vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 17:44:38 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

Merge commit '584d674be6' into concedo_experimental

Browse source 
# Conflicts:
#	.github/workflows/nix-flake-update.yml
#	Makefile
#	Package.swift
#	ggml-cuda.cu
#	tests/test-quantize-fns.cpp
This commit is contained in:
Concedo 2024-01-14 16:29:44 +08:00
commit dc7bc0cb50
31 changed files with 2138 additions and 386 deletions
Download patch file Download diff file Expand all files Collapse all files

106
llama.cpp
View file

@ -2234,6 +2234,7 @@ struct llama_model_loader {
case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break;
case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break;
case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -2596,7 +2597,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K";
case LLAMA_FTYPE_MOSTLY_Q2_K: return "Q2_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large";
@ -2606,6 +2608,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium";
case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K";
case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
default: return "unknown, may not work";
}
@ -2840,6 +2843,7 @@ static void llm_load_hparams(
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
switch (hparams.n_layer) {
case 24: model.type = e_model::MODEL_1B; break;
case 32: model.type = e_model::MODEL_3B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
@ -3175,7 +3179,15 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown");
LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type));
LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str());
LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9);
if (ml.n_elements >= 1e12) {
LLAMA_LOG_INFO("%s: model params = %.2f T\n", __func__, ml.n_elements*1e-12);
} else if (ml.n_elements >= 1e9) {
LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9);
} else if (ml.n_elements >= 1e6) {
LLAMA_LOG_INFO("%s: model params = %.2f M\n", __func__, ml.n_elements*1e-6);
} else {
LLAMA_LOG_INFO("%s: model params = %.2f K\n", __func__, ml.n_elements*1e-3);
}
if (ml.n_bytes < GiB) {
LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements);
} else {
@ -4121,7 +4133,6 @@ static void llm_build_k_shift(
struct ggml_cgraph * graph,
llm_rope_type type,
int64_t n_ctx,
int n_rot,
float freq_base,
float freq_scale,
const llm_build_cb & cb) {
@ -4129,14 +4140,13 @@ static void llm_build_k_shift(
const int64_t n_head_kv = hparams.n_head_kv;
const int64_t n_embd_head_k = hparams.n_embd_head_k;
const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
const int32_t n_rot = hparams.n_rot;
const int32_t n_orig_ctx = cparams.n_yarn_orig_ctx;
const float ext_factor = cparams.yarn_ext_factor;
const float attn_factor = cparams.yarn_attn_factor;
const float beta_fast = cparams.yarn_beta_fast;
const float beta_slow = cparams.yarn_beta_slow;
GGML_ASSERT(n_embd_head_k % n_rot == 0);
struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx);
cb(K_shift, "K_shift", -1);
@ -4540,7 +4550,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -4578,14 +4588,14 @@ struct llm_build_context {
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
hparams.n_rot, 0, 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,
n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -4708,6 +4718,7 @@ struct llm_build_context {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4725,7 +4736,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -4751,12 +4762,12 @@ struct llm_build_context {
case MODEL_7B:
Qcur = ggml_rope_custom(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
hparams.n_rot, 0, 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,
n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
break;
@ -4829,6 +4840,7 @@ struct llm_build_context {
const int64_t n_embd_head = hparams.n_embd_head_v;
const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4846,7 +4858,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -4887,13 +4899,13 @@ struct llm_build_context {
// using mode = 2 for neox mode
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
ctx0, Qcur, inp_pos, hparams.n_rot, 2, 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_embd_head, 2, 0, n_orig_ctx,
ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -5050,15 +5062,14 @@ struct llm_build_context {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
const int64_t n_rot = n_embd_head_k / 2;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head/2 == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
cb(inpL, "imp_embd", -1);
cb(inpL, "inp_embd", -1);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
@ -5069,7 +5080,7 @@ struct llm_build_context {
cb(KQ_mask, "KQ_mask", -1);
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -5129,7 +5140,7 @@ struct llm_build_context {
