
bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes
bool no_kv_offload = false; // disable KV offloading
+ bool warmup = true; // warmup run
std::string cache_type_k = "f16"; // KV cache data type for the K
std::string cache_type_v = "f16"; // KV cache data type for the V
When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support of missing backend operations can be added in another PR.
+Note: to debug the inference graph: you can use [eval-callback](../examples/eval-callback).
+
## GGUF specification
https://github.com/ggerganov/ggml/blob/master/docs/gguf.md
add_subdirectory(benchmark)
add_subdirectory(convert-llama2c-to-ggml)
add_subdirectory(embedding)
+ add_subdirectory(eval-callback)
add_subdirectory(finetune)
add_subdirectory(gritlm)
add_subdirectory(gguf-split)
--- /dev/null
+set(TARGET eval-callback)
+add_executable(${TARGET} eval-callback.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TEST_TARGET test-eval-callback)
+add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42)
+set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
--- /dev/null
+# llama.cpp/examples/eval-callback
+
+A simple example which demonstrates how to use callback during the inference.
+It simply prints to the console all operations and tensor data.
+
+Usage:
+
+```shell
+eval-callback \
+ --hf-repo ggml-org/models \
+ --hf-file phi-2/ggml-model-q4_0.gguf \
+ --model phi-2-q4_0.gguf \
+ --prompt hello \
+ --seed 42 \
+ -ngl 33
+```
+
+Will print:
+
+```shell
+llm_load_tensors: offloaded 33/33 layers to GPU
+...
+llama_new_context_with_model: n_ctx = 512
+...
+llama_new_context_with_model: CUDA0 compute buffer size = 105.00 MiB
+llama_new_context_with_model: CUDA_Host compute buffer size = 6.01 MiB
+llama_new_context_with_model: graph nodes = 1225
+llama_new_context_with_model: graph splits = 2
+ggml_debug: inp_embd = (f32) GET_ROWS(token_embd.weight{2560, 51200, 1, 1}, inp_tokens{1, 1, 1, 1}}) = {2560, 1, 1, 1}
+ [
+ [
+ [ -0.0181, 0.0272, 0.0272, ...],
+ ],
+ ]
+ggml_debug: norm-0 = (f32) NORM(CUDA0#inp_embd#0{2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+ [
+ [
+ [ -0.6989, 1.0636, 1.0636, ...],
+ ],
+ ]
+ggml_debug: norm_w-0 = (f32) MUL(norm-0{2560, 1, 1, 1}, blk.0.attn_norm.weight{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+ [
+ [
+ [ -0.1800, 0.2817, 0.2632, ...],
+ ],
+ ]
+ggml_debug: attn_norm-0 = (f32) ADD(norm_w-0{2560, 1, 1, 1}, blk.0.attn_norm.bias{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+ [
+ [
+ [ -0.1863, 0.2970, 0.2604, ...],
+ ],
+ ]
+ggml_debug: wqkv-0 = (f32) MUL_MAT(blk.0.attn_qkv.weight{2560, 7680, 1, 1}, attn_norm-0{2560, 1, 1, 1}}) = {7680, 1, 1, 1}
+ [
+ [
+ [ -1.1238, 1.2876, -1.8086, ...],
+ ],
+ ]
+ggml_debug: bqkv-0 = (f32) ADD(wqkv-0{7680, 1, 1, 1}, blk.0.attn_qkv.bias{7680, 1, 1, 1}}) = {7680, 1, 1, 1}
+ [
+ [
+ [ -1.1135, 1.4604, -1.9226, ...],
+ ],
+ ]
+ggml_debug: bqkv-0 (view) = (f32) VIEW(bqkv-0{7680, 1, 1, 1}, }) = {2560, 1, 1, 1}
+ [
+ [
+ [ -1.1135, 1.4604, -1.9226, ...],
+ ],
+ ]
+ggml_debug: Qcur-0 = (f32) CONT(bqkv-0 (view){2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+ [
+ [
+ [ -1.1135, 1.4604, -1.9226, ...],
+ ],
+ ]
+ggml_debug: Qcur-0 (reshaped) = (f32) RESHAPE(Qcur-0{2560, 1, 1, 1}, }) = {80, 32, 1, 1}
+ [
+ [
+ [ -1.1135, 1.4604, -1.9226, ...],
+ [ -0.3608, 0.5076, -1.8866, ...],
+ [ 1.7643, 0.0273, -2.1065, ...],
+ ...
+ ],
+ ]
+ggml_debug: Qcur-0 = (f32) ROPE(Qcur-0 (reshaped){80, 32, 1, 1}, CUDA0#inp_pos#0{1, 1, 1, 1}}) = {80, 32, 1, 1}
+ [
+ [
+ [ -1.1135, 1.4604, -1.9226, ...],
+ [ -0.3608, 0.5076, -1.8866, ...],
+ [ 1.7643, 0.0273, -2.1065, ...],
+ ...
+ ],
+ ]
+```
--- /dev/null
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <cstdio>
+#include <random>
+#include <string>
+#include <tuple>
+#include <vector>
+
+/**
+ * This the arbitrary data which will be passed to each callback.
+ * Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor.
+ */
+struct callback_data {
+ std::vector<uint8_t> data;
+};
+
+static std::string ggml_ne_string(const ggml_tensor * t) {
+ std::string str;
+ for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+ str += std::to_string(t->ne[i]);
+ if (i + 1 < GGML_MAX_DIMS) {
+ str += ", ";
+ }
+ }
+ return str;
+}
+
+static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+ float sum = 0;
+ for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+ printf(" [\n");
+ for (int64_t i2 = 0; i2 < ne[2] && i2 < n; i2++) {
+ printf(" [\n");
+ for (int64_t i1 = 0; i1 < ne[1] && i1 < n; i1++) {
+ printf(" [");
+ for (int64_t i0 = 0; i0 < ne[0] && i0 < n; i0++) {
+ size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+ float v;
+ if (type == GGML_TYPE_F16) {
+ v = ggml_fp16_to_fp32(*(ggml_fp16_t *) data + i);
+ } else if (type == GGML_TYPE_F32) {
+ v = *(float *) data + i;
+ } else if (type == GGML_TYPE_I32) {
+ v = (float) *(int32_t *) data + i;
+ } else if (type == GGML_TYPE_I16) {
+ v = (float) *(int16_t *) data + i;
+ } else if (type == GGML_TYPE_I8) {
+ v = (float) *(int8_t *) data + i;
+ } else {
+ GGML_ASSERT(false);
+ }
+ printf("%8.4f", v);
+ sum += v;
+ if (i0 < ne[0] - 1 && i0 < n - 1) printf(", ");
+ }
+ if (ne[0] > n) printf(", ...");
+ printf("],\n");
+ }
+ if (ne[1] > n) printf(" ...\n");
+ printf(" ],\n");
+ }
+ if (ne[2] > n) printf(" ...\n");
+ printf(" ]\n");
+ printf(" sum = %f\n", sum);
+ }
+}
+
+/**
+ * GGML operations callback during the graph execution.
+ *
+ * @param t current tensor
+ * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
+ * if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
+ * see ggml_backend_sched_eval_callback
+ * @param user_data user data to pass at each call back
+ * @return true to receive data or continue the graph, false otherwise
+ */
+static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
+ auto * cb_data = (callback_data *) user_data;
+
+ const struct ggml_tensor * src0 = t->src[0];
+ const struct ggml_tensor * src1 = t->src[1];
+
+ if (ask) {
+ return true; // Always retrieve data
+ }
+
+ char src1_str[128] = {0};
+ if (src1) {
+ sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
+ }
+
+ printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
+ t->name, ggml_type_name(t->type), ggml_op_name(t->op),
+ src0->name, ggml_ne_string(src0).c_str(),
+ src1 ? src1_str : "",
+ ggml_ne_string(t).c_str());
+
+
+ // copy the data from the GPU memory if needed
+ const bool is_host = ggml_backend_buffer_is_host(t->buffer);
+
+ if (!is_host) {
+ auto n_bytes = ggml_nbytes(t);
+ cb_data->data.resize(n_bytes);
+ ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+ }
+
+ if (!ggml_is_quantized(t->type)) {
+ uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
+ ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
+ }
+
+ return true;
+}
+
+static bool run(llama_context * ctx, const gpt_params & params) {
+ const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+
+ std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+ if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+ fprintf(stderr, "%s : failed to eval\n", __func__);
+ return false;
+ }
+
+ return true;
+}
+
+int main(int argc, char ** argv) {
+
+ callback_data cb_data;
+
+ gpt_params params;
+ if (!gpt_params_parse(argc, argv, params)) {
+ return 1;
+ }
+
+ print_build_info();
+
+ std::mt19937 rng(params.seed);
+ if (params.random_prompt) {
+ params.prompt = gpt_random_prompt(rng);
+ }
+
+ llama_backend_init();
+ llama_numa_init(params.numa);
+
+ // pass the callback to the backend scheduler
+ // it will be executed for each node during the graph computation
+ params.cb_eval = ggml_debug;
+ params.cb_eval_user_data = &cb_data;
+ params.warmup = false;
+
+ // init
+ llama_model * model;
+ llama_context * ctx;
+ 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;
+ }
+
+ // print system information
+ {
+ fprintf(stderr, "\n");
+ fprintf(stderr, "%s\n", get_system_info(params).c_str());
+ }
+
+ bool OK = run(ctx, params);
+ if (!OK) {
+ return 1;
+ }
+
+ llama_print_timings(ctx);
+
+ llama_free(ctx);
+ llama_free_model(model);
+
+ llama_backend_free();
+
+ return 0;
+}
llama_backend_init();
llama_numa_init(params.numa);
- llama_model_params mparams = llama_model_params_from_gpt_params(params);
-
- llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
- if (model == NULL) {
- fprintf(stderr, "%s: error: unable to load model\n", __func__);
- return 1;
- }
-
- llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
- cparams.cb_eval = ik_collect_imatrix;
- cparams.cb_eval_user_data = NULL;
-
- llama_context * ctx = llama_new_context_with_model(model, cparams);
- if (ctx == NULL) {
- fprintf(stderr, "%s: error: unable to create context\n", __func__);
+ params.cb_eval = ik_collect_imatrix;
+ params.cb_eval_user_data = NULL;
+ params.warmup = false;
+
+ // init
+ llama_model * model;
+ llama_context * ctx;
+ 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;
}
add_new_bigram(bigram.left, left_symbol.next); // right side of current symbol
}
- // add the fnished tokens to the final list keeping correct order for next and prev
+ // add the finished tokens to the final list keeping correct order for next and prev
for (auto & sym : symbols) {
if (sym.n > 0) {
sym.prev = final_prev_index;
overview / pkg / ggml / sources / llama.cpp / commitdiff