Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 17:25:07 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/docker.yml
#	.github/workflows/editorconfig.yml
#	examples/run/run.cpp
#	examples/server/README.md
#	scripts/sync-ggml.last
This commit is contained in:
Concedo
2025-01-09 16:50:29 +08:00
parent 5cce8a5fbc 8d59d91171
commit 1b49dc305f
18 changed files with 888 additions and 107 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-adapter.cpp
+18 -6
View File
@@ -242,6 +242,10 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
} else {
ab_map[name].b = cur;
}
} else if (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");
}
@@ -251,6 +255,7 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
for (auto & it : ab_map) {
const std::string & name = it.first;
llama_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");
@@ -259,16 +264,23 @@ static void llama_lora_adapter_init_impl(struct llama_model & model, const char
// device buft and device ctx
auto * model_tensor = llama_model_get_tensor(model, name.c_str());
if (!model_tensor) {
throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model");
throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
}
struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
// validate tensor shape
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");
}
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)");
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)");
}
}
// save tensor to adapter
src/llama-adapter.h
+7
View File
@@ -45,6 +45,13 @@ struct llama_lora_weight {
struct ggml_tensor * a = nullptr;
struct ggml_tensor * b = nullptr;
// get actual scale based on rank and alpha
float get_scale(float alpha, float adapter_scale) {
const float rank = (float) b->ne[0];
const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
return scale;
}
llama_lora_weight() = default;
llama_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {}
};
src/llama-quant.cpp
+13 -14
View File
@@ -7,14 +7,12 @@
#include <algorithm>
#include <cmath>
#include <cstring>
#include <cinttypes>
#include <fstream>
#include <mutex>
#include <thread>
#include <unordered_map>
// TODO: replace with ggml API call
#define QK_K 256
static void zeros(std::ofstream & file, size_t n) {
char zero = 0;
for (size_t i = 0; i < n; ++i) {
@@ -154,8 +152,10 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
if (qs.params->output_tensor_type < GGML_TYPE_COUNT) {
new_type = qs.params->output_tensor_type;
} else {
int nx = tensor->ne[0];
if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
const int64_t nx = tensor->ne[0];
const int64_t qk_k = ggml_blck_size(new_type);
if (arch == LLM_ARCH_FALCON || nx % qk_k != 0) {
new_type = GGML_TYPE_Q8_0;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS ||
@@ -367,20 +367,19 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
// if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
//}
bool convert_incompatible_tensor = false;
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K || new_type == GGML_TYPE_IQ4_XS ||
new_type == GGML_TYPE_IQ2_XS || new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_S ||
new_type == GGML_TYPE_IQ3_XXS || new_type == GGML_TYPE_IQ1_S || new_type == GGML_TYPE_IQ3_S ||
new_type == GGML_TYPE_IQ1_M) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for %s", __func__, nx, ny, QK_K, ggml_type_name(new_type));
{
const int64_t nx = tensor->ne[0];
const int64_t ny = tensor->ne[1];
const int64_t qk_k = ggml_blck_size(new_type);
if (nx % qk_k != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %" PRId64 " x %" PRId64 " are not divisible by %" PRId64 ", required for %s", __func__, nx, ny, qk_k, ggml_type_name(new_type));
convert_incompatible_tensor = true;
} else {
++qs.n_k_quantized;
}
}
if (convert_incompatible_tensor) {
switch (new_type) {
case GGML_TYPE_TQ1_0:
src/llama.cpp
+18 -3
View File
@@ -2652,6 +2652,21 @@ static struct ggml_tensor * llm_build_inp_embd(
ggml_set_input(lctx.inp_tokens);
inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
// apply lora for embedding tokens if needed
for (auto & it : lctx.lora_adapters) {
struct llama_lora_weight * lora = it.first->get_weight(tok_embd);
if (lora == nullptr) {
continue;
}
const float adapter_scale = it.second;
const float scale = lora->get_scale(it.first->alpha, adapter_scale);
struct ggml_tensor * inpL_delta = ggml_scale(ctx, ggml_mul_mat(
ctx, lora->b, // non-transposed lora_b
ggml_get_rows(ctx, lora->a, lctx.inp_tokens)
), scale);
inpL = ggml_add(ctx, inpL, inpL_delta);
}
} else {
lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
inpL = lctx.inp_embd;
@@ -2724,9 +2739,8 @@ static struct ggml_tensor * llm_build_lora_mm(
if (lora == nullptr) {
continue;
}
const float alpha = it.first->alpha;
const float rank = (float) lora->b->ne[0];
const float scale = alpha ? it.second * alpha / rank : it.second;
const float adapter_scale = it.second;
const float scale = lora->get_scale(it.first->alpha, adapter_scale);
struct ggml_tensor * ab_cur = ggml_mul_mat(
ctx0, lora->b,
ggml_mul_mat(ctx0, lora->a, cur)
@@ -4092,6 +4106,7 @@ struct llm_build_context {
// feed-forward network
if (model.layers[il].ffn_gate_inp == nullptr) {
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, cb, il);