vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/nix/scope.nix
#	.github/workflows/nix-ci-aarch64.yml
#	.github/workflows/nix-ci.yml
#	README.md
#	scripts/sync-ggml.last
This commit is contained in:
Concedo 2024-02-24 18:30:51 +08:00
commit 359a14d3c2
10 changed files with 251 additions and 72 deletions
Download patch file Download diff file Expand all files Collapse all files

83
llama.cpp
View file

@ -533,7 +533,6 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
@ -4126,7 +4125,12 @@ static bool llm_load_tensors(
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, false);
// same as tok_embd, duplicated to allow offloading
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
}
for (int i = 0; i < n_layer; ++i) {
@ -4135,14 +4139,23 @@ static bool llm_load_tensors(
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, false);
layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, false);
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, false);
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, false);
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, false);
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, false);
// AWQ ScaleActivation layer
layer.ffn_act = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, false);
@ -6243,7 +6256,7 @@ struct llm_build_context {
attn_norm = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm,
NULL,
model.layers[il].attn_norm_b,
LLM_NORM, cb, il);
cb(attn_norm, "attn_norm", il);
@ -6254,6 +6267,11 @@ struct llm_build_context {
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
if (model.layers[il].bqkv){
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
}
if (hparams.f_clamp_kqv > 0.0f) {
cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
cb(cur, "wqkv_clamped", il);
@ -6270,7 +6288,7 @@ struct llm_build_context {
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
model.layers[il].wo, NULL,
model.layers[il].wo, model.layers[il].bo,
Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
cb(cur, "kqv_out", il);
}
@ -6283,13 +6301,13 @@ struct llm_build_context {
{
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm,
NULL,
model.layers[il].ffn_norm_b,
LLM_NORM, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL,
model.layers[il].ffn_up, model.layers[il].ffn_up_b,
NULL, NULL,
model.layers[il].ffn_down, NULL,
model.layers[il].ffn_down, model.layers[il].ffn_down_b,
model.layers[il].ffn_act,
LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
@ -6306,7 +6324,7 @@ struct llm_build_context {
cur = llm_build_norm(ctx0, cur, hparams,
model.output_norm,
NULL,
model.output_norm_b,
LLM_NORM, cb, -1);
cb(cur, "result_norm", -1);
@ -7506,6 +7524,7 @@ struct llm_build_context {
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb);
cb(inpL, "inp_embd", -1);
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
cb(inpL, "inp_scaled", -1);
@ -7547,6 +7566,7 @@ struct llm_build_context {
n_embd_head_k, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
cb(Qcur, "Qcur_scaled", il);
@ -7561,6 +7581,7 @@ struct llm_build_context {
Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f, cb, il);
cb(cur, "kqv_out", il);
}
struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
cb(sa_out, "sa_out", il);
@ -10802,7 +10823,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
return std::make_pair(i_layer, n_layer);
};
if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
// for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings
// with the quantization of the output tensor
if (name == tn(LLM_TENSOR_OUTPUT, "weight") ||
(LLM_TENSOR_NAMES.at(arch).find(LLM_TENSOR_OUTPUT) == LLM_TENSOR_NAMES.at(arch).end() && name == "token_embd.weight")) {
int nx = tensor->ne[0];
if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
new_type = GGML_TYPE_Q8_0;
@ -13085,6 +13109,37 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<|assistant|>\n";
}
} else if (tmpl.find("bos_token + message['role']") != std::string::npos) {
// mlabonne/AlphaMonarch-7B template (the <s> is included inside history)
for (auto message : chat) {
std::string bos = (message == chat.front()) ? "" : "<s>"; // skip BOS for first message
ss << bos << message->role << "\n" << message->content << "</s>\n";
}
if (add_ass) {
ss << "<s>assistant\n";
}
} else if (tmpl.find("<start_of_turn>") != std::string::npos) {
// google/gemma-7b-it
std::string system_prompt = "";
for (auto message : chat) {
std::string role(message->role);
if (role == "system") {
// there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
system_prompt = trim(message->content);
continue;
}
// in gemma, "assistant" is "model"
role = role == "assistant" ? "model" : message->role;
ss << "<start_of_turn>" << role << "\n";
if (!system_prompt.empty() && role != "model") {
ss << system_prompt << "\n\n";
system_prompt = "";
}
ss << trim(message->content) << "<end_of_turn>\n";
}
if (add_ass) {
ss << "<start_of_turn>model\n";
}
} else {
// template not supported
return -1;