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/backend/CANN.md
#	examples/model-conversion/Makefile
#	examples/model-conversion/scripts/causal/compare-embeddings-logits.sh
#	examples/model-conversion/scripts/causal/convert-model.sh
#	examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.py
#	examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh
#	examples/model-conversion/scripts/causal/run-converted-model.sh
#	examples/model-conversion/scripts/embedding/compare-embeddings-logits.sh
#	examples/model-conversion/scripts/embedding/convert-model.sh
#	examples/model-conversion/scripts/embedding/modelcard.template
#	examples/model-conversion/scripts/embedding/run-converted-model.sh
#	examples/model-conversion/scripts/utils/create-collection-add-model.sh
#	examples/model-conversion/scripts/utils/inspect-converted-model.sh
#	examples/model-conversion/scripts/utils/inspect-org-model.py
#	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-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/common.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	ggml/src/ggml-opencl/ggml-opencl.cpp
#	ggml/src/ggml-sycl/ggml-sycl.cpp
#	src/llama-context.cpp
#	tests/test-backend-ops.cpp
#	tests/test-chat.cpp
This commit is contained in:
Concedo
2025-09-05 13:25:34 +08:00
parent 979e2113e2 5d6688de08
commit f0d4128e9f
43 changed files with 1408 additions and 256 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-arch.cpp
+22
View File
@@ -45,6 +45,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_GEMMA2, "gemma2" },
{ LLM_ARCH_GEMMA3, "gemma3" },
{ LLM_ARCH_GEMMA3N, "gemma3n" },
{ LLM_ARCH_GEMMA_EMBEDDING, "gemma-embedding" },
{ LLM_ARCH_STARCODER2, "starcoder2" },
{ LLM_ARCH_MAMBA, "mamba" },
{ LLM_ARCH_MAMBA2, "mamba2" },
@@ -1038,6 +1039,27 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_LAUREL_POST_NORM, "blk.%d.laurel_post_norm" },
},
},
{
LLM_ARCH_GEMMA_EMBEDDING,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" },
},
},
{
LLM_ARCH_STARCODER2,
{
src/llama-arch.h
+1
View File
@@ -49,6 +49,7 @@ enum llm_arch {
LLM_ARCH_GEMMA2,
LLM_ARCH_GEMMA3,
LLM_ARCH_GEMMA3N,
LLM_ARCH_GEMMA_EMBEDDING,
LLM_ARCH_STARCODER2,
LLM_ARCH_MAMBA,
LLM_ARCH_MAMBA2,
src/llama-context.cpp
+5 -1
View File
@@ -285,6 +285,9 @@ llama_context::llama_context(
const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
// avoid reserving graphs with zero outputs
n_outputs = 1;
LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);
// resolve automatic Flash Attention use
@@ -1367,7 +1370,8 @@ llm_graph_result * llama_context::get_gf_res_reserve() const {
}
ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx, bool split_only) {
// LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
//LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
GGML_ASSERT_CONTINUE(n_outputs >= 1);
if (n_tokens % n_seqs != 0) {
n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
src/llama-graph.cpp
+50 -9
View File
@@ -258,6 +258,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
}
}
static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
(swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
(swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
(swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);
LLAMA_LOG_DEBUG(" ");
for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
LLAMA_LOG_DEBUG("%2d", j);
}
LLAMA_LOG_DEBUG("\n");
for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
LLAMA_LOG_DEBUG(" %2d ", i);
for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
float val = data[i * n_kv + j];
if (val == -INFINITY) {
LLAMA_LOG_DEBUG(" ∞");
} else {
LLAMA_LOG_DEBUG(" 0");
}
}
LLAMA_LOG_DEBUG("\n");
}
}
void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
const int64_t n_kv = ubatch->n_tokens;
const int64_t n_tokens = ubatch->n_tokens;
@@ -277,21 +307,32 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
const llama_seq_id s0 = ubatch->seq_id[i0][0];
// TODO: reimplement this like in llama_kv_cache
if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
if (hparams.use_alibi) {
f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
