vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into falcon

Browse source
This commit is contained in:
Georgi Gerganov 2023-08-22 20:15:01 +03:00
commit b19c6e4640
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
6 changed files with 199 additions and 82 deletions
Download patch file Download diff file Expand all files Collapse all files

29
convert.py
View file

@ -69,7 +69,10 @@ SAFETENSORS_DATA_TYPES: Dict[str, DataType] = {
'I32': DT_I32,
}
class GGMLFileType(enum.Enum):
# TODO: match this with `llama_ftype`
# TODO: rename to LLAMAFileType
# TODO: move to `gguf.py`
class GGMLFileType(enum.IntEnum):
AllF32 = 0
MostlyF16 = 1 # except 1d tensors
@ -101,6 +104,8 @@ class Params:
n_head_kv: int
f_norm_eps: float
ftype: Optional[GGMLFileType] = None
@staticmethod
def find_n_mult(n_ff: int, n_embd: int) -> int:
# hardcoded magic range
@ -738,6 +743,9 @@ class OutputFile:
self.gguf.add_head_count_kv (params.n_head_kv)
self.gguf.add_layer_norm_rms_eps (params.f_norm_eps)
if params.ftype:
self.gguf.add_file_type(params.ftype)
def add_meta_vocab(self, vocab: Vocab) -> None:
tokens = []
scores = []
@ -1020,6 +1028,12 @@ def main(args_in: Optional[List[str]] = None) -> None:
" - LLaMA v2: --ctx 4096\n")
params.n_ctx = args.ctx
if args.outtype:
params.ftype = {
"f32": GGMLFileType.AllF32,
"f16": GGMLFileType.MostlyF16,
}[args.outtype]
print(f"params = {params}")
vocab: Vocab
@ -1040,11 +1054,14 @@ def main(args_in: Optional[List[str]] = None) -> None:
vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
vocab = load_vocab(vocab_dir, args.vocabtype)
model = model_plus.model
model = convert_model_names(model, params)
output_type = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, output_type)
outfile = args.outfile or default_outfile(model_plus.paths, output_type)
model = model_plus.model
model = convert_model_names(model, params)
ftype = pick_output_type(model, args.outtype)
model = convert_to_output_type(model, ftype)
outfile = args.outfile or default_outfile(model_plus.paths, ftype)
params.ftype = ftype
print(f"Writing {outfile}, format {ftype}")
OutputFile.write_all(outfile, params, model, vocab)
print(f"Wrote {outfile}")

37
examples/embedding/embedding.cpp
View file

@ -72,23 +72,30 @@ int main(int argc, char ** argv) {
fprintf(stderr, "\n");
}
if (params.embedding){
if (embd_inp.size() > 0) {
if (llama_eval(ctx, embd_inp.data(), embd_inp.size(), n_past, params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return 1;
}
}
const int n_embd = llama_n_embd(ctx);
const auto embeddings = llama_get_embeddings(ctx);
for (int i = 0; i < n_embd; i++) {
printf("%f ", embeddings[i]);
}
printf("\n");
if (embd_inp.size() > (size_t)params.n_ctx) {
fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
__func__, embd_inp.size(), params.n_ctx);
return 1;
}
while (!embd_inp.empty()) {
int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
if (llama_eval(ctx, embd_inp.data(), n_tokens, n_past, params.n_threads)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
return 1;
}
n_past += n_tokens;
embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
}
const int n_embd = llama_n_embd(ctx);
const auto embeddings = llama_get_embeddings(ctx);
for (int i = 0; i < n_embd; i++) {
printf("%f ", embeddings[i]);
}
printf("\n");
llama_print_timings(ctx);
llama_free(ctx);
llama_free_model(model);

4
gguf.py
View file

@ -26,6 +26,7 @@ KEY_GENERAL_DESCRIPTION = "general.description"
KEY_GENERAL_LICENSE = "general.license"
KEY_GENERAL_SOURCE_URL = "general.source.url"
KEY_GENERAL_SOURCE_HF_REPO = "general.source.hugginface.repository"
KEY_GENERAL_FILE_TYPE = "general.file_type"
# LLM
KEY_CONTEXT_LENGTH = "{arch}.context_length"
@ -595,6 +596,9 @@ class GGUFWriter:
def add_source_hf_repo(self, repo: str):
self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo)
def add_file_type(self, ftype: int):
self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype)
def add_name(self, name: str):
self.add_string(KEY_GENERAL_NAME, name)

