ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) { return nullptr;
}
return tensors[il];
}
ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const {
ggml_tensor * layer_dir = tensor_for(il); if (layer_dir != nullptr) {
cur = ggml_add(ctx, cur, layer_dir);
}
// make tensors
tensors.reserve(hparams.n_layer);
tensors.push_back(nullptr); // there's never a tensor for layer 0 for (size_t il = 1; il < hparams.n_layer; il++) {
ggml_backend_buffer_type_t buft = model.select_buft(il);
ggml_context * ctx = ctx_for_buft(buft); if (!ctx) {
LLAMA_LOG_ERROR("%s: failed to allocate context for control vector\n", __func__); returnfalse;
}
ggml_tensor * tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hparams.n_embd);
tensors.push_back(tensor);
}
// allocate tensors / buffers and zero
bufs.reserve(ctx_map.size()); for (auto it : ctx_map) {
ggml_backend_buffer_type_t buft = it.first;
ggml_context * ctx = it.second;
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (!buf) {
LLAMA_LOG_ERROR("%s: failed to allocate buffer for control vector\n", __func__); returnfalse;
}
ggml_backend_buffer_clear(buf, 0);
bufs.emplace_back(buf);
}
if (data == nullptr) { // disable the current control vector (but leave allocated for later)
layer_start = -1;
layer_end = -1; returntrue;
}
if (n_embd != (int) hparams.n_embd) {
LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__); returnfalse;
}
if (tensors.empty()) { if (!init(model)) { returnfalse;
}
}
layer_start = il_start;
layer_end = il_end;
for (size_t il = 1; il < hparams.n_layer; il++) {
assert(tensors[il] != nullptr);
const size_t off = n_embd * (il - 1); // buffer doesn't have data for layer 0, since it's never present if (off + n_embd <= len) {
ggml_backend_tensor_set(tensors[il], data + off, 0, n_embd * ggml_element_size(tensors[il]));
}
}
gguf_context_ptr ctx_gguf { gguf_init_from_file(path_lora, meta_gguf_params) }; if (!ctx_gguf) { throw std::runtime_error("failed to load lora adapter file from " + std::string(path_lora));
}
ggml_context_ptr ctx { ctx_init };
// check metadata
{ auto get_kv_str = [&](const std::string & key) -> std::string { int id = gguf_find_key(ctx_gguf.get(), key.c_str()); return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf.get(), id));
}; auto get_kv_f32 = [&](const std::string & key) -> float { int id = gguf_find_key(ctx_gguf.get(), key.c_str()); return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf.get(), id);
};
LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
auto general_type = get_kv_str(llm_kv(LLM_KV_GENERAL_TYPE)); if (general_type != "adapter") { throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
}
auto general_arch_str = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE)); auto general_arch = llm_arch_from_string(general_arch_str); if (general_arch != model.arch) { throw std::runtime_error("model arch and LoRA arch mismatch");
}
auto adapter_type = get_kv_str(llm_kv(LLM_KV_ADAPTER_TYPE)); if (adapter_type != "lora") { throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
}
for (ggml_tensor * cur = ggml_get_first_tensor(ctx.get()); cur; cur = ggml_get_next_tensor(ctx.get(), cur)) {
std::string name(cur->name); if (str_endswith(name, ".lora_a")) {
replace_all(name, ".lora_a", ""); if (ab_map.find(name) == ab_map.end()) {
ab_map[name] = llama_adapter_lora_weight(cur, nullptr);
} else {
ab_map[name].a = cur;
}
} elseif (str_endswith(name, ".lora_b")) {
replace_all(name, ".lora_b", ""); if (ab_map.find(name) == ab_map.end()) {
ab_map[name] = llama_adapter_lora_weight(nullptr, cur);
} else {
ab_map[name].b = cur;
}
} elseif (str_endswith(name, "_norm.weight")) { // TODO: add support for norm vector // for now, we don't really care because most adapters still work fine without it continue;
} else { throw std::runtime_error("LoRA tensor '" + name + "' has unexpected suffix");
}
}
// get extra buffer types of the CPU // TODO: a more general solution for non-CPU extra buft should be imlpemented in the future // ref: https://github.com/ggml-org/llama.cpp/pull/12593#pullrequestreview-2718659948
std::vector<ggml_backend_buffer_type_t> buft_extra;
{ auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); if (!cpu_dev) { throw std::runtime_error(format("%s: no CPU backend found", __func__));
} auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
if (ggml_backend_dev_get_extra_bufts_fn) {
ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); while (extra_bufts && *extra_bufts) {
buft_extra.emplace_back(*extra_bufts);
++extra_bufts;
}
}
}
// add tensors for (auto & it : ab_map) { const std::string & name = it.first;
llama_adapter_lora_weight & w = it.second; bool is_token_embd = str_endswith(name, "token_embd.weight");
if (!w.a || !w.b) { throw std::runtime_error("LoRA tensor pair for '" + name + "' is missing one component");
}
// device buft and device ctx constauto * model_tensor = model.get_tensor(name.c_str()); if (!model_tensor) { throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
}
auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer);
// do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case for (auto & ex : buft_extra) { if (ex == buft) {
LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); if (!cpu_dev) { throw std::runtime_error(format("%s: no CPU backend found", __func__));
}
buft = ggml_backend_dev_buffer_type(cpu_dev);
break;
}
}
LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
ggml_context * dev_ctx = ctx_for_buft(buft); // validate tensor shape if (is_token_embd) { // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd() if (model_tensor->ne[0] != w.b->ne[1] || model_tensor->ne[1] != w.a->ne[1]) { throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
}
} else { if (model_tensor->ne[0] != w.a->ne[0] || model_tensor->ne[1] != w.b->ne[1]) { throw std::runtime_error("tensor '" + name + "' has incorrect shape (hint: maybe wrong base model?)");
} if (w.a->ne[1] != w.b->ne[0]) { throw std::runtime_error("lora_a tensor is not transposed (hint: adapter from \"finetune\" example is no longer supported)");
}
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.