Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '9f102a1407ed5d73b8c954f32edab50f8dfa3f58' into concedo_experimental

Browse Source 
# Conflicts:
#	.devops/intel.Dockerfile
#	.github/ISSUE_TEMPLATE/010-bug-compilation.yml
#	.github/ISSUE_TEMPLATE/011-bug-results.yml
#	.github/pull_request_template.md
#	CODEOWNERS
#	README.md
#	common/CMakeLists.txt
#	ggml/src/ggml-hexagon/ggml-hexagon.cpp
#	ggml/src/ggml-hexagon/htp/binary-ops.c
#	ggml/src/ggml-hexagon/htp/hex-dma.c
#	ggml/src/ggml-hexagon/htp/hex-dma.h
#	ggml/src/ggml-hexagon/htp/hex-dump.h
#	ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c
#	ggml/src/ggml-hexagon/htp/hvx-utils.h
#	ggml/src/ggml-hexagon/htp/main.c
#	ggml/src/ggml-hexagon/htp/ssm-conv.c
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/cvt.cl
#	ggml/src/ggml-rpc/ggml-rpc.cpp
#	scripts/snapdragon/adb/run-bench.sh
#	scripts/sync_vendor.py
#	tests/test-backend-ops.cpp
#	tools/llama-bench/llama-bench.cpp
This commit is contained in:
Concedo
2026-03-25 23:45:41 +08:00
parent 39938e19d3 9f102a1407
commit c00fe0af5a
32 changed files with 1302 additions and 447 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+2 -2
View File
@@ -2564,7 +2564,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_TOKEN_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_POS_EMBD, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_TOKEN_TYPES, {LLM_TENSOR_LAYER_INPUT, GGML_OP_GET_ROWS}},
{LLM_TENSOR_TOKEN_EMBD_NORM, {LLM_TENSOR_LAYER_INPUT, GGML_OP_MUL}},
{LLM_TENSOR_TOKEN_EMBD_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // do the norms on the first layer (not the input layer)
{LLM_TENSOR_OUTPUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
{LLM_TENSOR_CLS_OUT, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
@@ -2725,7 +2725,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
{LLM_TENSOR_LAUREL_POST_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
// this tensor is loaded for T5, but never used
{LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
{LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_INPUT, GGML_OP_IM2COL}},
{LLM_TENSOR_CONV1D, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_IM2COL}},
{LLM_TENSOR_POS_NET_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_POS_NET_NORM1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_POS_NET_NORM2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
src/llama-context.cpp
+7 -8
View File
@@ -350,14 +350,6 @@ llama_context::llama_context(
if (cparams.pipeline_parallel) {
LLAMA_LOG_INFO("%s: pipeline parallelism enabled\n", __func__);
if (!graph_reuse_disable) {
// TODO: figure out a way to make graph reuse work with pipeline parallelism
// ref: https://github.com/ggml-org/llama.cpp/pull/20463
LLAMA_LOG_WARN("%s: graph reuse is currently not compatible with pipeline parallelism - disabling\n", __func__);
graph_reuse_disable = true;
}
}
sched_reserve();
@@ -1199,6 +1191,13 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
if (!graph_reuse_disable && res->can_reuse(gparams)) {
//LLAMA_LOG_DEBUG("%s: reusing previous graph\n", __func__);
// with pipeline parallelism, the previous graph_compute_async may still be running
// on the GPU. we must synchronize before set_inputs to avoid overwriting input tensors
// that the previous compute is still reading.