// RoPE the first n_rot of q/k, pass the other half, and concat.
struct ggml_tensor * qrot = ggml_view_3d(
ctx0, tmpq, n_rot, n_head, n_tokens,
ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
0
@ -5137,7 +5148,7 @@ struct llm_build_context {
cb(qrot, "qrot", il);
struct ggml_tensor * krot = ggml_view_3d(
ctx0, tmpk, n_rot, n_head, n_tokens,
ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpk) * n_embd_head,
ggml_element_size(tmpk) * n_embd_head * n_head,
0
@ -5146,29 +5157,29 @@ struct llm_build_context {
// get the second half of tmpq, e.g tmpq[n_rot:, :, :]
struct ggml_tensor * qpass = ggml_view_3d(
ctx0, tmpq, n_rot, n_head, n_tokens,
ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpq) * n_embd_head,
ggml_element_size(tmpq) * n_embd_head * n_head,
ggml_element_size(tmpq) * n_rot
ggml_element_size(tmpq) * hparams.n_rot
);
cb(qpass, "qpass", il);
struct ggml_tensor * kpass = ggml_view_3d(
ctx0, tmpk, n_rot, n_head, n_tokens,
ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
ggml_element_size(tmpk) * n_embd_head,
ggml_element_size(tmpk) * n_embd_head * n_head,
ggml_element_size(tmpk) * n_rot
ggml_element_size(tmpk) * hparams.n_rot
);
cb(kpass, "kpass", il);
struct ggml_tensor * qrotated = ggml_rope_custom(
ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx,
ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(qrotated, "qrotated", il);
struct ggml_tensor * krotated = ggml_rope_custom(
ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx,
ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(krotated, "krotated", il);
@ -5565,7 +5576,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -5678,7 +5689,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -5710,13 +5721,13 @@ struct llm_build_context {
// using mode = 2 for neox mode
Qcur = ggml_rope_custom(
ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
ctx0, Qcur, inp_pos, hparams.n_rot, 2, 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_embd_head, 2, 0, n_orig_ctx,
ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur", il);
@ -5795,7 +5806,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -5891,6 +5902,7 @@ struct llm_build_context {
const int64_t n_embd_head = hparams.n_embd_head_v;
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -5908,7 +5920,7 @@ struct llm_build_context {
// shift the entire K-cache if needed
if (do_rope_shift) {
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
}
for (int il = 0; il < n_layer; ++il) {
@ -5934,13 +5946,13 @@ 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,
ctx0, ggml_reshape_3d(ctx0, Qcur, hparams.n_rot, n_head, n_tokens), inp_pos,
n_embd_head, 2, 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,
ctx0, ggml_reshape_3d(ctx0, Kcur, hparams.n_rot, n_head_kv, n_tokens), inp_pos,
n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
@ -9277,10 +9289,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
// TODO: explore better strategies
new_type = GGML_TYPE_Q8_0;
}
} else if (name.find("ffn_down.weight") != std::string::npos) {
} else if (name.find("ffn_down") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q5_K
: arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
@ -9289,14 +9304,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
if (arch == LLM_ARCH_FALCON) {
new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q6_K :
use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
} else {
if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
}
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) {
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) {
new_type = GGML_TYPE_Q5_K;
}
++qs.i_feed_forward_w2;
@ -9314,9 +9329,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
}
// IK: let's remove this, else Q2_K is almost the same as Q3_K_S
//else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
//}
// This can be used to reduce the size of the Q5_K_S model.
// The associated PPL increase is fully in line with the size reduction
//else {
@ -9365,6 +9381,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@ -9374,6 +9391,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break;
case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -9422,7 +9440,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
++qs.n_attention_wv;
}
else if (name.find("ffn_down.weight") != std::string::npos) {
else if (name.find("ffn_down") != std::string::npos) {
++qs.n_feed_forward_w2;
}
}
@ -11262,7 +11280,7 @@ void llama_print_timings(struct llama_context * ctx) {
__func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n",
__func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
LLAMA_LOG_INFO("%s: total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (timings.t_end_ms - timings.t_start_ms), (timings.n_p_eval + timings.n_eval));
}
void llama_reset_timings(struct llama_context * ctx) {