2023-03-25 19:26:40 +01:00
# include "common.h"
# include "llama.h"
2023-03-28 18:48:20 +02:00
# include <cmath>
2023-08-28 17:59:39 +02:00
# include <cstdio>
# include <cstring>
2023-04-16 12:13:00 +02:00
# include <ctime>
2023-07-22 14:21:24 +02:00
# include <sstream>
2023-08-25 18:05:02 +02:00
# include <thread>
# include <mutex>
2024-01-18 18:18:21 +01:00
# include <atomic>
2023-08-28 17:59:39 +02:00
# include <vector>
2024-01-18 12:46:27 +01:00
# include <array>
# include <fstream>
# include <sstream>
2023-03-28 18:48:20 +02:00
2023-06-16 20:23:53 +02:00
# if defined(_MSC_VER)
# pragma warning(disable: 4244 4267) // possible loss of data
# endif
2023-08-28 17:59:39 +02:00
struct results_perplexity {
std : : vector < llama_token > tokens ;
double ppl_value ;
std : : vector < float > logits ;
std : : vector < float > probs ;
} ;
struct results_log_softmax {
double log_softmax ;
float logit ;
float prob ;
} ;
2023-09-15 21:38:27 +02:00
static void write_logfile (
const llama_context * ctx , const gpt_params & params , const llama_model * model ,
const struct results_perplexity & results
) {
2023-08-28 17:59:39 +02:00
if ( params . logdir . empty ( ) ) {
return ;
}
if ( params . hellaswag ) {
fprintf ( stderr , " %s: warning: logging results is not implemented for HellaSwag. No files will be written. \n " , __func__ ) ;
return ;
}
const std : : string timestamp = get_sortable_timestamp ( ) ;
const bool success = create_directory_with_parents ( params . logdir ) ;
if ( ! success ) {
fprintf ( stderr , " %s: warning: failed to create logdir %s, cannot write logfile \n " ,
__func__ , params . logdir . c_str ( ) ) ;
return ;
}
const std : : string logfile_path = params . logdir + timestamp + " .yml " ;
FILE * logfile = fopen ( logfile_path . c_str ( ) , " w " ) ;
if ( logfile = = NULL ) {
fprintf ( stderr , " %s: failed to open logfile %s \n " , __func__ , logfile_path . c_str ( ) ) ;
return ;
}
fprintf ( logfile , " binary: main \n " ) ;
char model_desc [ 128 ] ;
llama_model_desc ( model , model_desc , sizeof ( model_desc ) ) ;
dump_non_result_info_yaml ( logfile , params , ctx , timestamp , results . tokens , model_desc ) ;
fprintf ( logfile , " \n " ) ;
fprintf ( logfile , " ###################### \n " ) ;
fprintf ( logfile , " # Perplexity Results # \n " ) ;
fprintf ( logfile , " ###################### \n " ) ;
fprintf ( logfile , " \n " ) ;
dump_vector_float_yaml ( logfile , " logits " , results . logits ) ;
fprintf ( logfile , " ppl_value: %f \n " , results . ppl_value ) ;
dump_vector_float_yaml ( logfile , " probs " , results . probs ) ;
llama_dump_timing_info_yaml ( logfile , ctx ) ;
fclose ( logfile ) ;
}
2023-09-15 21:38:27 +02:00
static std : : vector < float > softmax ( const std : : vector < float > & logits ) {
2023-03-28 18:48:20 +02:00
std : : vector < float > probs ( logits . size ( ) ) ;
2023-03-25 19:26:40 +01:00
float max_logit = logits [ 0 ] ;
2023-09-28 18:04:36 +02:00
for ( float v : logits ) {
