vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/nix/package.nix
#	.github/labeler.yml
#	.gitignore
#	CMakeLists.txt
#	Makefile
#	Package.swift
#	README.md
#	ci/run.sh
#	docs/build.md
#	examples/CMakeLists.txt
#	flake.lock
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	grammars/README.md
#	requirements/requirements-convert_hf_to_gguf.txt
#	requirements/requirements-convert_hf_to_gguf_update.txt
#	scripts/check-requirements.sh
#	scripts/compare-llama-bench.py
#	scripts/gen-unicode-data.py
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
#	scripts/sync-ggml.sh
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tests/test-tokenizer-random.py
This commit is contained in:
Concedo 2024-07-11 16:36:16 +08:00
commit 2cad736260
85 changed files with 12568 additions and 445 deletions
Download patch file Download diff file Expand all files Collapse all files

316
src/llama.cpp
View file

@ -62,6 +62,12 @@
#include <io.h>
#endif
#if __cplusplus >= 202000L
#define LU8(x) (const char*)(u8##x)
#else
#define LU8(x) u8##x
#endif
#include <algorithm>
#include <array>
#include <cassert>
@ -253,6 +259,7 @@ enum llm_arch {
LLM_ARCH_OPENELM,
LLM_ARCH_ARCTIC,
LLM_ARCH_DEEPSEEK2,
LLM_ARCH_CHATGLM,
LLM_ARCH_BITNET,
LLM_ARCH_T5,
LLM_ARCH_JAIS,
@ -296,6 +303,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_OPENELM, "openelm" },
{ LLM_ARCH_ARCTIC, "arctic" },
{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
{ LLM_ARCH_CHATGLM, "chatglm" },
{ LLM_ARCH_BITNET, "bitnet" },
{ LLM_ARCH_T5, "t5" },
{ LLM_ARCH_JAIS, "jais" },
@ -1229,6 +1237,21 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
},
},
{
LLM_ARCH_CHATGLM,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
{ LLM_TENSOR_ROPE_FREQS, "rope_freqs" },
{ 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_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
},
},
{
LLM_ARCH_BITNET,
{
@ -2115,9 +2138,11 @@ enum e_model {
MODEL_2_8B,
MODEL_3B,
MODEL_4B,
MODEL_6B,
MODEL_6_9B,
MODEL_7B,
MODEL_8B,
MODEL_9B,
MODEL_11B,
MODEL_12B,
MODEL_13B,
@ -2143,7 +2168,6 @@ enum e_model {
MODEL_16x12B,
MODEL_10B_128x3_66B,
MODEL_57B_A14B,
MODEL_9B,
MODEL_27B,
};
@ -3804,6 +3828,9 @@ struct llama_model_loader {
case GGML_TYPE_IQ4_NL: ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL; break;
case GGML_TYPE_IQ4_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS; break;
case GGML_TYPE_IQ3_S: ftype = LLAMA_FTYPE_MOSTLY_IQ3_S; break;
case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
default:
{
LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@ -4510,6 +4537,9 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_MOSTLY_IQ4_XS: return "IQ4_XS - 4.25 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_S: return "IQ3_S - 3.4375 bpw";
case LLAMA_FTYPE_MOSTLY_IQ3_M: return "IQ3_S mix - 3.66 bpw";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
default: return "unknown, may not work";
}
@ -4543,9 +4573,11 @@ static const char * llama_model_type_name(e_model type) {
case MODEL_2_8B: return "2.8B";
case MODEL_3B: return "3B";
case MODEL_4B: return "4B";
case MODEL_6B: return "6B";
case MODEL_6_9B: return "6.9B";
case MODEL_7B: return "7B";
case MODEL_8B: return "8B";
case MODEL_9B: return "9B";
case MODEL_11B: return "11B";
case MODEL_12B: return "12B";
case MODEL_13B: return "13B";
@ -4571,7 +4603,6 @@ static const char * llama_model_type_name(e_model type) {
case MODEL_16x12B: return "16x12B";
case MODEL_10B_128x3_66B: return "10B+128x3.66B";
case MODEL_57B_A14B: return "57B.A14B";
case MODEL_9B: return "9B";
case MODEL_27B: return "27B";
default: return "?B";
}
@ -4678,16 +4709,6 @@ static void llm_load_hparams(
// non-transformer models do not have attention heads
if (hparams.n_head() > 0) {
// sanity check for n_rot (optional)
hparams.n_rot = hparams.n_embd / hparams.n_head();
ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
if (hparams.n_rot != hparams.n_embd / hparams.n_head()) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd / hparams.n_head()));
}
}
// gpt-neox n_rot = rotary_pct * (n_embd / n_head)
// gpt-j n_rot = rotary_dim
@ -4696,6 +4717,17 @@ static void llm_load_hparams(
hparams.n_embd_head_v = hparams.n_embd / hparams.n_head();
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false);
// sanity check for n_rot (optional)
hparams.n_rot = hparams.n_embd_head_k;
ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
if (model.arch == LLM_ARCH_LLAMA || model.arch == LLM_ARCH_FALCON) {
if (hparams.n_rot != hparams.n_embd_head_k) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
}
}
} else {
hparams.n_rot = 0;
hparams.n_embd_head_k = 0;
@ -5176,6 +5208,15 @@ static void llm_load_hparams(
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_CHATGLM:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
switch (hparams.n_layer) {
case 28: model.type = e_model::MODEL_6B; break;
case 40: model.type = e_model::MODEL_9B; break;
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
case LLM_ARCH_BITNET:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@ -5309,9 +5350,7 @@ static void llm_load_vocab(
if (merges_keyidx == -1) {
throw std::runtime_error("cannot find tokenizer merges in model file\n");
}
