Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/check-vendor.yml
#	.github/workflows/close-issue.yml
#	.github/workflows/editorconfig.yml
#	.github/workflows/gguf-publish.yml
#	.github/workflows/labeler.yml
#	.github/workflows/pre-tokenizer-hashes.yml
#	.github/workflows/python-check-requirements.yml
#	.github/workflows/python-lint.yml
#	.github/workflows/python-type-check.yml
#	.github/workflows/server.yml
#	.github/workflows/update-ops-docs.yml
#	README.md
#	docs/build.md
#	examples/model-conversion/scripts/utils/perplexity-gen.sh
#	examples/model-conversion/scripts/utils/perplexity-run-simple.sh
#	examples/model-conversion/scripts/utils/perplexity-run.sh
#	examples/model-conversion/scripts/utils/quantize.sh
#	examples/model-conversion/scripts/utils/run-embedding-server.sh
#	ggml/src/ggml-cpu/ggml-cpu.c
#	ggml/src/ggml-hexagon/htp/flash-attn-ops.c
#	ggml/src/ggml-opencl/CMakeLists.txt
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-opencl/kernels/cvt.cl
#	ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	scripts/compare-llama-bench.py
#	tests/test-backend-ops.cpp
#	tests/test-gguf.cpp
#	tools/cli/README.md
#	tools/completion/README.md
#	tools/server/README.md
This commit is contained in:
Concedo
2026-01-27 23:06:13 +08:00
parent 3a84623ea5 68ac3acb43
commit f6ece6fd37
40 changed files with 3108 additions and 477 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-context.cpp
+6 -12
View File
@@ -803,7 +803,7 @@ 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));
}
const uint32_t n_embd_out = model.hparams.get_n_embd_out();
const uint32_t n_embd_out = model.hparams.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());
@@ -1289,7 +1289,7 @@ 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();
const uint32_t n_embd_out = hparams.n_embd_out();
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));
@@ -1698,7 +1698,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
{
// extract token embeddings
GGML_ASSERT(embd != nullptr);
const uint32_t n_embd_out = hparams.get_n_embd_out();
const uint32_t n_embd_out = hparams.n_embd_out();
float * embd_out = embd + n_outputs_prev*n_embd_out;
if (n_outputs) {
@@ -1831,7 +1831,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs, const llama_batch & ba
const auto n_batch = cparams.n_batch;
const auto n_vocab = vocab.n_tokens();
const auto n_embd_out = hparams.get_n_embd_out();
const auto n_embd_out = hparams.n_embd_out();
bool has_logits = true;
bool has_embd = cparams.embeddings;
@@ -2183,13 +2183,6 @@ llm_graph_cb llama_context::graph_get_cb() const {
ggml_set_name(cur, name);
}
if (!cparams.offload_kqv) {
if (strcmp(name, "kqv_merged_cont") == 0) {
// all nodes between the KV store and the attention output are run on the CPU
ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu);
}
}
// norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
// FIXME: fix in ggml_backend_sched
const bool full_offload = model.n_gpu_layers() > model.hparams.n_layer;
@@ -2569,6 +2562,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
}
}
// [TAG_CONTEXT_STATE_LOGITS]
// write logits
{
//LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
@@ -2913,7 +2907,7 @@ void llama_context::opt_epoch_iter(
};
ctx_compute_opt = ggml_init(params);
}
ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_inp_tokens(), res->get_logits());
ggml_opt_alloc(opt_ctx, train);
res->set_inputs(&ubatch);
src/llama-graph.cpp
+164 -23
View File
@@ -23,7 +23,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
}
if (ubatch->embd) {
const int64_t n_embd = embd->ne[0];
GGML_ASSERT(n_embd == embd->ne[0]);
const int64_t n_tokens = ubatch->n_tokens;
ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd));
@@ -33,8 +34,8 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
bool res = true;
res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
res &= (!embd && !params.ubatch.embd) || (embd && embd->ne[1] == params.ubatch.n_tokens);
res &= (!params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
res &= (!params.ubatch.embd) || (embd && embd->ne[1] == params.ubatch.n_tokens);
return res;
}
@@ -406,6 +407,27 @@ bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
return res;
}
void llm_graph_input_attn_k::set_input(const llama_ubatch * ubatch) {
mctx->set_input_k_idxs(self_k_idxs, ubatch);
mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
}
bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
this->mctx = mctx;
bool res = true;
res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
res &= self_kq_mask->ne[0] == mctx->get_n_kv();
res &= self_kq_mask->ne[1] == params.ubatch.n_tokens;
return res;
}