max_logit = std : : max ( max_logit , v ) ;
}
2023-03-25 19:26:40 +01:00
double sum_exp = 0.0 ;
for ( size_t i = 0 ; i < logits . size ( ) ; i + + ) {
// Subtract the maximum logit value from the current logit value for numerical stability
2023-03-28 18:48:20 +02:00
const float logit = logits [ i ] - max_logit ;
const float exp_logit = expf ( logit ) ;
2023-03-25 19:26:40 +01:00
sum_exp + = exp_logit ;
probs [ i ] = exp_logit ;
}
2023-09-28 18:04:36 +02:00
for ( size_t i = 0 ; i < probs . size ( ) ; i + + ) {
probs [ i ] / = sum_exp ;
}
2023-03-25 19:26:40 +01:00
return probs ;
}
2023-09-15 21:38:27 +02:00
static results_log_softmax log_softmax ( int n_vocab , const float * logits , int tok ) {
2023-08-25 18:05:02 +02:00
float max_logit = logits [ 0 ] ;
2023-09-28 18:04:36 +02:00
for ( int i = 1 ; i < n_vocab ; + + i ) {
max_logit = std : : max ( max_logit , logits [ i ] ) ;
}
2023-08-25 18:05:02 +02:00
double sum_exp = 0.0 ;
2023-09-28 18:04:36 +02:00
for ( int i = 0 ; i < n_vocab ; + + i ) {
sum_exp + = expf ( logits [ i ] - max_logit ) ;
}
2023-08-28 17:59:39 +02:00
return { logits [ tok ] - max_logit - log ( sum_exp ) , logits [ tok ] , expf ( logits [ tok ] - max_logit ) / ( float ) sum_exp } ;
2023-08-25 18:05:02 +02:00
}
2024-01-22 15:10:14 +01:00
static inline int nearest_int ( float fval ) {
//assert(fval <= 4194303.f);
float val = fval + 12582912.f ;
int i ; memcpy ( & i , & val , sizeof ( int ) ) ;
return ( i & 0x007fffff ) - 0x00400000 ;
}
static double log_softmax ( int n_vocab , const float * logits , uint16_t * log_prob , int tok ) {
float max_logit = logits [ 0 ] ;
float min_logit = logits [ 0 ] ;
for ( int i = 1 ; i < n_vocab ; + + i ) {
max_logit = std : : max ( max_logit , logits [ i ] ) ;
min_logit = std : : min ( min_logit , logits [ i ] ) ;
}
min_logit = std : : max ( min_logit , max_logit - 16 ) ;
double sum_exp = 0.0 ;
for ( int i = 0 ; i < n_vocab ; + + i ) {
sum_exp + = expf ( logits [ i ] - max_logit ) ;
}
const float log_sum_exp = log ( sum_exp ) ;
const float min_log_prob = min_logit - max_logit - log_sum_exp ;
const float scale = ( max_logit - min_logit ) / 65535.f ;
float * d = ( float * ) log_prob ;
d [ 0 ] = scale ;
d [ 1 ] = min_log_prob ;
log_prob + = 4 ;
if ( scale ) {
const float inv_scale = 1 / scale ;
for ( int i = 0 ; i < n_vocab ; + + i ) {
log_prob [ i ] = logits [ i ] > min_logit ? nearest_int ( inv_scale * ( logits [ i ] - min_logit ) ) : 0 ;
}
} else {
std : : memset ( log_prob , 0 , n_vocab * sizeof ( uint16_t ) ) ;
}
return max_logit + log_sum_exp - logits [ tok ] ;
}
2023-09-15 21:38:27 +02:00
static void process_logits (
int n_vocab , const float * logits , const int * tokens , int n_token , std : : vector < std : : thread > & workers ,
double & nll , double & nll2 , float * logit_history , float * prob_history
) {
2023-08-25 18:05:02 +02:00
std : : mutex mutex ;
int counter = 0 ;
2023-08-28 17:59:39 +02:00