const int n_merges = gguf_get_arr_n(ctx, merges_keyidx);
for (int i = 0; i < n_merges; i++) {
const std::string word = gguf_get_arr_str(ctx, merges_keyidx, i);
if (!OldBPETokenizerMode)
@ -5463,6 +5502,10 @@ static void llm_load_vocab(
tokenizer_pre == "poro-chat") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
vocab.tokenizer_clean_spaces = false;
} else if (
tokenizer_pre == "chatglm-bpe") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHATGLM4;
vocab.special_bos_id = -1;
} else if (
tokenizer_pre == "viking") {
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING;
@ -5596,7 +5639,6 @@ static void llm_load_vocab(
vocab.special_eot_id = 107;
}
}
try {
vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
} catch (const std::exception & e) {
@ -7513,6 +7555,36 @@ static bool llm_load_tensors(
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
}
} break;
case LLM_ARCH_CHATGLM:
{
model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
// output
{
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab});
}
for (int i = 0; i < n_layer; ++i) {
ggml_context * ctx_layer = ctx_for_layer(i);
ggml_context * ctx_split = ctx_for_layer_split(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 + (hparams.n_embd_head_k << 2)});
layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + (hparams.n_embd_head_k << 2)});
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});
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff * 2});
layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
}
} break;
default:
throw std::runtime_error("unknown architecture");
}
@ -7737,6 +7809,7 @@ enum llm_ffn_op_type {
LLM_FFN_GELU,
LLM_FFN_RELU,
LLM_FFN_RELU_SQR,
LLM_FFN_SWIGLU,
};
enum llm_ffn_gate_type {
@ -7941,6 +8014,19 @@ static struct ggml_tensor * llm_build_ffn(
cur = ggml_sqr(ctx, cur);
cb(cur, "ffn_sqr(relu)", il);
} break;
case LLM_FFN_SWIGLU:
{
// Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
int64_t split_point = cur->ne[0] / 2;
struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
x0 = ggml_silu(ctx, x0);
cb(cur, "ffn_silu", il);
cur = ggml_mul(ctx, x0, x1);
cb(cur, "ffn_mul", il);
} break;
}
if (type_gate == LLM_FFN_PAR) {
@ -8129,7 +8215,7 @@ static struct ggml_tensor * llm_build_kqv(
struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
cb(kq, "kq", il);
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
// for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
// ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@ -10789,19 +10875,12 @@ struct llm_build_context {
// special-case: the up and gate tensors are merged into a single tensor
// TOOD: support into llm_build_ffn
{
struct ggml_tensor* up = ggml_mul_mat(ctx0, model.layers[il].ffn_up, cur);
cb(up, "ffn_up", il);
auto g = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), 0));
auto y = ggml_cont(ctx0, ggml_view_2d(ctx0, up, up->ne[0] / 2, up->ne[1], ggml_row_size(up->type, up->ne[0]), up->nb[1] / 2));
y = ggml_mul(ctx0, y, ggml_silu(ctx0, g));
cb(y, "ffn_gate", il);
auto down = ggml_mul_mat(ctx0, model.layers[il].ffn_down, y);
cb(down, "ffn_down", il);
cur = down;
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL, NULL,
NULL, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
}
@ -11571,7 +11650,7 @@ struct llm_build_context {
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, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
@ -11580,7 +11659,7 @@ struct llm_build_context {
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, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
@ -11684,7 +11763,7 @@ struct llm_build_context {
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, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);
@ -11693,7 +11772,7 @@ struct llm_build_context {
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, n_ctx_orig, freq_base, freq_scale,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Kcur, "Kcur", il);
@ -13214,6 +13293,8 @@ struct llm_build_context {
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
} else {
GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
struct ggml_tensor * embd_enc = llm_build_inp_embd_enc();
struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
@ -13493,6 +13574,120 @@ struct llm_build_context {
return gf;
}
struct ggml_cgraph * build_chatglm() {
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
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);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
// inp_pos - contains the positions
struct ggml_tensor * inp_pos = build_inp_pos();
// KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * inpSA = inpL;
cur = llm_build_norm(ctx0, inpL, hparams,
model.layers[il].attn_norm,
NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "attn_norm", il);
// self-attention
{
struct ggml_tensor * Qcur = nullptr;
struct ggml_tensor * Kcur = nullptr;
struct ggml_tensor * Vcur = nullptr;
cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
cb(cur, "wqkv", il);
cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
cb(cur, "bqkv", il);
Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
//printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
Qcur = ggml_rope_ext(
ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Qcur, "Qcur_rope", il);
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, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(Kcur, "Kcur_rope", il);
cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
model.layers[il].wo, NULL,
Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
}
if (il == n_layer - 1) {
// skip computing output for unused tokens
struct ggml_tensor * inp_out_ids = build_inp_out_ids();
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
// Add the input
struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// FF
{
cur = llm_build_norm(ctx0, ffn_inp, hparams,
model.layers[il].ffn_norm,
NULL,
LLM_NORM_RMS, cb, il);
cb(cur, "ffn_norm", il);
cur = llm_build_ffn(ctx0, cur,
model.layers[il].ffn_up, NULL, NULL,
NULL, NULL, NULL,
model.layers[il].ffn_down, NULL, NULL,
NULL,
LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
cb(cur, "ffn_out", il);
}
inpL = ggml_add(ctx0, cur, ffn_inp);
cb(inpL, "l_out", il);
}
cur = llm_build_norm(ctx0, inpL, hparams,
model.output_norm,
NULL,
LLM_NORM_RMS, cb, -1);
cb(cur, "result_norm", -1);
cur = ggml_mul_mat(ctx0, model.output, cur);
cb(cur, "result_output", -1);
ggml_build_forward_expand(gf, cur);
return gf;
}
};
static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@ -13724,6 +13919,10 @@ static struct ggml_cgraph * llama_build_graph(
{
result = llm.build_deepseek2();
} break;
case LLM_ARCH_CHATGLM:
{
result = llm.build_chatglm();
} break;
case LLM_ARCH_BITNET:
{
result = llm.build_bitnet();
@ -15560,6 +15759,11 @@ struct llm_tokenizer_bpe {
" ?[^(\\s|.,!?…。，、।۔،)]+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_CHATGLM4:
regex_exprs = {
"(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
};
break;
case LLAMA_VOCAB_PRE_TYPE_VIKING:
regex_exprs = {
"\\p{N}",
@ -16488,7 +16692,6 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
if (add_special) {
tokenizer.append_bos(output);
}
for (const auto & fragment : fragment_buffer) {
if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
@ -17915,6 +18118,10 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ3_S;
}
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
new_type == GGML_TYPE_Q4_0_8_8) {
new_type = GGML_TYPE_Q4_0;
}
}
} else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || ftype == LLAMA_FTYPE_MOSTLY_IQ1_S ||
ftype == LLAMA_FTYPE_MOSTLY_IQ2_S || ftype == LLAMA_FTYPE_MOSTLY_IQ2_M || ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) {
@ -18227,6 +18434,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_IQ4_XS: default_type = GGML_TYPE_IQ4_XS; break;
case LLAMA_FTYPE_MOSTLY_IQ3_S: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_IQ3_M: default_type = GGML_TYPE_IQ3_S; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -18537,6 +18747,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
f32_data = (float *) f32_conv_buf.data();
}
int chunk_size_multiplier = 1;
if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
}
LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
fflush(stdout);
@ -18549,7 +18767,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const int64_t nrows = tensor->ne[1];
static const int64_t min_chunk_size = 32 * 512;
const int64_t chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
chunk_size_multiplier;
const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
@ -19482,6 +19701,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
case LLM_ARCH_OLMO:
case LLM_ARCH_ARCTIC:
case LLM_ARCH_DEEPSEEK2:
case LLM_ARCH_CHATGLM:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2
@ -21225,7 +21445,6 @@ int32_t llama_tokenize(
bool add_special,
bool parse_special) {
auto res = llama_tokenize_internal(model->vocab, std::string(text, text_len), add_special, parse_special);
if (n_tokens_max < (int) res.size()) {
// LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
return -((int) res.size());
@ -21650,12 +21869,31 @@ static int32_t llama_chat_apply_template_internal(
if (add_ass) {
ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
}
} else if (tmpl == "minicpm" || tmpl_contains(u8"<用户>")) {
} else if (tmpl == "chatglm3" || tmpl_contains("[gMASK]sop")) {
// chatglm3-6b
ss << "[gMASK]" << "sop";
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>" << "\n " << message->content;
}
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == "chaglm4" || tmpl_contains("[gMASK]<sop>")) {
ss << "[gMASK]" << "<sop>";
for (auto message : chat) {
std::string role(message->role);
ss << "<|" << role << "|>" << "\n" << message->content;
}
if (add_ass) {
ss << "<|assistant|>";
}
} else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) {
// MiniCPM-3B-OpenHermes-2.5-v2-GGUF
for (auto message : chat) {
std::string role(message->role);
if (role == "user") {
ss << u8"<用户>";
ss << LU8("<用户>");
ss << trim(message->content);
ss << "<AI>";
} else {
@ -21671,7 +21909,7 @@ static int32_t llama_chat_apply_template_internal(
} else if (role == "user") {
ss << "User: " << message->content << "\n\n";
} else if (role == "assistant") {
ss << "Assistant: " << message->content << u8"<｜end▁of▁sentence｜>";
ss << "Assistant: " << message->content << LU8("<｜end▁of▁sentence｜>");
}
}
if (add_ass) {