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
2024-09-08 11:44:55 +08:00
committed by GitHub
parent 8a7baa1bbf 06d4fba851
commit 06db1f2cf5
31 changed files with 689 additions and 25128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
llm/ext_server/CMakeLists.txt vendored
View File

@@ -2,7 +2,7 @@ set(TARGET ollama_llama_server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
set(LLAMA_SERVER_LDFLAGS $ENV{LLAMA_SERVER_LDFLAGS})
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
add_executable(${TARGET} server.cpp utils.hpp httplib.h)
install(TARGETS ${TARGET} RUNTIME)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>

24596
llm/ext_server/json.hpp vendored
View File

File diff suppressed because it is too large Load Diff

25
llm/ext_server/server.cpp vendored
View File

@@ -262,7 +262,7 @@ struct server_slot {
char buffer[512];
double t_token = t_prompt_processing / n_prompt_tokens_processed;
double n_tokens_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed;
sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
snprintf(buffer, sizeof(buffer), "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
t_prompt_processing, n_prompt_tokens_processed,
t_token, n_tokens_second);
LOG_DEBUG(buffer, {
@@ -276,7 +276,7 @@ struct server_slot {
t_token = t_token_generation / n_decoded;
n_tokens_second = 1e3 / t_token_generation * n_decoded;
sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
snprintf(buffer, sizeof(buffer), "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)",
t_token_generation, n_decoded,
t_token, n_tokens_second);
LOG_DEBUG(buffer, {
@@ -288,7 +288,7 @@ struct server_slot {
{"n_tokens_second", n_tokens_second},
});
sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation);
snprintf(buffer, sizeof(buffer), " total time = %10.2f ms", t_prompt_processing + t_token_generation);
LOG_DEBUG(buffer, {
{"slot_id", id},
{"task_id", task_id},
@@ -425,7 +425,7 @@ struct llama_server_context
n_ctx = llama_n_ctx(ctx);
add_bos_token = llama_should_add_bos_token(model);
add_bos_token = llama_add_bos_token(model);
return true;
}
@@ -1031,7 +1031,7 @@ struct llama_server_context
continue;
}
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
LOG_TEE("Error processing the given image");
return false;
}
@@ -2014,7 +2014,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf("options:\n");
printf(" -h, --help show this help message and exit\n");
printf(" -v, --verbose verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.cpuparams.n_threads);
printf(" -tb N, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads)\n");
printf(" --threads-http N number of threads in the http server pool to process requests (default: max(hardware concurrency - 1, --parallel N + 2))\n");
printf(" -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx);
@@ -2287,7 +2287,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
invalid_param = true;
break;
}
params.n_threads = std::stoi(argv[i]);
params.cpuparams.n_threads = std::stoi(argv[i]);
}
else if (arg == "--grp-attn-n" || arg == "-gan")
{
@@ -2315,7 +2315,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
invalid_param = true;
break;
}
params.n_threads_batch = std::stoi(argv[i]);
params.cpuparams_batch.n_threads = std::stoi(argv[i]);
}
else if (arg == "--threads-http")
{
@@ -2626,6 +2626,11 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, g
params.kv_overrides.back().key[0] = 0;
}
postprocess_cpu_params(params.cpuparams, nullptr);
postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
postprocess_cpu_params(params.draft_cpuparams, &params.cpuparams);
postprocess_cpu_params(params.draft_cpuparams_batch, &params.cpuparams_batch);
if (invalid_param)
{
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
@@ -2775,8 +2780,8 @@ int main(int argc, char **argv) {
{"commit", LLAMA_COMMIT}});
LOG_INFO("system info", {
{"n_threads", params.n_threads},
{"n_threads_batch", params.n_threads_batch},
{"n_threads", params.cpuparams.n_threads},
{"n_threads_batch", params.cpuparams_batch.n_threads},
{"total_threads", std::thread::hardware_concurrency()},
{"system_info", llama_print_system_info()},
});

2
llm/generate/gen_darwin.sh
View File

@@ -19,7 +19,7 @@ sign() {
fi
}
COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DGGML_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
case "${GOARCH}" in
"amd64")