auto compute = [ & mutex , & counter , & nll , & nll2 , logit_history , prob_history , n_vocab , logits , tokens , n_token ] ( ) {
2023-09-28 18:04:36 +02:00
double local_nll = 0 ;
double local_nll2 = 0 ;
2023-08-25 18:05:02 +02:00
while ( true ) {
std : : unique_lock < std : : mutex > lock ( mutex ) ;
int i = counter + + ;
if ( i > = n_token ) {
nll + = local_nll ; nll2 + = local_nll2 ;
break ;
}
lock . unlock ( ) ;
2023-08-28 17:59:39 +02:00
const results_log_softmax results = log_softmax ( n_vocab , logits + i * n_vocab , tokens [ i + 1 ] ) ;
const double v = - results . log_softmax ;
2023-08-25 18:05:02 +02:00
local_nll + = v ;
local_nll2 + = v * v ;
2023-08-28 17:59:39 +02:00
logit_history [ i ] = results . logit ;
prob_history [ i ] = results . prob ;
2023-08-25 18:05:02 +02:00
}
} ;
2023-09-28 18:04:36 +02:00
for ( auto & w : workers ) {
w = std : : thread ( compute ) ;
}
2023-08-25 18:05:02 +02:00
compute ( ) ;
2023-09-28 18:04:36 +02:00
for ( auto & w : workers ) {
w . join ( ) ;
}
2023-08-25 18:05:02 +02:00
}
2024-01-22 15:10:14 +01:00
static void process_logits ( std : : ostream & out , int n_vocab , const float * logits , const int * tokens , int n_token ,
std : : vector < std : : thread > & workers , std : : vector < uint16_t > & log_probs , double & nll , double & nll2 ) {
std : : mutex mutex ;
const int nv = 2 * ( ( n_vocab + 1 ) / 2 ) + 4 ;
int counter = 0 ;
auto compute = [ & mutex , & counter , & log_probs , & nll , & nll2 , n_vocab , logits , tokens , n_token , nv ] ( ) {
double local_nll = 0 ;
double local_nll2 = 0 ;
while ( true ) {
std : : unique_lock < std : : mutex > lock ( mutex ) ;
int i = counter + + ;
if ( i > = n_token ) {
nll + = local_nll ; nll2 + = local_nll2 ;
break ;
}
lock . unlock ( ) ;
const double v = log_softmax ( n_vocab , logits + i * n_vocab , log_probs . data ( ) + i * nv , tokens [ i + 1 ] ) ;
local_nll + = v ;
local_nll2 + = v * v ;
}
} ;
for ( auto & w : workers ) {
w = std : : thread ( compute ) ;
}
compute ( ) ;
for ( auto & w : workers ) {
w . join ( ) ;
}
out . write ( ( const char * ) log_probs . data ( ) , n_token * nv * sizeof ( uint16_t ) ) ;
}
struct kl_divergence_result {
double sum_nll = 0 ;
double sum_nll2 = 0 ;
double sum_kld = 0 ;
double sum_kld2 = 0 ;
double sum_nll_diff = 0 ;
double sum_nll_diff2 = 0 ;
2024-01-23 14:17:20 +01:00
size_t n_same_top = 0 ;
2024-01-22 15:10:14 +01:00
size_t count = 0 ;
} ;
2024-01-23 14:17:20 +01:00
static double log_softmax ( int n_vocab , const float * logits , const uint16_t * base_log_prob , int tok , kl_divergence_result & kld ) {
2024-01-22 15:10:14 +01:00
float max_logit = logits [ 0 ] ;
2024-01-23 14:17:20 +01:00
int imax = 0 ;
2024-01-22 15:10:14 +01:00
for ( int i = 1 ; i < n_vocab ; + + i ) {
2024-01-23 14:17:20 +01:00
if ( logits [ i ] > max_logit ) {
max_logit = logits [ i ] ;
imax = i ;
}
2024-01-22 15:10:14 +01:00
}
double sum_exp = 0.0 ;
for ( int i = 0 ; i < n_vocab ; + + i ) {
sum_exp + = expf ( logits [ i ] - max_logit ) ;
}
const float log_sum_exp = log ( sum_exp ) ;