} else {
f = 0.0f;
}
break;
if (s0 != s1) {
continue; // skip different sequences
}
if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
continue; // skip future tokens for causal attention
}
if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
continue; // skip masked tokens for SWA
}
// TODO: reimplement this like in llama_kv_cache_unified
if (hparams.use_alibi) {
f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
} else {
f = 0.0f;
}
}
data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
}
}
}
if (debug) {
print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
}
}
void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
src/llama-graph.h
+8
View File
@@ -78,6 +78,11 @@ struct llm_graph_params;
class llm_graph_input_i {
public:
llm_graph_input_i() {
const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
}
virtual ~llm_graph_input_i() = default;
virtual void set_input(const llama_ubatch * ubatch) = 0;
@@ -90,6 +95,9 @@ public:
GGML_UNUSED(params);
return false;
}
protected:
// env: LLAMA_GRAPH_INPUT_DEBUG
int debug = 0;
};
using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
src/llama-hparams.cpp
+37
View File
@@ -1,6 +1,7 @@
#include "llama-hparams.h"
#include "ggml.h"
#include <cassert>
void llama_hparams::set_swa_pattern(uint32_t n_pattern, bool dense_first) {
if (dense_first) {
@@ -178,3 +179,39 @@ uint32_t llama_hparams::n_layer_kv() const {
return res;
}
bool llama_hparams::is_masked_swa(llama_pos p0, llama_pos p1) const {
assert(p0 >= 0 && p1 >= 0);
switch (swa_type) {
case LLAMA_SWA_TYPE_NONE:
{
} break;
case LLAMA_SWA_TYPE_STANDARD:
{
if (p1 - p0 >= (int32_t) n_swa) {
return true;
}
} break;
case LLAMA_SWA_TYPE_CHUNKED:
{
const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
if (p0 < pos_chunk_start) {
return true;
}
} break;
case LLAMA_SWA_TYPE_SYMMETRIC:
{
const int32_t half_n_swa = (int32_t) n_swa / 2;
const int32_t pos_diff = p1 - p0;
// Mask if outside the symmetric window
if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
return true;
}
} break;
}
return false;
}
src/llama-hparams.h
+6 -3
View File
@@ -16,9 +16,10 @@ enum llama_expert_gating_func_type {
};
enum llama_swa_type {
LLAMA_SWA_TYPE_NONE = 0,
LLAMA_SWA_TYPE_STANDARD = 1,
LLAMA_SWA_TYPE_CHUNKED = 2,
LLAMA_SWA_TYPE_NONE = 0,
LLAMA_SWA_TYPE_STANDARD = 1,
LLAMA_SWA_TYPE_CHUNKED = 2,
LLAMA_SWA_TYPE_SYMMETRIC = 3,
};
struct llama_hparams_posnet {
@@ -227,6 +228,8 @@ struct llama_hparams {
// number of layers for which has_kv() returns true
uint32_t n_layer_kv() const;
bool is_masked_swa(llama_pos p0, llama_pos p1) const;
};
static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");
src/llama-kv-cache-iswa.cpp
+2 -2
View File
@@ -64,14 +64,14 @@ llama_kv_cache_iswa::llama_kv_cache_iswa(
kv_base = std::make_unique<llama_kv_cache>(
model, type_k, type_v,
v_trans, offload, unified, size_base, n_seq_max, n_pad,
0, LLAMA_SWA_TYPE_NONE, filter_base, reuse);
0, filter_base, reuse);
LLAMA_LOG_INFO("%s: creating SWA KV cache, size = %u cells\n", __func__, size_swa);
kv_swa = std::make_unique<llama_kv_cache>(
model, type_k, type_v,
v_trans, offload, unified, size_swa, n_seq_max, n_pad,
hparams.n_swa, hparams.swa_type, filter_swa, reuse);
hparams.n_swa, filter_swa, reuse);
}
void llama_kv_cache_iswa::clear(bool data) {
src/llama-kv-cache.cpp
+2 -25
View File
@@ -27,11 +27,10 @@ llama_kv_cache::llama_kv_cache(
uint32_t n_seq_max,
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
const layer_filter_cb & filter,
const layer_reuse_cb & reuse) :
model(model), hparams(model.hparams), v_trans(v_trans),
n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa), swa_type(swa_type) {
n_seq_max(n_seq_max), n_stream(unified ? 1 : n_seq_max), n_pad(n_pad), n_swa(n_swa) {
GGML_ASSERT(kv_size % n_pad == 0);
@@ -1393,29 +1392,7 @@ ggml_cgraph * llama_kv_cache::build_graph_shift(llm_graph_result * res, llama_co
}
bool llama_kv_cache::is_masked_swa(llama_pos p0, llama_pos p1) const {
assert(p0 >= 0 && p1 >= 0);
switch (swa_type) {
case LLAMA_SWA_TYPE_NONE:
{
} break;
case LLAMA_SWA_TYPE_STANDARD:
{