void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
@@ -634,7 +656,8 @@ int64_t llm_graph_result::get_max_nodes() const {
}
void llm_graph_result::reset() {
t_tokens = nullptr;
t_inp_tokens = nullptr;
t_inp_embd = nullptr;
t_logits = nullptr;
t_embd = nullptr;
t_embd_pooled = nullptr;
@@ -1338,17 +1361,29 @@ 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_inp();
const int64_t n_embd_inp = hparams.n_embd_inp();
const int64_t n_embd = hparams.n_embd;
auto inp = std::make_unique<llm_graph_input_embd>();
assert(n_embd_inp >= n_embd);
ggml_tensor * cur = nullptr;
auto inp = std::make_unique<llm_graph_input_embd>(n_embd_inp);
if (ubatch.token) {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
//cb(inp->tokens, "inp_tokens", -1);
ggml_set_input(inp->tokens);
res->t_tokens = inp->tokens;
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
cb(inp->tokens, "inp_tokens", -1);
ggml_set_input(inp->tokens);
res->t_inp_tokens = inp->tokens;
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd_inp, ubatch.n_tokens);
cb(inp->embd, "inp_embd", -1);
ggml_set_input(inp->embd);
// select one of the 2 inputs, based on the batch contents
// ref: https://github.com/ggml-org/llama.cpp/pull/18550
std::array<ggml_tensor *, 2> inps;
// token embeddings path (ubatch.token != nullptr)
{
auto & cur = inps[0];
cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
@@ -1369,19 +1404,36 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
cur = ggml_add(ctx0, cur, inpL_delta);
}
} else {
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
ggml_set_input(inp->embd);
if (n_embd_inp != n_embd) {
cur = ggml_pad(ctx0, cur, hparams.n_embd_inp() - n_embd, 0, 0, 0);
}
}
// vector embeddings path (ubatch.embd != nullptr)
{
auto & cur = inps[1];
cur = inp->embd;
}
assert(ggml_are_same_shape (inps[0], inps[1]));
assert(ggml_are_same_stride(inps[0], inps[1]));
ggml_tensor * cur = ggml_build_forward_select(gf, inps.data(), inps.size(), ubatch.token ? 0 : 1);
if (n_embd_inp != n_embd) {
cur = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0);
}
res->t_inp_embd = cur;
// For Granite architecture
if (hparams.f_embedding_scale != 0.0f) {
cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
}
cb(cur, "inp_embd", -1);
cb(cur, "embd", -1);
res->add_input(std::move(inp));
@@ -1480,7 +1532,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
//}
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;
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);
ggml_set_input(cur);
@@ -1565,11 +1617,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
v = ggml_transpose(ctx0, v);
}
// TODO: update llama_kv_cache to not store V cache in the MLA case and automatically return a view of K
if (v_mla) {
v = ggml_view_4d(ctx0, k, v->ne[0], v->ne[1], v->ne[2], v->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
}
// this can happen when KV cache is not used (e.g. an embedding model with non-causal attn)
if (k->type == GGML_TYPE_F32) {
k = ggml_cast(ctx0, k, GGML_TYPE_F16);
@@ -1583,6 +1630,11 @@ ggml_tensor * llm_graph_context::build_attn_mha(
hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
cb(cur, LLAMA_TENSOR_NAME_FATTN, il);
if (!cparams.offload_kqv) {
// all nodes between the KV store and the attention output are run on the CPU
ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu);
}
ggml_flash_attn_ext_add_sinks(cur, sinks);
ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
@@ -1792,9 +1844,11 @@ ggml_tensor * llm_graph_context::build_attn(
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * sinks,
ggml_tensor * v_mla,
ggml_tensor * v_mla, // TODO: remove
float kq_scale,
int il) const {
GGML_ASSERT(v_mla == nullptr);
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
// expand k later to enable rope fusion which directly writes into k-v cache
@@ -1837,6 +1891,93 @@ ggml_tensor * llm_graph_context::build_attn(
return cur;
}
static std::unique_ptr<llm_graph_input_attn_k> build_attn_inp_k_impl(
ggml_context * ctx0,
const llama_ubatch & ubatch,
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_kv_cache_context * mctx_cur) {
auto inp = std::make_unique<llm_graph_input_attn_k>(hparams, cparams, mctx_cur);
{
GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
const auto n_kv = mctx_cur->get_n_kv();
const auto n_tokens = ubatch.n_tokens;
const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;
inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, n_tokens/n_stream, 1, n_stream);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
}
return inp;
}
llm_graph_input_attn_k * llm_graph_context::build_attn_inp_k() const {
const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);
auto inp = build_attn_inp_k_impl(ctx0, ubatch, hparams, cparams, mctx_cur);
return (llm_graph_input_attn_k *) res->add_input(std::move(inp));
}
ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_k * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur,
ggml_tensor * k_cur,
ggml_tensor * v_cur,
ggml_tensor * kq_b,
ggml_tensor * sinks,
ggml_tensor * v_mla,
float kq_scale,
int il) const {
// these nodes are added to the graph together so that they are not reordered
// by doing so, the number of splits in the graph is reduced
// expand k later to enable rope fusion which directly writes into k-v cache
ggml_build_forward_expand(gf, q_cur);
ggml_build_forward_expand(gf, v_cur);
ggml_build_forward_expand(gf, k_cur);
const auto * mctx_cur = inp->mctx;
// store to KV cache
{
const auto & k_idxs = inp->get_k_idxs();
ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
}
const auto & kq_mask = inp->get_kq_mask();
ggml_tensor * q = q_cur;
ggml_tensor * k = mctx_cur->get_k(ctx0, il);
ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);
ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
cb(cur, "kqv_out", il);
if (wo) {
cur = build_lora_mm(wo, cur);
if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
// GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
}
}
if (wo_b) {
cur = ggml_add(ctx0, cur, wo_b);
}
return cur;
}
ggml_tensor * llm_graph_context::build_attn(
llm_graph_input_attn_kv_iswa * inp,
ggml_tensor * wo,
src/llama-graph.h
+54 -3
View File
@@ -106,7 +106,7 @@ using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
class llm_graph_input_embd : public llm_graph_input_i {
public:
llm_graph_input_embd() = default;
llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
virtual ~llm_graph_input_embd() = default;
void set_input(const llama_ubatch * ubatch) override;
@@ -115,6 +115,8 @@ public:
ggml_tensor * tokens = nullptr; // I32 [n_batch]
ggml_tensor * embd = nullptr; // F32 [n_embd, n_batch]
const int64_t n_embd = 0;
};
class llm_graph_input_pos : public llm_graph_input_i {
@@ -315,6 +317,39 @@ public:
const llama_kv_cache_context * mctx;
};
// V-less input for the KV cache
// ref: https://github.com/ggml-org/llama.cpp/pull/19067
class llm_graph_input_attn_k : public llm_graph_input_i {
public:
llm_graph_input_attn_k(
const llama_hparams & hparams,
const llama_cparams & cparams,
const llama_kv_cache_context * mctx) :
hparams(hparams),
cparams(cparams),
mctx(mctx) {
}
~llm_graph_input_attn_k() = default;
void set_input(const llama_ubatch * ubatch) override;
bool can_reuse(const llm_graph_params & params) override;
ggml_tensor * get_k_idxs() const { return self_k_idxs; }
ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch/n_stream, 1, n_stream]
const llama_hparams hparams;
const llama_cparams cparams;
const llama_kv_cache_context * mctx;
};
class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
public:
llm_graph_input_attn_kv_iswa(
@@ -566,7 +601,7 @@ public:
virtual ~llm_graph_result() = default;
ggml_tensor * get_tokens() const { return t_tokens; }
ggml_tensor * get_inp_tokens() const { return t_inp_tokens; }
ggml_tensor * get_logits() const { return t_logits; }
ggml_tensor * get_embd() const { return t_embd; }
ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
@@ -593,7 +628,8 @@ public:
void set_params(const llm_graph_params & params);
// important graph nodes
ggml_tensor * t_tokens = nullptr;
ggml_tensor * t_inp_tokens = nullptr;
ggml_tensor * t_inp_embd = nullptr; // [n_embd_inp, n_tokens]
ggml_tensor * t_logits = nullptr;
ggml_tensor * t_embd = nullptr;
ggml_tensor * t_embd_pooled = nullptr;
@@ -830,6 +866,21 @@ struct llm_graph_context {
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] // TODO: remove
float kq_scale,
int il) const;
llm_graph_input_attn_k * build_attn_inp_k() const;
ggml_tensor * build_attn(
llm_graph_input_attn_k * inp,
ggml_tensor * wo,
ggml_tensor * wo_b,
ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
ggml_tensor * kq_b,
ggml_tensor * sinks, // [n_head_q]
ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
float kq_scale,
int il) const;
src/llama-hparams.cpp
+17 -2
View File
@@ -72,8 +72,8 @@ uint32_t llama_hparams::n_embd_inp() const {
return n_embd_inp;
}
uint32_t llama_hparams::get_n_embd_out() const {
return n_embd_out > 0 ? n_embd_out : n_embd;
uint32_t llama_hparams::n_embd_out() const {
return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
}
uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
@@ -175,6 +175,21 @@ bool llama_hparams::is_swa(uint32_t il) const {
GGML_ABORT("fatal error");
}
bool llama_hparams::is_mla() const {
assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||
(n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));
return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;
}
uint32_t llama_hparams::n_embd_head_k_mla() const {
return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k;
}
uint32_t llama_hparams::n_embd_head_v_mla() const {
return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v;
}
bool llama_hparams::has_kv(uint32_t il) const {
if (n_layer_kv_from_start >= 0) {
if (il < (uint32_t) n_layer_kv_from_start) {
src/llama-hparams.h
+10 -4
View File
@@ -53,8 +53,8 @@ struct llama_hparams {
uint32_t n_rel_attn_bkts = 0;
// note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
uint32_t n_embd_head_k_mla = 0;
uint32_t n_embd_head_v_mla = 0;
uint32_t n_embd_head_k_mla_impl = 0;
uint32_t n_embd_head_v_mla_impl = 0;
// for WavTokenizer
struct llama_hparams_posnet posnet;
@@ -164,7 +164,7 @@ struct llama_hparams {
uint32_t n_cls_out = 1;
// output embedding dimension (0 = use n_embd)
uint32_t n_embd_out = 0;
uint32_t n_embd_out_impl = 0;
// llama4 smallthinker
uint32_t n_moe_layer_step = 0;
@@ -239,7 +239,7 @@ struct llama_hparams {
uint32_t n_embd_inp() const;
// dimension of output embeddings
uint32_t get_n_embd_out() const;
uint32_t n_embd_out() const;
// dimension of key embeddings across all k-v heads
uint32_t n_embd_k_gqa(uint32_t il = 0) const;
@@ -269,6 +269,12 @@ struct llama_hparams {
bool is_swa(uint32_t il) const;
// note: currently only support if either all or none of the layers are MLA
bool is_mla() const;
uint32_t n_embd_head_k_mla() const;
uint32_t n_embd_head_v_mla() const;
bool has_kv(uint32_t il) const;
// number of layers for which has_kv() returns true
src/llama-kv-cache.cpp
+28 -8
View File
@@ -97,6 +97,8 @@ llama_kv_cache::llama_kv_cache(
__func__, hparams.n_embd_v_gqa_max());
}
const bool is_mla = hparams.is_mla();
for (uint32_t il = 0; il < hparams.n_layer; il++) {
if (!hparams.has_kv(il)) {
LLAMA_LOG_DEBUG("%s: layer %3d: does not have KV cache\n", __func__, il);
@@ -130,18 +132,21 @@ llama_kv_cache::llama_kv_cache(
throw std::runtime_error("failed to create ggml context for kv cache");
}
ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
const bool has_k = true;
const bool has_v = !is_mla;
ggml_format_name(k, "cache_k_l%d", il);
ggml_format_name(v, "cache_v_l%d", il);
ggml_tensor * k = has_k ? ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream) : nullptr;
ggml_tensor * v = has_v ? ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream) : nullptr;
has_k && ggml_format_name(k, "cache_k_l%d", il);
has_v && ggml_format_name(v, "cache_v_l%d", il);
std::vector<ggml_tensor *> k_stream;
std::vector<ggml_tensor *> v_stream;
for (uint32_t s = 0; s < n_stream; ++s) {
k_stream.push_back(ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]));
v_stream.push_back(ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]));
k_stream.push_back(has_k ? ggml_view_2d(ctx, k, n_embd_k_gqa, kv_size, k->nb[1], s*k->nb[2]) : nullptr);
v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
}
map_layer_ids[il] = layers.size();
@@ -647,7 +652,10 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co
const auto & layer = layers[il];
ggml_backend_tensor_copy(layer.k_stream[ssrc], layer.k_stream[sdst]);
ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
if (layer.v_stream[ssrc]) {
ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
}
}
}
}
@@ -1516,7 +1524,7 @@ size_t llama_kv_cache::size_v_bytes() const {
size_t size_v_bytes = 0;
for (const auto & layer : layers) {
size_v_bytes += ggml_nbytes(layer.v);
size_v_bytes += layer.v ? ggml_nbytes(layer.v) : 0;
}
return size_v_bytes;
@@ -1798,6 +1806,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[cr.strm];
if (!v) {
continue;
}
// Write value type
const int32_t v_type_i = (int32_t) v->type;
@@ -1824,6 +1835,9 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[cr.strm];
if (!v) {
continue;
}
// Write value type
const int32_t v_type_i = (int32_t) v->type;
@@ -2027,6 +2041,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[strm];
if (!v) {
continue;
}
// Read type of value
int32_t v_type_i_ref;
@@ -2068,6 +2085,9 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
auto * v = layer.v_stream[strm];
if (!v) {
continue;
}
// Read type of value
int32_t v_type_i_ref;
src/llama-model-saver.cpp
+2 -2
View File
@@ -146,8 +146,8 @@ void llama_model_saver::add_kv_from_model() {
add_kv(LLM_KV_VOCAB_SIZE, vocab.n_tokens());
add_kv(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train);
add_kv(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd);
if (hparams.n_embd_out > 0) {
add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out);
if (hparams.n_embd_out_impl > 0) {