164
k_quants.c
View file

@ -77,6 +77,11 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
}
return 1/iscale;
}
bool return_early = false;
if (rmse_type < 0) {
rmse_type = -rmse_type;
return_early = true;
}
int weight_type = rmse_type%2;
float sumlx = 0;
float suml2 = 0;
@ -89,56 +94,9 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t *
suml2 += w*l*l;
}
float scale = sumlx/suml2;
if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
float best = scale * sumlx;
for (int itry = 0; itry < 3; ++itry) {
iscale = 1/scale;
float slx = 0;
float sl2 = 0;
bool changed = false;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
l = MAX(-nmax, MIN(nmax-1, l));
if (l + nmax != L[i]) { changed = true; }
float w = weight_type == 1 ? x[i] * x[i] : 1.f;
slx += w*x[i]*l;
sl2 += w*l*l;
}
if (!changed || sl2 == 0 || slx*slx <= best*sl2) { break; }
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale * x[i]);
L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
}
sumlx = slx; suml2 = sl2;
scale = sumlx/suml2;
best = scale * sumlx;
}
for (int itry = 0; itry < 5; ++itry) {
int n_changed = 0;
for (int i = 0; i < n; ++i) {
float w = weight_type == 1 ? x[i]*x[i] : 1;
int l = L[i] - nmax;
float slx = sumlx - w*x[i]*l;
if (slx > 0) {
float sl2 = suml2 - w*l*l;
int new_l = nearest_int(x[i] * sl2 / slx);
new_l = MAX(-nmax, MIN(nmax-1, new_l));
if (new_l != l) {
slx += w*x[i]*new_l;
sl2 += w*new_l*new_l;
if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
L[i] = nmax + new_l; sumlx = slx; suml2 = sl2;
scale = sumlx / suml2; best = scale * sumlx;
++n_changed;
}
}
}
}
if (!n_changed) { break; }
}
if (rmse_type < 3) {
return scale;
}
for (int is = -4; is <= 4; ++is) {
for (int is = -9; is <= 9; ++is) {
if (is == 0) {
continue;
}
@ -221,12 +179,17 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t *
return 1/iscale;
}
static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, int ntry) {
static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
int ntry, float alpha) {
float min = x[0];
float max = x[0];
float sum_x = 0;
float sum_x2 = 0;
for (int i = 1; i < n; ++i) {
if (x[i] < min) min = x[i];
if (x[i] > max) max = x[i];
sum_x += x[i];
sum_x2 += x[i]*x[i];
}
if (max == min) {
for (int i = 0; i < n; ++i) L[i] = 0;
@ -254,7 +217,7 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
for (int i = 0; i < n; ++i) {
sum += x[i] - scale*L[i];
}
min = sum/n;
min = alpha*min + (1 - alpha)*sum/n;
if (min > 0) min = 0;
iscale = 1/scale;
if (!did_change) break;
@ -263,6 +226,82 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t
return scale;
}
static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
float rmin, float rdelta, int nstep, bool use_mad) {
float min = x[0];
float max = x[0];
float sum_w = weights[0];
float sum_x = sum_w * x[0];
for (int i = 1; i < n; ++i) {
if (x[i] < min) min = x[i];
if (x[i] > max) max = x[i];
float w = weights[i];
sum_w += w;
sum_x += w * x[i];
}
if (min > 0) min = 0;
if (max == min) {
for (int i = 0; i < n; ++i) L[i] = 0;
*the_min = -min;
return 0.f;
}
float iscale = nmax/(max - min);
float scale = 1/iscale;
float best_mad = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
L[i] = MAX(0, MIN(nmax, l));
float diff = scale * L[i] + min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
best_mad += w * diff;
}
if (nstep < 1) {
*the_min = -min;
return scale;
}
for (int is = 0; is <= nstep; ++is) {
iscale = (rmin + rdelta*is + nmax)/(max - min);
float sum_l = 0, sum_l2 = 0, sum_xl = 0;
for (int i = 0; i < n; ++i) {
int l = nearest_int(iscale*(x[i] - min));
l = MAX(0, MIN(nmax, l));
Laux[i] = l;
float w = weights[i];
sum_l += w*l;
sum_l2 += w*l*l;
sum_xl += w*l*x[i];
}
float D = sum_w * sum_l2 - sum_l * sum_l;
if (D > 0) {
float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
float this_min = (sum_l2 * sum_x - sum_l * sum_xl)/D;
if (this_min > 0) {
this_min = 0;
this_scale = sum_xl / sum_l2;
}
float mad = 0;
for (int i = 0; i < n; ++i) {
float diff = this_scale * Laux[i] + this_min - x[i];
diff = use_mad ? fabsf(diff) : diff * diff;
float w = weights[i];
mad += w * diff;
}
if (mad < best_mad) {
for (int i = 0; i < n; ++i) {
L[i] = Laux[i];
}
best_mad = mad;
scale = this_scale;
min = this_min;
}
}
}
*the_min = -min;
return scale;
}
#if QK_K == 256
static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
if (j < 4) {
@ -281,6 +320,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
const int nb = k / QK_K;
uint8_t L[QK_K];
uint8_t Laux[16];
float weights[16];
float mins[QK_K/16];
float scales[QK_K/16];
@ -291,7 +332,8 @@ void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/16; ++j) {
scales[j] = make_qkx1_quants(16, 3, x + 16*j, L + 16*j, &mins[j], 5);
for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;
@ -637,6 +679,8 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
const int nb = k / QK_K;
uint8_t L[QK_K];
uint8_t Laux[32];
float weights[32];
float mins[QK_K/32];
float scales[QK_K/32];
@ -645,7 +689,12 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/32; ++j) {
scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 5);
//scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
float sum_x2 = 0;
for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
float av_x = sqrtf(sum_x2/32);
for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;
@ -798,6 +847,8 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
uint8_t L[QK_K];
float mins[QK_K/32];
float scales[QK_K/32];
float weights[32];
uint8_t Laux[32];
#else
int8_t L[QK_K];
float scales[QK_K/16];
@ -810,7 +861,12 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
float max_scale = 0; // as we are deducting the min, scales are always positive
float max_min = 0;
for (int j = 0; j < QK_K/32; ++j) {
scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 5);
//scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
float sum_x2 = 0;
for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
float av_x = sqrtf(sum_x2/32);
for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
float scale = scales[j];
if (scale > max_scale) {
max_scale = scale;