if (p1 - p0 >= (int32_t) n_swa) {
return true;
}
} break;
case LLAMA_SWA_TYPE_CHUNKED:
{
const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
if (p0 < pos_chunk_start) {
return true;
}
} break;
}
return false;
return hparams.is_masked_swa(p0, p1);
}
void llama_kv_cache::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
src/llama-kv-cache.h
-3
View File
@@ -89,7 +89,6 @@ public:
uint32_t n_seq_max,
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
const layer_filter_cb & filter,
const layer_reuse_cb & reuse);
@@ -212,8 +211,6 @@ private:
// env: LLAMA_KV_CACHE_DEBUG
int debug = 0;
const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
std::vector<ggml_context_ptr> ctxs;
std::vector<ggml_backend_buffer_ptr> bufs;
src/llama-memory-hybrid.cpp
-2
View File
@@ -17,7 +17,6 @@ llama_memory_hybrid::llama_memory_hybrid(
uint32_t kv_size,
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
/* recurrent */
ggml_type type_r,
ggml_type type_s,
@@ -41,7 +40,6 @@ llama_memory_hybrid::llama_memory_hybrid(
n_seq_max,
n_pad,
n_swa,
swa_type,
filter_attn == nullptr ?
[&](int32_t il) { return !hparams.is_recurrent(il); }
: filter_attn,
src/llama-memory-hybrid.h
-1
View File
@@ -27,7 +27,6 @@ public:
uint32_t kv_size,
uint32_t n_pad,
uint32_t n_swa,
llama_swa_type swa_type,
/* recurrent */
ggml_type type_r,
ggml_type type_s,
src/llama-model.cpp
+158 -3
View File
@@ -1115,7 +1115,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 18: type = LLM_TYPE_537M; break;
case 18: type = LLM_TYPE_270M; break;
case 26: type = LLM_TYPE_1B; break;
case 34: type = LLM_TYPE_4B; break;
case 48: type = LLM_TYPE_12B; break;
@@ -1147,6 +1147,26 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}
} break;
case LLM_ARCH_GEMMA_EMBEDDING:
{
hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
hparams.set_swa_pattern(6);
hparams.causal_attn = false; // embeddings do not use causal attention
hparams.rope_freq_base_train_swa = 10000.0f;
hparams.rope_freq_scale_train_swa = 1.0f;
ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type);
switch (hparams.n_layer) {
case 24: type = LLM_TYPE_0_3B; break;
default: type = LLM_TYPE_UNKNOWN;
}
hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k));
} break;
case LLM_ARCH_STARCODER2:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -3580,6 +3600,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
} break;
case LLM_ARCH_GEMMA3:
case LLM_ARCH_GEMMA_EMBEDDING:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -11145,6 +11166,136 @@ struct llm_build_gemma3n_iswa : public llm_graph_context {
}
};
struct llm_build_gemma_embedding_iswa : public llm_graph_context {
llm_build_gemma_embedding_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_k;
ggml_tensor * cur;
ggml_tensor * inpL;
inpL = build_inp_embd(model.tok_embd);
// important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings)
if (ubatch.token) {
inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
cb(inpL, "inp_scaled", -1);
}
// inp_pos - contains the positions
ggml_tensor * inp_pos = build_inp_pos();
auto * inp_attn = build_attn_inp_no_cache();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
const float freq_base_l = model.get_rope_freq_base (cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
// norm
cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
cb(Qcur, "Qcur", il);
ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
cb(Kcur, "Kcur", il);
ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
cb(Qcur, "Qcur_normed", il);
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow);
Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
cb(Kcur, "Kcur_normed", il);
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
// ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
cur = build_attn(inp_attn,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
}
cur = build_norm(cur,
model.layers[il].attn_post_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "attn_post_norm", il);
ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
cb(sa_out, "sa_out", il);
cur = build_norm(sa_out,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "ffn_norm", il);