add_kv(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out_impl);
}
add_kv(LLM_KV_BLOCK_COUNT, hparams.n_layer);
add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
src/llama-model.cpp
+15 -16
View File
@@ -618,7 +618,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train);
ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd);
ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out, false);
ml.get_key(LLM_KV_EMBEDDING_LENGTH_OUT, hparams.n_embd_out_impl, false);
ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer);
ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false);
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
@@ -1803,15 +1803,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case LLM_ARCH_DEEPSEEK2:
{
// lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
if (!is_lite) {
ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
}
ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false);
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale, false);
@@ -1842,6 +1843,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
switch (hparams.n_layer) {
case 27: type = LLM_TYPE_16B; break;
case 47: type = LLM_TYPE_30B_A3B; break;
case 60: type = LLM_TYPE_236B; break;
case 61: type = LLM_TYPE_671B; break;
default: type = LLM_TYPE_UNKNOWN;
@@ -5062,14 +5064,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
} break;
case LLM_ARCH_DEEPSEEK2:
{
// lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
const bool is_mla = hparams.is_mla();
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
@@ -5094,13 +5093,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
if (!is_lite) {
if (q_lora_rank > 0) {
layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
}
layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
if (!is_lite) {
if (q_lora_rank > 0) {
layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
} else {
@@ -6750,7 +6749,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
// for LFM2-ColBert-350M
dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.get_n_embd_out()}, TENSOR_NOT_REQUIRED);
dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
} break;
case LLM_ARCH_SMALLTHINKER:
{
@@ -7470,8 +7469,8 @@ void llama_model::print_info() const {
LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead);
LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q);
LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv);
LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla);
LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla);
LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla());
LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla());
LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp);
LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared);
LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale);
@@ -8316,7 +8315,7 @@ int32_t llama_model_n_embd_inp(const llama_model * model) {
}
int32_t llama_model_n_embd_out(const llama_model * model) {
return model->hparams.get_n_embd_out();
return model->hparams.n_embd_out();
}
int32_t llama_model_n_layer(const llama_model * model) {
src/llama.cpp
+50 -16
View File
@@ -334,8 +334,12 @@ static void llama_params_fit_impl(
__func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
}
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
__func__, hp_nct, n_ctx_min);
if (n_ctx_min == UINT32_MAX) {
LLAMA_LOG_INFO("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
} else {
LLAMA_LOG_INFO("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
__func__, hp_nct, n_ctx_min);
}
}
} else {
LLAMA_LOG_INFO("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
@@ -1114,25 +1118,55 @@ int32_t llama_chat_apply_template(
// model split
//
int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count) {
int32_t llama_split_path(
char * split_path,
size_t maxlen,
const char * path_prefix,
int32_t split_no,
int32_t split_count) {
static const char * const SPLIT_PATH_FORMAT = "%s-%05d-of-%05d.gguf";
if (snprintf(split_path, maxlen, SPLIT_PATH_FORMAT, path_prefix, split_no + 1, split_count)) {
return strlen(split_path);
const int written = snprintf(
split_path,
maxlen,
SPLIT_PATH_FORMAT,
path_prefix,
split_no + 1,
split_count
);
if (written < 0 || (size_t) written >= maxlen) {
return 0;
}
return 0;
return (int32_t) written;
}
int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count) {
std::string str_split_path(split_path);
char postfix[32];
snprintf(postfix, 32, "-%05d-of-%05d.gguf", split_no + 1, split_count);
std::string str_postfix(postfix);
int32_t llama_split_prefix(
char * split_prefix,
size_t maxlen,
const char * split_path,