if (cparams.pipeline_parallel) {
ggml_backend_sched_synchronize(sched.get());
}
n_reused++;
} else {
res->reset();
src/llama-memory-recurrent.cpp
+2 -5
View File
@@ -928,11 +928,8 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell
llama_seq_id seq_id;
io.read_to(&seq_id, sizeof(seq_id));
// TODO: llama_memory_recurrent should have a notion of max sequences
//if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) {
if (seq_id < 0) {
//LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx));
LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, inf)\n", __func__, seq_id);
if (seq_id < 0 || (uint32_t) seq_id >= this->n_seq_max) {
LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, this->n_seq_max);
return false;
}
src/llama-model.cpp
+15 -15
View File
@@ -3375,8 +3375,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"), {hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
}
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {n_embd}, 0);
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
@@ -3423,7 +3423,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_MODERN_BERT:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -3506,8 +3506,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); // word_embeddings
type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0); // token_type_embeddings
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); // LayerNorm
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0); //LayerNorm bias
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0); // LayerNorm
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {n_embd}, 0); // LayerNorm bias
cls = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, 1}, TENSOR_NOT_REQUIRED);
cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"), {1}, TENSOR_NOT_REQUIRED);
@@ -3558,8 +3558,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
case LLM_ARCH_BLOOM:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {n_embd}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -5938,8 +5938,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// Block 0, LN0
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {n_embd}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -6053,8 +6053,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// Block 0, LN0
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {n_embd}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
@@ -6225,8 +6225,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hparams.n_embd, n_vocab}, 0);
conv1d = create_tensor(tn(LLM_TENSOR_CONV1D, "weight"), {7, hparams.n_embd, hparams.posnet.n_embd}, 0);
conv1d_b = create_tensor(tn(LLM_TENSOR_CONV1D, "bias"), {1, hparams.posnet.n_embd}, 0);
conv1d = create_tensor(tn(LLM_TENSOR_CONV1D, "weight", 0), {7, hparams.n_embd, hparams.posnet.n_embd}, 0);
conv1d_b = create_tensor(tn(LLM_TENSOR_CONV1D, "bias", 0), {1, hparams.posnet.n_embd}, 0);
// posnet
{
@@ -6291,8 +6291,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
GGML_ASSERT(hparams.posnet.n_embd == hparams.convnext.n_embd);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {hparams.posnet.n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {hparams.posnet.n_embd}, 0);
tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {hparams.posnet.n_embd}, 0);
tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias", 0), {hparams.posnet.n_embd}, 0);
// convnext
{
src/llama.cpp
+4 -4
View File
@@ -388,14 +388,14 @@ static void llama_params_fit_impl(
case LAYER_FRACTION_ATTN: {
static std::array<std::string, n_strings> patterns;
if (patterns[il].empty()) {
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|gate|down).*";
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|up|gate_up|down).*";
}
return patterns[il].c_str();
}
case LAYER_FRACTION_UP: {
static std::array<std::string, n_strings> patterns;
if (patterns[il].empty()) {
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|down).*";
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|gate_up|down).*";
}
return patterns[il].c_str();
}
@@ -409,7 +409,7 @@ static void llama_params_fit_impl(
case LAYER_FRACTION_MOE: {
static std::array<std::string, n_strings> patterns;
if (patterns[il].empty()) {
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate)_(ch|)exps";
patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate_up|gate)_(ch|)exps";
}
return patterns[il].c_str();
}
@@ -503,7 +503,7 @@ static void llama_params_fit_impl(
int64_t global_surplus_cpu_moe = 0;
if (hp_nex > 0) {
const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate)_(ch|)exps"; // matches all MoE tensors
const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate_up|gate)_(ch|)exps"; // matches all MoE tensors
ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
tensor_buft_overrides[1] = {nullptr, nullptr};
src/models/bert.cpp
+2 -2
View File
@@ -28,8 +28,8 @@ llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params
cb(inpL, "inp_embd", -1);
// embed layer norm
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
cb(inpL, "inp_norm", -1);
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
cb(inpL, "inp_norm", 0);
auto * inp_attn = build_attn_inp_no_cache();
src/models/bloom.cpp
+2 -2
View File
@@ -16,8 +16,8 @@ llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_para
inpL = build_norm(inpL,
model.tok_norm,
model.tok_norm_b,
LLM_NORM, -1);
cb(inpL, "inp_norm", -1);
LLM_NORM, 0);
cb(inpL, "inp_norm", 0);
ggml_tensor * inp_out_ids = build_inp_out_ids();
src/models/modern-bert.cpp
+2 -2
View File
@@ -15,8 +15,8 @@ llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const ll
cb(inpL, "inp_embd", -1);
// embed layer norm
inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, -1);
cb(inpL, "inp_norm", -1);
inpL = build_norm(inpL, model.tok_norm, nullptr, LLM_NORM, 0);
cb(inpL, "inp_norm", 0);
ggml_tensor * inp_out_ids = build_inp_out_ids();
src/models/rwkv6.cpp
+1 -1
View File
@@ -8,7 +8,7 @@ llm_build_rwkv6::llm_build_rwkv6(const llama_model & model, const llm_graph_para
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
auto * rs_inp = build_rs_inp();
src/models/rwkv7.cpp
+1 -1
View File
@@ -9,7 +9,7 @@ llm_build_rwkv7::llm_build_rwkv7(const llama_model & model, const llm_graph_para
ggml_tensor * v_first = nullptr;
inpL = build_inp_embd(model.tok_embd);
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, 0);
auto * rs_inp = build_rs_inp();
src/models/wavtokenizer-dec.cpp
+1 -1
View File
@@ -93,7 +93,7 @@ llm_build_wavtokenizer_dec::llm_build_wavtokenizer_dec(const llama_model & model
cur = build_norm(cur,
model.tok_norm,
model.tok_norm_b,
LLM_NORM, -1);
LLM_NORM, 0);
cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));