Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

still not really working right

Browse Source
This commit is contained in:
LostRuins Concedo
2025-11-09 01:57:48 +08:00
parent e6ca0aa8d0 eeee367de5
commit d6a2ad8455
48 changed files with 3193 additions and 2705 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+9 -3
View File
@@ -21,6 +21,8 @@ llama_context::llama_context(
llama_context_params params) :
model(model),
balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
// TODO warning when creating llama_context with awkward ctx size that is not a power of 2,
// may need to be backend-dependent
LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
t_start_us = model.t_start_us;
@@ -112,10 +114,14 @@ llama_context::llama_context(
}
}
// ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732
cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256);
if (cparams.kv_unified) {
cparams.n_ctx_seq = cparams.n_ctx;
} else {
cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max;
cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256);
if (cparams.n_ctx_seq == 0) {
throw std::runtime_error("n_ctx_seq == 0");
@@ -823,7 +829,7 @@ int llama_context::encode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_inp();
const int64_t n_vocab = model.vocab.n_tokens();
// note: during encode, we always pass the full sequence starting from pos = 0
@@ -992,7 +998,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
const auto & hparams = model.hparams;
const int64_t n_vocab = vocab.n_tokens();
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_inp();
// when computing embeddings, all tokens are output
const bool output_all = cparams.embeddings;
@@ -2150,7 +2156,7 @@ void llama_context::opt_epoch_iter(
batch.logits [pos_batch] = true;
}
if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) {
LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
return;
}
src/llama-graph.cpp
+2 -2
View File
@@ -1142,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
// input embeddings with optional lora
ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd = hparams.n_embd_inp();
auto inp = std::make_unique<llm_graph_input_embd>();
@@ -1279,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
// return cur;
//}
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp();
const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
src/llama-hparams.cpp
+10
View File
@@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
return n_head/n_head_kv;
}
uint32_t llama_hparams::n_embd_inp() const {
uint32_t n_embd_inp = n_embd;
if (n_deepstack_layers > 0) {
n_embd_inp += n_embd * n_deepstack_layers;
}
return n_embd_inp;
}
uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
const uint32_t n_head_kv = this->n_head_kv(il);
src/llama-hparams.h
+3
View File
@@ -227,6 +227,9 @@ struct llama_hparams {
uint32_t n_gqa(uint32_t il = 0) const;
// dimension of main + auxiliary input embeddings
uint32_t n_embd_inp() const;
// dimension of key embeddings across all k-v heads
uint32_t n_embd_k_gqa(uint32_t il = 0) const;
src/llama-kv-cache-iswa.cpp
+3 -1
View File
@@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
const uint32_t size_base = kv_size;
uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad));
// note: the SWA cache is always padded to 256 for performance
// https://github.com/ggml-org/llama.cpp/issues/17037
uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256);
//kcpp: pad the swa kv cache as well, similar to extra_context_handle_fragmentation
size_swa += 128;
src/llama-model.cpp
+8 -16
View File
@@ -121,6 +121,7 @@
#include "models/wavtokenizer-dec.cpp"
#include "models/xverse.cpp"
#include "models/graph-context-mamba.cpp"
#include "models/pangu-embedded.cpp"
#if defined(GGML_USE_CLBLAST)
# include "ggml_v3b-opencl.h"
@@ -376,8 +377,8 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
} break;
case GGML_OP_IM2COL:
{
const int n_embd = hparams.n_embd;
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1);
const int n_embd_inp = hparams.n_embd_inp();
ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1);
op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16);
} break;
case GGML_OP_SCALE:
@@ -1139,9 +1140,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case 64: type = LLM_TYPE_32B; break;
default: type = LLM_TYPE_UNKNOWN;
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_QWEN3MOE:
{
@@ -1165,9 +1163,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case 94: type = LLM_TYPE_235B_A22B; break;
default: type = LLM_TYPE_UNKNOWN;
}
// since vision model stacks deepstack features along feature dim
// we also create a fake "n_embd" for text model to be the main embd + deepstack embds
hparams.n_embd *= hparams.n_deepstack_layers + 1;
} break;
case LLM_ARCH_PHI2:
{
@@ -3494,10 +3489,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_QWEN3:
case LLM_ARCH_QWEN3VL:
{
// for model loading, the weights only have the main embd
// so we need to divide by the number of deepstack layers + 1
// n_embd is const int so we declare a new variable
int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
@@ -3533,10 +3524,6 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_QWEN3MOE:
case LLM_ARCH_QWEN3VLMOE:
{
// for model loading, the weights only have the main embd
// so we need to divide by the number of deepstack layers + 1
// n_embd is const int so we declare a new variable
int64_t n_embd = hparams.n_embd / (hparams.n_deepstack_layers + 1);
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
@@ -6689,6 +6676,7 @@ void llama_model::print_info() const {
if (!hparams.vocab_only) {
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
LLAMA_LOG_INFO("%s: n_embd_inp = %u\n", __func__, hparams.n_embd_inp());
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str());
LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str());
@@ -7537,6 +7525,10 @@ int32_t llama_model_n_embd(const llama_model * model) {
return model->hparams.n_embd;
}
int32_t llama_model_n_embd_inp(const llama_model * model) {
return model->hparams.n_embd_inp();
}
int32_t llama_model_n_layer(const llama_model * model) {
return model->hparams.n_layer;
}
src/models/qwen3vl-moe.cpp
+1 -2
View File
@@ -1,9 +1,8 @@
#include "models.h"
llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
src/models/qwen3vl.cpp
+1 -4
View File
@@ -1,13 +1,10 @@
#include "models.h"
llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_full = hparams.n_embd; // main embd + deepstack embds
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = n_embd_full / (n_deepstack_layers + 1);
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);