int32_t split_no,
int32_t split_count) {
// check if split_prefix ends with postfix
int size_prefix = str_split_path.size() - str_postfix.size();
if (size_prefix > 0 && str_split_path.find(str_postfix, size_prefix) != std::string::npos) {
snprintf(split_prefix, std::min((size_t) size_prefix + 1, maxlen), "%s", split_path);
return size_prefix;
const std::string str_split_path(split_path);
char postfix[32];
snprintf(postfix, sizeof(postfix), "-%05d-of-%05d.gguf", split_no + 1, split_count);
const std::string str_postfix(postfix);
if (str_split_path.size() <= str_postfix.size()) {
return 0;
}
const size_t size_prefix = str_split_path.size() - str_postfix.size();
if (str_split_path.compare(size_prefix, std::string::npos, str_postfix) == 0) {
const size_t copy_len = std::min(size_prefix + 1, maxlen);
snprintf(split_prefix, copy_len, "%s", split_path);
return (int32_t) size_prefix;
}
return 0;
src/models/deepseek2.cpp
+10 -9
View File
@@ -2,14 +2,11 @@
llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
llm_graph_context(params) {
// lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0);
const bool is_mla = hparams.is_mla();
// note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k;
const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v;
const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
const int64_t n_embd_head_qk_rope = hparams.n_rot;
const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
@@ -43,7 +40,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_kv();
auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr;
ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -57,6 +55,9 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
// self_attention
{
ggml_tensor * q = NULL;
const bool is_lite = model.layers[il].wq;
if (!is_lite) {
q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
cb(q, "q", il);
@@ -145,7 +146,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
}
// note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
cur = build_attn(inp_attn,
cur = build_attn(inp_attn_k,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
} else {
@@ -182,7 +183,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
}
// note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
cur = build_attn(inp_attn,
cur = build_attn(inp_attn_kv,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
}
src/models/gemma3n-iswa.cpp
+2 -2
View File
@@ -245,12 +245,12 @@ ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) {
// equivalent to get_per_layer_inputs() in python code
// output shape: [n_embd_altup, n_layer, n_tokens]
ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() {
auto inp = std::make_unique<llm_graph_input_embd>();
auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
ggml_tensor * inp_per_layer;
if (ubatch.token) {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
ggml_set_input(inp->tokens);
res->t_tokens = inp->tokens;
res->t_inp_tokens = inp->tokens;
inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup));
src/models/qwen3vl-moe.cpp
+5 -14
View File
@@ -2,7 +2,8 @@
llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = hparams.n_embd;
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);
@@ -16,17 +17,6 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -120,8 +110,9 @@ llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
if (ubatch.embd && (size_t)il < n_deepstack_layers) {
cur = ggml_add(ctx0, cur, deepstack_features[il]);
if (il < (int) n_deepstack_layers) {
ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
cur = ggml_add(ctx0, cur, ds);
cb(cur, "deepstack_out", il);
}
src/models/qwen3vl.cpp
+5 -14
View File
@@ -2,7 +2,8 @@
llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const size_t n_deepstack_layers = hparams.n_deepstack_layers;
const int64_t n_embd = hparams.n_embd;
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);
@@ -16,17 +17,6 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
int sections[4];
std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
if (ubatch.embd) {
// Image input: split main embd and deepstack embds
ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0);
for (size_t i = 0; i < n_deepstack_layers; i++) {
deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float));
}
inpL = inpL_main;
}
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
@@ -113,8 +103,9 @@ llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
if (ubatch.embd && (size_t)il < n_deepstack_layers) {
cur = ggml_add(ctx0, cur, deepstack_features[il]);
if (il < (int) n_deepstack_layers) {
ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
cur = ggml_add(ctx0, cur, ds);
cb(cur, "deepstack_out", il);
}