6
llm/ggml.go
View File

@@ -360,11 +360,13 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
switch llm.KV().Architecture() {
case "llama":
fullOffload = 4 * batch * (1 + 4*embedding + context*(1+heads))
fullOffload = max(
4*batch*(1+4*embedding+context*(1+heads)),
4*batch*(embedding+vocab),
)
partialOffload = 4 * batch * embedding
partialOffload += max(
// 4*batch*(4+6*embedding+context*(2*heads)+llm.KV().GQA()),
4*batch*(1+embedding+max(context, embedding))+embedding*embedding*9/16+4*context*(batch*heads+embeddingHeads*headsKV),
4*batch*(embedding+vocab)+embedding*vocab*105/128,
)

2
llm/llama.cpp

Submodule llm/llama.cpp updated: 1e6f6554aa...8962422b1c

26
llm/memory.go
View File

@@ -7,6 +7,7 @@ import (
"strings"
"github.com/ollama/ollama/api"
"github.com/ollama/ollama/envconfig"
"github.com/ollama/ollama/format"
"github.com/ollama/ollama/gpu"
)
@@ -94,6 +95,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
// Overflow that didn't fit into the GPU
var overflow uint64
overhead := envconfig.GpuOverhead()
availableList := make([]string, len(gpus))
for i, gpu := range gpus {
availableList[i] = format.HumanBytes2(gpu.FreeMemory)
@@ -164,8 +166,22 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
gzo = gpuZeroOverhead
}
// Only include GPUs that can fit the graph, gpu minimum, the layer buffer and at least more layer
if gpus[i].FreeMemory < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
slog.Debug("gpu has too little memory to allocate any layers", "gpu", gpus[i])
if (gpus[i].FreeMemory - overhead) < gzo+max(graphPartialOffload, graphFullOffload)+gpus[i].MinimumMemory+2*layerSize {
slog.Debug("gpu has too little memory to allocate any layers",
"id", gpus[i].ID,
"library", gpus[i].Library,
"variant", gpus[i].Variant,
"compute", gpus[i].Compute,
"driver", fmt.Sprintf("%d.%d", gpus[i].DriverMajor, gpus[i].DriverMinor),
"name", gpus[i].Name,
"total", format.HumanBytes2(gpus[i].TotalMemory),
"available", format.HumanBytes2(gpus[i].FreeMemory),
"minimum_memory", gpus[i].MinimumMemory,
"layer_size", format.HumanBytes2(layerSize),
"gpu_zer_overhead", format.HumanBytes2(gzo),
"partial_offload", format.HumanBytes2(graphPartialOffload),
"full_offload", format.HumanBytes2(graphFullOffload),
)
continue
}
gpusWithSpace = append(gpusWithSpace, gs{i, &gpus[i]})
@@ -196,7 +212,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
for j := len(gpusWithSpace); j > 0; j-- {
g := gpusWithSpace[i%j]
used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
if g.g.FreeMemory > used+layerSize {
if (g.g.FreeMemory - overhead) > used+layerSize {
gpuAllocations[g.i] += layerSize
layerCounts[g.i]++
layerCount++
@@ -219,7 +235,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
for j := len(gpusWithSpace); j > 0; j-- {
g := gpusWithSpace[layerCount%j]
used := gpuAllocations[g.i] + max(graphPartialOffload, graphFullOffload)
if g.g.FreeMemory > used+memoryLayerOutput {
if (g.g.FreeMemory - overhead) > used+memoryLayerOutput {
gpuAllocations[g.i] += memoryLayerOutput
layerCounts[g.i]++
layerCount++
@@ -306,6 +322,7 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
}
func (m MemoryEstimate) log() {
overhead := envconfig.GpuOverhead()
slog.Info(
"offload to "+m.inferenceLibrary,
slog.Group(
@@ -323,6 +340,7 @@ func (m MemoryEstimate) log() {
"memory",
// memory available by GPU for offloading
"available", m.availableList,
"gpu_overhead", format.HumanBytes2(overhead),
slog.Group(
"required",
// memory required for full offloading

10
llm/patches/05-default-pretokenizer.diff
View File

@@ -1,8 +1,8 @@
diff --git a/src/llama.cpp b/src/llama.cpp
index a207451f..2ddf431d 100644
index 88355971..dd7d41ed 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -5347,16 +5347,7 @@ static void llm_load_vocab(
@@ -6083,16 +6083,7 @@ static void llm_load_vocab(
if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
vocab.tokenizer_add_space_prefix = false;
vocab.tokenizer_clean_spaces = true;
@@ -20,9 +20,9 @@ index a207451f..2ddf431d 100644
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
} else if (
tokenizer_pre == "llama3" ||
@@ -5443,7 +5434,8 @@ static void llm_load_vocab(
tokenizer_pre == "codeshell") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CODESHELL;
@@ -6188,7 +6179,8 @@ static void llm_load_vocab(
tokenizer_pre == "exaone") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_EXAONE;
} 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__);

