vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo 2024-05-22 15:04:31 +08:00
commit 4a3c8c190b
25 changed files with 829 additions and 1596 deletions
Download patch file Download diff file Expand all files Collapse all files

288
llama.cpp
View file

@ -326,6 +326,7 @@ enum llm_kv {
LLM_KV_ROPE_SCALE_LINEAR,
LLM_KV_ROPE_SCALING_TYPE,
LLM_KV_ROPE_SCALING_FACTOR,
LLM_KV_ROPE_SCALING_ATTN_FACTOR,
LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,
LLM_KV_ROPE_SCALING_FINETUNED,
@ -403,6 +404,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" },
{ LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" },
{ LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" },
{ LLM_KV_ROPE_SCALING_ATTN_FACTOR, "%s.rope.scaling.attn_factor" },
{ LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" },
{ LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" },
@ -458,6 +460,8 @@ enum llm_tensor {
LLM_TENSOR_OUTPUT,
LLM_TENSOR_OUTPUT_NORM,
LLM_TENSOR_ROPE_FREQS,
LLM_TENSOR_ROPE_FACTORS_LONG,
LLM_TENSOR_ROPE_FACTORS_SHORT,
LLM_TENSOR_ATTN_Q,
LLM_TENSOR_ATTN_K,
LLM_TENSOR_ATTN_V,
@ -825,18 +829,20 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{
LLM_ARCH_PHI3,
{
{ 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_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_OUTPUT_NORM, "output_norm" },
{ LLM_TENSOR_OUTPUT, "output" },
{ LLM_TENSOR_ROPE_FACTORS_LONG, "rope_factors_long" },
{ LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
{ LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
{ LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" },
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
},
},
{
@ -1776,6 +1782,7 @@ struct llama_hparams {
float f_norm_eps;
float f_norm_rms_eps;
float rope_attn_factor = 1.0f;
float rope_freq_base_train;
float rope_freq_scale_train;
uint32_t n_yarn_orig_ctx;
@ -1824,6 +1831,7 @@ struct llama_hparams {
if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true;
if (!is_float_close(this->f_norm_rms_eps, other.f_norm_rms_eps, EPSILON)) return true;
if (!is_float_close(this->rope_attn_factor, other.rope_attn_factor, EPSILON)) return true;
if (!is_float_close(this->rope_freq_base_train, other.rope_freq_base_train, EPSILON)) return true;
if (!is_float_close(this->rope_freq_scale_train, other.rope_freq_scale_train, EPSILON)) return true;
@ -2133,6 +2141,10 @@ struct llama_model {
struct ggml_tensor * output;
struct ggml_tensor * output_b;
// long rope factors
struct ggml_tensor * rope_long = nullptr;
struct ggml_tensor * rope_short = nullptr;
std::vector<llama_layer> layers;
llama_split_mode split_mode;
@ -3337,6 +3349,39 @@ struct llama_model_loader {
return get_arr_n(llm_kv(kid), result, required);
}
template<typename T>
bool get_arr(const std::string & key, std::vector<T> & result, const bool required = true) {
const int kid = gguf_find_key(meta, key.c_str());
if (kid < 0) {
if (required) {
throw std::runtime_error(format("key not found in model: %s", key.c_str()));
}
return false;
}
struct GGUFMeta::ArrayInfo arr_info =
GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta, kid);
if (arr_info.gt != GGUF_TYPE_FLOAT32 && arr_info.gt != GGUF_TYPE_INT32) {
throw std::runtime_error(format("%s is not a float32 or int32 array", key.c_str()));
}
// GGML_ASSERT(gguf_type_size(arr_info.gt) == sizeof(T));
GGML_ASSERT((arr_info.gt != GGUF_TYPE_FLOAT32 || std::is_same<T, float>::value));
GGML_ASSERT((arr_info.gt != GGUF_TYPE_INT32 || std::is_same<T, int>::value));
result.resize(arr_info.length);
result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
return true;
}
template<typename T>
bool get_arr(const enum llm_kv kid, T& result, const bool required = true) {
return get_arr(llm_kv(kid), result, required);
}
template<typename T>
bool get_key(const std::string & key, T & result, const bool required = true) {
auto it = kv_overrides.find(key);
@ -3893,6 +3938,8 @@ static void llm_load_hparams(
}
hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale;
ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false);
// sanity check for n_rot (optional)
{
hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head;
@ -4950,6 +4997,7 @@ static bool llm_load_tensors(
// create tensors for the weights
{
const int64_t n_embd = hparams.n_embd;
const int64_t n_embd_head = n_embd / hparams.n_head;
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
const int64_t n_embd_gqa = n_embd_v_gqa;
@ -5661,6 +5709,9 @@ static bool llm_load_tensors(
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab });
model.rope_long = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight"), { n_embd_head/2 }, false);
model.rope_short = ml.create_tensor(ctx_input, tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight"), { n_embd_head/2 }, false);
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd });
@ -5671,12 +5722,12 @@ static bool llm_load_tensors(
ggml_context* ctx_layer = ctx_for_layer(i);
ggml_context* ctx_split = ctx_for_layer_split(i);
auto& layer = model.layers[i];
auto & layer = model.layers[i];
layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd });
layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, false);
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd });
layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd });
@ -6891,17 +6942,20 @@ struct llm_build_context {
cb(lctx.inp_K_shift, "K_shift", -1);
ggml_set_input(lctx.inp_K_shift);
struct ggml_tensor * rope_factors = build_rope_factors();
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * tmp =
// we rotate only the first n_rot dimensions
ggml_rope_custom_inplace(ctx0,
ggml_rope_ext_inplace(ctx0,
ggml_view_3d(ctx0, kv_self.k_l[il],