// feed-forward network
{
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, NULL,
model.layers[il].ffn_gate, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_GELU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
}
cur = build_norm(cur,
model.layers[il].ffn_post_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "ffn_post_norm", -1);
cur = ggml_add(ctx0, cur, sa_out);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
res->t_embd = cur;
ggml_build_forward_expand(gf, cur);
}
};
// TODO: move up next to build_starcoder
struct llm_build_starcoder2 : public llm_graph_context {
llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
@@ -18581,6 +18732,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
case LLM_ARCH_NOMIC_BERT_MOE:
case LLM_ARCH_NEO_BERT:
case LLM_ARCH_WAVTOKENIZER_DEC:
case LLM_ARCH_GEMMA_EMBEDDING:
case LLM_ARCH_DREAM:
case LLM_ARCH_LLADA:
{
@@ -18629,7 +18781,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
/* attn_kv_size */ cparams.n_ctx,
/* attn_n_pad */ padding,
/* attn_n_swa */ hparams.n_swa,
/* attn_swa_type */ hparams.swa_type,
/* recurrent_type_k */ GGML_TYPE_F32,
/* recurrent_type_v */ GGML_TYPE_F32,
/* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
@@ -18699,7 +18850,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
cparams.n_seq_max,
padding,
hparams.n_swa,
hparams.swa_type,
nullptr,
nullptr);
}
@@ -18861,6 +19011,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
{
llm = std::make_unique<llm_build_gemma3n_iswa>(*this, params);
} break;
case LLM_ARCH_GEMMA_EMBEDDING:
{
llm = std::make_unique<llm_build_gemma_embedding_iswa>(*this, params);
} break;
case LLM_ARCH_STARCODER2:
{
llm = std::make_unique<llm_build_starcoder2>(*this, params);
@@ -19261,6 +19415,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
case LLM_ARCH_GEMMA2:
case LLM_ARCH_GEMMA3:
case LLM_ARCH_GEMMA3N:
case LLM_ARCH_GEMMA_EMBEDDING:
case LLM_ARCH_STARCODER2:
case LLM_ARCH_OPENELM:
case LLM_ARCH_GPTNEOX:
src/llama-model.h
-1
View File
@@ -39,7 +39,6 @@ enum llm_type {
LLM_TYPE_410M,
LLM_TYPE_450M,
LLM_TYPE_475M,
LLM_TYPE_537M,
LLM_TYPE_558M,
LLM_TYPE_700M,
LLM_TYPE_770M,
src/llama-sampling.cpp
+65 -2
View File
@@ -604,10 +604,73 @@ static const char * llama_sampler_dist_name(const struct llama_sampler * /*smpl*
static void llama_sampler_dist_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_dist *) smpl->ctx;
// sorting is not necessary here
llama_sampler_softmax_impl(cur_p, false);
// edge cases
if (cur_p->size == 0) {
cur_p->selected = -1;
return;
}
cur_p->selected = 0;
if (cur_p->size == 1) {
cur_p->data[0].p = 1.0f;
return;
}
// max logit for numerical stability
float max_l = cur_p->data[0].logit;
if (!cur_p->sorted) {
for (size_t i = 1; i < cur_p->size; ++i) {
max_l = std::max(max_l, cur_p->data[i].logit);
}
}
// apply softmax to obtain the probabilities
double sum_cum = 0.0f;
for (size_t i = 0; i < cur_p->size; ++i) {
float p = expf(cur_p->data[i].logit - max_l);
cur_p->data[i].p = p;
sum_cum += p;
}
#if 1
// sample from the obtained probabilities and normalize the probs in a single pass
// this is ~3x faster on Mac with full gpt-oss vocab than the version below
//
std::uniform_real_distribution<double> dist(0.0f, 1.0f);
const double rnd = dist(ctx->rng);
double sum_run = 0.0f;
const double sum_tgt = sum_cum*rnd;
bool found = false;
for (size_t i = 0; i < cur_p->size; ++i) {
if (!found) {
// accumulate probs until we reach the target sum
sum_run += cur_p->data[i].p;
if (sum_run >= sum_tgt) {
cur_p->selected = i;
found = true;
}
}
// normalize probs
cur_p->data[i].p /= sum_cum;
}
// fallback to the last token (don't think this can happen)
assert(found);
if (!found) {
cur_p->selected = cur_p->size - 1;
}
#else
// for clarity, this is the same as above but does one pass for normalization and one extra pass for sampling
for (size_t i = 0; i < cur_p->size; ++i) {
cur_p->data[i].p /= sum_cum;
}
cur_p->selected = llama_sample_dist(cur_p, ctx->rng);
#endif
}
static struct llama_sampler * llama_sampler_dist_clone(const struct llama_sampler * smpl) {