45
llama.cpp
View file

@ -1219,6 +1219,16 @@ struct llama_model_loader {
} break;
}
// this is a way to mark that we have "guessed" the file type
ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED);
{
const int kid = gguf_find_key(ctx_gguf, "general.file_type");
if (kid >= 0) {
ftype = (llama_ftype) gguf_get_val_u32(ctx_gguf, kid);
}
}
for (int i = 0; i < n_kv; i++) {
const char * name = gguf_get_key(ctx_gguf, i);
const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
@ -1421,7 +1431,11 @@ struct llama_model_loader {
// load LLaMA models
//
const char * llama_model_ftype_name(enum llama_ftype ftype) {
std::string llama_model_ftype_name(enum llama_ftype ftype) {
if (ftype & LLAMA_FTYPE_GUESSED) {
return llama_model_ftype_name((enum llama_ftype) (ftype & ~LLAMA_FTYPE_GUESSED)) + " (guessed)";
}
switch (ftype) {
case LLAMA_FTYPE_ALL_F32: return "all F32";
case LLAMA_FTYPE_MOSTLY_F16: return "mostly F16";
@ -1643,7 +1657,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base);
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale);
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));
LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str());
LLAMA_LOG_INFO("%s: model size = %.2f B\n", __func__, ml.n_elements*1e-9);
// general kv
@ -3710,6 +3724,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
// copy the KV pairs from the input file
gguf_set_kv (ctx_out, model_loader->ctx_gguf);
gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
gguf_set_val_u32(ctx_out, "general.file_type", ftype);
#ifdef GGML_USE_K_QUANTS
int n_attention_wv = 0;
@ -3809,24 +3824,40 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
new_type = GGML_TYPE_Q6_K;
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
(i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
++i_attention_wv;
} else if (name.find("ffn_down.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
//else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < n_feed_forward_w2/8) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_feed_forward_w2 < 4) new_type = GGML_TYPE_Q5_K;
++i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_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;
}
// 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 {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
//}
bool convert_incompatible_tensor = false;
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
@ -4557,7 +4588,7 @@ int llama_model_n_embd(const struct llama_model * model) {
}
int llama_model_type(const struct llama_model * model, char * buf, size_t buf_size) {
return snprintf(buf, buf_size, "LLaMA %s %s", llama_model_type_name(model->type), llama_model_ftype_name(model->ftype));
return snprintf(buf, buf_size, "LLaMA %s %s", llama_model_type_name(model->type), llama_model_ftype_name(model->ftype).c_str());
}
int llama_model_quantize(

2
llama.h
View file

@ -103,6 +103,8 @@ extern "C" {
LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,// except 1d tensors
LLAMA_FTYPE_MOSTLY_Q6_K = 18,// except 1d tensors
LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
};
typedef struct llama_token_data {