n_embd_head_k, n_head_kv, n_ctx,
ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
0),
lctx.inp_K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
lctx.inp_K_shift, rope_factors, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(tmp, "K_shifted", il);
ggml_build_forward_expand(gf, tmp);
}
@ -7004,6 +7058,17 @@ struct llm_build_context {
return lctx.inp_pos;
}
struct ggml_tensor * build_rope_factors() {
// choose long/short freq factors based on the context size
const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
if (n_ctx_pre_seq > hparams.n_yarn_orig_ctx) {
return model.rope_long;
}
return model.rope_short;
}
struct ggml_tensor * build_inp_out_ids() {
lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
cb(lctx.inp_out_ids, "inp_out_ids", -1);
@ -7111,15 +7176,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -7241,13 +7306,13 @@ struct llm_build_context {
switch (model.type) {
case MODEL_7B:
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -7353,15 +7418,15 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -7474,14 +7539,14 @@ struct llm_build_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
// using mode = 2 for neox mode
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -7597,15 +7662,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -7749,15 +7814,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -8102,15 +8167,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -8542,15 +8607,15 @@ struct llm_build_context {
}
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos,
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos,
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -8662,14 +8727,14 @@ struct llm_build_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
// using mode = 2 for neox mode
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -8773,15 +8838,15 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -8887,15 +8952,15 @@ struct llm_build_context {
Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -9039,8 +9104,8 @@ struct llm_build_context {
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);
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
@ -9050,8 +9115,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -9122,6 +9187,9 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
// rope freq factors for 128k context
struct ggml_tensor * rope_factors = build_rope_factors();
for (int il = 0; il < n_layer; ++il) {
auto residual = inpL;
@ -9158,8 +9226,8 @@ struct llm_build_context {
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);
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Qcur = ggml_rope_ext(
ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
@ -9167,8 +9235,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, Kcur, inp_pos, n_rot, rope_type, 0, n_orig_ctx,
Kcur = ggml_rope_ext(
ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -9274,14 +9342,14 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens), inp_pos, nullptr,
n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr,
n_embd_head, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
@ -9482,15 +9550,15 @@ struct llm_build_context {
cb(tmpk, "tmpk", il);
cb(Vcur, "Vcur", il);
struct ggml_tensor * Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos,
struct ggml_tensor * Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
struct ggml_tensor * Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos,
struct ggml_tensor * Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -9598,15 +9666,15 @@ struct llm_build_context {
// cb(Vcur, "Vcur", il);
// }
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -9715,15 +9783,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -9845,15 +9913,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -9965,8 +10033,8 @@ struct llm_build_context {
struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
cb(Vcur, "Vcur", il);
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr,
n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
@ -9974,8 +10042,8 @@ struct llm_build_context {
Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
cb(Qcur, "Qcur_scaled", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
n_embd_head_k, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
@ -10085,15 +10153,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -10375,15 +10443,15 @@ struct llm_build_context {
cb(Kcur, "Kcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -10506,15 +10574,15 @@ struct llm_build_context {
cb(Vcur, "Vcur", il);
}
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur", il);
Kcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
Kcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
@ -12789,15 +12857,16 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
// tokenizer.encode('', add_special_tokens=True) returns [1]
// tokenizer.encode('', add_special_tokens=False) returns []
if (add_special && vocab.special_add_bos != 0) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
}
static const bool rtrim = true; //TODO: as param
bool is_prev_special = false;
bool special_token_rtrim = false;
if (add_special && vocab.special_add_bos != 0) {
GGML_ASSERT(vocab.special_bos_id != -1);
output.push_back(vocab.special_bos_id);
is_prev_special = true;
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
// without adding this leading whitespace, we do not get the same results as the original tokenizer
@ -15719,6 +15788,7 @@ struct llama_context * llama_new_context_with_model(
cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
}
cparams.yarn_attn_factor *= hparams.rope_attn_factor;
cparams.causal_attn = hparams.causal_attn;
if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {