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:
#	docs/ops.md
#	docs/ops/WebGPU.csv
#	examples/model-conversion/logits.cpp
#	examples/retrieval/retrieval.cpp
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-webgpu/ggml-webgpu.cpp
#	ggml/src/ggml-webgpu/wgsl-shaders/unary_op.wgsl
This commit is contained in:
Concedo
2026-01-06 20:49:01 +08:00
parent 246ce4babd bd2a93d475
commit 3108fe740c
23 changed files with 182 additions and 63 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+13 -10
View File
@@ -768,7 +768,8 @@ float * llama_context::get_embeddings_ith(int32_t i) {
throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
}
return embd + j*model.hparams.n_embd;
const uint32_t n_embd_out = model.hparams.get_n_embd_out();
return embd + j*n_embd_out;
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
@@ -1204,9 +1205,10 @@ int llama_context::encode(const llama_batch & batch_inp) {
{
// extract token embeddings
GGML_ASSERT(embd != nullptr);
const uint32_t n_embd_out = hparams.get_n_embd_out();
GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size);
ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float));
GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size);
ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float));
} break;
case LLAMA_POOLING_TYPE_MEAN:
case LLAMA_POOLING_TYPE_CLS:
@@ -1610,12 +1612,13 @@ int llama_context::decode(const llama_batch & batch_inp) {
{
// extract token embeddings
GGML_ASSERT(embd != nullptr);
float * embd_out = embd + n_outputs_prev*n_embd;
const uint32_t n_embd_out = hparams.get_n_embd_out();
float * embd_out = embd + n_outputs_prev*n_embd_out;
if (n_outputs) {
GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size);
ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float));
GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd_out <= (int64_t) embd_size);
ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd_out*sizeof(float));
}
} break;
case LLAMA_POOLING_TYPE_MEAN:
@@ -1740,9 +1743,9 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());
const auto n_batch = cparams.n_batch;
const auto n_vocab = vocab.n_tokens();
const auto n_embd = hparams.n_embd;
const auto n_batch = cparams.n_batch;
const auto n_vocab = vocab.n_tokens();
const auto n_embd_out = hparams.get_n_embd_out();
bool has_logits = true;
bool has_embd = cparams.embeddings;
@@ -1783,7 +1786,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
// Allocate CPU logits buffer only if needed by sequences in this batch
logits_size = (has_logits && cpu_logits) ? n_vocab*n_outputs_max : 0;
embd_size = has_embd ? n_embd*n_outputs_max : 0;
embd_size = has_embd ? n_embd_out*n_outputs_max : 0;
// TODO: avoid this branching by working with the worst-case
if (!has_sampling) {