28
llm/patches/06-embeddings.diff
View File

@@ -1,37 +1,36 @@
diff --git a/src/llama.cpp b/src/llama.cpp
index 1fe2b9f7..a43312a7 100644
index 88355971..d7db689b 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
@@ -15906,7 +15906,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
const auto n_embd = hparams.n_embd;
// TODO: use a per-batch flag for logits presence instead
- const bool has_logits = !cparams.embeddings;
+ const bool has_logits = cparams.causal_attn;
const bool has_embd = lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
@@ -13959,17 +13959,25 @@ static int llama_decode_internal(
@@ -16175,20 +16175,23 @@ static int llama_decode_internal(
// no output
res = nullptr;
embd = nullptr;
- } else if (cparams.embeddings) {
- res = nullptr; // do not extract logits for embedding case
- embd = gf->nodes[gf->n_nodes - 1];
- if (strcmp(embd->name, "result_embd_pooled") != 0) {
- embd = gf->nodes[gf->n_nodes - 2];
- res = nullptr; // do not extract logits for embedding case
- embd = nullptr;
+ }
+
+ if (cparams.embeddings) {
+ for (int i = gf->n_nodes - 1; i >= 0; --i) {
for (int i = gf->n_nodes - 1; i >= 0; --i) {
- if (strcmp(gf->nodes[i]->name, "result_embd_pooled") == 0) {
- embd = gf->nodes[i];
+ embd = gf->nodes[i];
+ if (strcmp(embd->name, "result_embd_pooled") == 0) {
+ break;
+ }
break;
}
}
GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
- } else {
+ } else {
- GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
} else {
embd = nullptr; // do not extract embeddings when not needed
GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
}
@@ -39,7 +38,6 @@ index 1fe2b9f7..a43312a7 100644
+ if (!cparams.causal_attn) {
+ res = nullptr; // do not extract logits when not needed
+ }
+
// 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);
ggml_backend_sched_alloc_graph(lctx.sched, gf);

350
llm/patches/09-lora.diff
View File

@@ -1,350 +0,0 @@
diff --git a/common/common.cpp b/common/common.cpp
index 2e8374d5..70d0afde 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2110,9 +2110,21 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
if (loaded_la.adapter == nullptr) {
fprintf(stderr, "%s: error: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
- llama_free(lctx);
- llama_free_model(model);
- return iparams;
+
+ // if that fails, try loading as ggla for compatibility
+ int err = llama_model_apply_lora_from_file(model,
+ la.path.c_str(),
+ la.scale,
+ nullptr,
+ params.n_threads);
+ if (err != 0) {
+ fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+ llama_free(lctx);
+ llama_free_model(model);
+ return iparams;
+ } else {
+ break;
+ }
}
iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
}
diff --git a/include/llama.h b/include/llama.h
index 93fd77ca..b0fb37a6 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -1160,6 +1160,20 @@ extern "C" {
LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+ // Apply a LoRA adapter to a loaded model
+ // path_base_model is the path to a higher quality model to use as a base for
+ // the layers modified by the adapter. Can be NULL to use the current loaded model.
+ // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+ // will be applied on top of the previous one
+ // Returns 0 on success
+ LLAMA_API int32_t llama_model_apply_lora_from_file(
+ const struct llama_model * model,
+ const char * path_lora,
+ float scale,
+ const char * path_base_model,
+ int32_t n_threads);
+
+
#ifdef __cplusplus
}
#endif
diff --git a/src/llama.cpp b/src/llama.cpp
index 80a0dd0f..9d7b0e17 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
fputs(text, stderr);
fflush(stderr);
}
+
+static int llama_apply_lora_from_file_internal(
+ const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
+) {
+ LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
+
+ const int64_t t_start_lora_us = ggml_time_us();
+
+ llama_file fin(path_lora, "rb");
+
+ // verify magic and version
+ {
+ uint32_t magic = fin.read_u32();
+ if (magic != LLAMA_FILE_MAGIC_GGLA) {
+ LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
+ return 1;
+ }
+
+ uint32_t format_version = fin.read_u32();
+ if (format_version != 1) {
+ LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
+ return 1;
+ }
+ }
+
+ int32_t lora_r = fin.read_u32();
+ int32_t lora_alpha = fin.read_u32();
+ float scaling = scale * (float)lora_alpha / (float)lora_r;
+
+ LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
+
+ // load base model
+ std::unique_ptr<llama_model_loader> ml;
+ if (path_base_model) {
+ LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
+ ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
+ ml->init_mappings(/*prefetch*/ false); // no prefetching
+ }
+
+ struct tensor_meta {
+ std::string name;
+ ggml_type type;
+ int32_t ne[2];
+ size_t offset;
+ };
+ std::map<std::string, tensor_meta> tensor_meta_map;
+
+ // load all tensor meta
+ while (true) {
+ if (fin.tell() == fin.size) {
+ // eof
+ break;
+ }
+
+ int32_t n_dims;
+ int32_t name_len;
+ int32_t ftype;
+
+ fin.read_raw(&n_dims, sizeof(n_dims));
+ fin.read_raw(&name_len, sizeof(name_len));
+ fin.read_raw(&ftype, sizeof(ftype));
+
+ if (n_dims != 1 && n_dims != 2) {
+ LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
+ return 1;
+ }
+
+ int32_t ne[2] = { 1, 1 };
+ for (int i = 0; i < n_dims; ++i) {
+ fin.read_raw(&ne[i], sizeof(ne[i]));
+ }
+
+ std::string name;
+ {
+ GGML_ASSERT(name_len < GGML_MAX_NAME);
+ char buf[GGML_MAX_NAME];
+ fin.read_raw(buf, name_len);
+ name = std::string(buf, name_len);
+ }
+
+ // check for lora suffix
+ std::string lora_suffix;
+ if (name.length() > 6) {
+ lora_suffix = name.substr(name.length() - 6);
+ }
+ if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
+ LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
+ return 1;
+ }
+
+ // tensor type
+ ggml_type wtype;
+ switch (ftype) {
+ case 0: wtype = GGML_TYPE_F32; break;
+ case 1: wtype = GGML_TYPE_F16; break;
+ default:
+ {
+ LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
+ __func__, ftype);
+ return 1;
+ }
+ }
+
+ // data offset
+ size_t offset = fin.tell();
+ offset = (offset + 31) & -32;
+
+ // skip tensor data
+ fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
+
+ tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
+ }
+
+ bool warned = false;
+ int n_tensors = 0;
+
+ // apply
+ ggml_backend_t backend_cpu = ggml_backend_cpu_init();
+ if (backend_cpu == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
+ return 1;
+ }
+ ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
+
+ std::vector<no_init<uint8_t>> read_buf;
+ for (const auto & it : model.tensors_by_name) {
+ const std::string & base_name = it.first;
+ ggml_tensor * model_t = it.second;
+
+ if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
+ tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
+ continue;
+ }
+
+ tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
+ tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
+
+ ggml_init_params lora_init_params = {
+ /* .mem_size */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
+ /* .mem_buffer */ nullptr,
+ /* .no_alloc */ true,
+ };
+ ggml_context * lora_ctx = ggml_init(lora_init_params);
+ if (lora_ctx == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ // create tensors
+ ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
+ ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
+ ggml_set_name(loraA, metaA.name.c_str());
+ ggml_set_name(loraB, metaB.name.c_str());
+
+ ggml_tensor * base_t;
+ if (ml) {
+ if (!ml->get_tensor_meta(base_name.c_str())) {
+ LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
+ return 1;
+ }
+ base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
+ } else {
+ base_t = ggml_dup_tensor(lora_ctx, model_t);
+ }
+ ggml_set_name(base_t, base_name.c_str());
+
+ // allocate in backend buffer
+ ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
+ if (lora_buf == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
+ return 1;
+ }
+
+ // load tensor data
+ auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
+ read_buf.resize(ggml_nbytes(tensor));
+ fin.seek(tensor_meta.offset, SEEK_SET);
+ fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
+ ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
+ };
+ load_tensor(metaA, loraA);
+ load_tensor(metaB, loraB);
+
+ // load base model tensor data
+ if (ml) {
+ ml->load_data_for(base_t);
+ } else {
+ ggml_backend_tensor_copy(model_t, base_t);
+ }
+
+ if (ggml_is_quantized(base_t->type) && !warned) {
+ LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
+ "use a f16 or f32 base model with --lora-base\n", __func__);
+ warned = true;
+ }
+
+ if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
+ LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
+ " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
+ ggml_free(lora_ctx);
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ auto build_lora_graph = [&]() {
+ // w = w + BA*s
+ ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
+ ggml_set_name(BA, "BA");
+
+ if (scaling != 1.0f) {
+ BA = ggml_scale(lora_ctx, BA, scaling);
+ ggml_set_name(BA, "BA_scaled");
+ }
+
+ ggml_tensor * r;
+ r = ggml_add_inplace(lora_ctx, base_t, BA);
+ ggml_set_name(r, "r_add");
+
+ if (base_t->type != model_t->type) {
+ // convert the result to the model type
+ r = ggml_cast(lora_ctx, r, model_t->type);
+ ggml_set_name(r, "r_cast");
+ }
+
+ return r;
+ };
+
+ ggml_cgraph * gf = ggml_new_graph(lora_ctx);
+ ggml_tensor * r = build_lora_graph();
+ ggml_build_forward_expand(gf, r);
+
+ ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
+ if (graph_buf == nullptr) {
+ LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
+ ggml_free(lora_ctx);
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_free(backend_cpu);
+ return 1;
+ }
+
+ ggml_backend_graph_compute(backend_cpu, gf);
+
+ ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
+
+#if 0
+ // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
+ //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
+
+ // sched compute
+ ggml_build_forward_expand(gf, build_graph());
+ ggml_backend_sched_init_measure(sched, gf);
+
+ // create the graph again, since the previous one was destroyed by the measure
+ ggml_graph_clear(gf);
+ ggml_build_forward_expand(gf, build_graph());
+ ggml_backend_sched_graph_compute(sched, gf);
+ ggml_backend_sched_free(sched);
+#endif
+
+ ggml_backend_buffer_free(lora_buf);
+ ggml_backend_buffer_free(graph_buf);
+ ggml_free(lora_ctx);
+
+ n_tensors++;
+ if (n_tensors % 4 == 0) {
+ LLAMA_LOG_INFO(".");
+ }
+ }
+
+ ggml_backend_free(backend_cpu);
+
+ const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
+ LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
+
+ return 0;
+}
+
+int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
+ try {
+ return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
+ } catch (const std::exception & err) {
+ LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
+ return 1;
+ }
+}
\ No newline at end of file

43
llm/patches/11-phi3-sliding-window.diff
View File

@@ -1,43 +0,0 @@
From 6eedae4cf2fcc8015dac79cb3f28f61fcabacab2 Mon Sep 17 00:00:00 2001
From: Michael Yang <mxyng@pm.me>
Date: Wed, 31 Jul 2024 14:57:04 -0700
Subject: [PATCH] phi3 sliding window
---
src/llama.cpp | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/src/llama.cpp b/src/llama.cpp
index a207451f..f2872d4e 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -4893,7 +4893,7 @@ static void llm_load_hparams(
} break;
case LLM_ARCH_PHI3:
{
- ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+ ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
@@ -10762,7 +10762,7 @@ struct llm_build_context {
struct ggml_tensor * inp_pos = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
- struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
+ struct ggml_tensor * KQ_mask = hparams.n_swa > 0 ? build_inp_KQ_mask_swa() : build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
auto residual = inpL;
@@ -10820,7 +10820,7 @@ struct llm_build_context {
cur = llm_build_kv(ctx0, lctx, kv_self, gf,
model.layers[il].wo, model.layers[il].bo,
- Kcur, Vcur, Qcur, KQ_mask_swa, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+ Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
}
if (il == n_layer - 1) {
--
2.45.2

7
llm/server.go
View File

@@ -98,7 +98,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
systemTotalMemory = systemMemInfo.TotalMemory
systemFreeMemory = systemMemInfo.FreeMemory
systemSwapFreeMemory = systemMemInfo.FreeSwap
slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
slog.Info("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
}
// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
@@ -584,8 +584,7 @@ func (s *llmServer) Ping(ctx context.Context) error {
func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
start := time.Now()
stallDuration := 5 * time.Minute // If no progress happens
finalLoadDuration := 5 * time.Minute // After we hit 100%, give the runner more time to come online
stallDuration := envconfig.LoadTimeout() // If no progress happens
stallTimer := time.Now().Add(stallDuration) // give up if we stall
slog.Info("waiting for llama runner to start responding")
@@ -637,7 +636,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
stallTimer = time.Now().Add(stallDuration)
} else if !fullyLoaded && int(s.loadProgress*100.0) >= 100 {
slog.Debug("model load completed, waiting for server to become available", "status", status.ToString())
stallTimer = time.Now().Add(finalLoadDuration)
stallTimer = time.Now().Add(stallDuration)
fullyLoaded = true
}
time.Sleep(time.Millisecond * 250)