const float * d = ( const float * ) base_log_prob ;
const float scale = d [ 0 ] ;
const float min_log_prob = d [ 1 ] ;
base_log_prob + = 4 ;
float nll = max_logit + log_sum_exp - logits [ tok ] ;
kld . sum_nll + = nll ;
kld . sum_nll2 + = nll * nll ;
nll + = ( scale * base_log_prob [ tok ] + min_log_prob ) ;
kld . sum_nll_diff + = nll ;
kld . sum_nll_diff2 + = nll * nll ;
max_logit + = log_sum_exp ;
double sum = 0 ;
2024-01-23 14:17:20 +01:00
int imax_base = - 1 ;
float p_log_base_max = 0 ;
2024-01-22 15:10:14 +01:00
for ( int i = 0 ; i < n_vocab ; + + i ) {
const float p_log_base = scale * base_log_prob [ i ] + min_log_prob ;
2024-01-23 14:17:20 +01:00
if ( i = = 0 | | p_log_base > p_log_base_max ) {
p_log_base_max = p_log_base ;
imax_base = i ;
}
2024-01-22 15:10:14 +01:00
if ( p_log_base > - 16.f ) {
const float p_base = expf ( p_log_base ) ;
sum + = p_base * ( p_log_base - logits [ i ] + max_logit ) ;
}
}
kld . sum_kld + = sum ;
kld . sum_kld2 + = sum * sum ;
+ + kld . count ;
2024-01-23 14:17:20 +01:00
if ( imax = = imax_base ) + + kld . n_same_top ;
return sum ;
2024-01-22 15:10:14 +01:00
}
static void process_logits ( int n_vocab , const float * logits , const int * tokens , int n_token ,
2024-01-23 14:17:20 +01:00
std : : vector < std : : thread > & workers , const std : : vector < uint16_t > & base_log_probs , kl_divergence_result & kld ,
float * kld_values ) {
2024-01-22 15:10:14 +01:00
std : : mutex mutex ;
const int nv = 2 * ( ( n_vocab + 1 ) / 2 ) + 4 ;
int counter = 0 ;
2024-01-23 14:17:20 +01:00
auto compute = [ & mutex , & counter , & base_log_probs , & kld , n_vocab , logits , tokens , n_token , nv , kld_values ] ( ) {
2024-01-22 15:10:14 +01:00
kl_divergence_result local_kld ;
while ( true ) {
std : : unique_lock < std : : mutex > lock ( mutex ) ;
int i = counter + + ;
if ( i > = n_token ) {
kld . sum_nll + = local_kld . sum_nll ;
kld . sum_nll2 + = local_kld . sum_nll2 ;
kld . sum_kld + = local_kld . sum_kld ;
kld . sum_kld2 + = local_kld . sum_kld2 ;
kld . sum_nll_diff + = local_kld . sum_nll_diff ;
kld . sum_nll_diff2 + = local_kld . sum_nll_diff2 ;
2024-01-23 14:17:20 +01:00
kld . n_same_top + = local_kld . n_same_top ;
2024-01-22 15:10:14 +01:00
kld . count + = local_kld . count ;
break ;
}
lock . unlock ( ) ;
2024-01-23 14:17:20 +01:00
double v = log_softmax ( n_vocab , logits + i * n_vocab , base_log_probs . data ( ) + i * nv , tokens [ i + 1 ] , local_kld ) ;
kld_values [ i ] = ( float ) v ;
2024-01-22 15:10:14 +01:00
}
} ;
for ( auto & w : workers ) {
w = std : : thread ( compute ) ;
}
compute ( ) ;
for ( auto & w : workers ) {
w . join ( ) ;
}
}
2023-09-15 21:38:27 +02:00
static results_perplexity perplexity_v2 ( llama_context * ctx , const gpt_params & params ) {
2024-02-18 21:39:30 +01:00
// Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
2023-08-23 11:56:42 +02:00
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
// Output: `perplexity: 13.5106 [114/114]`
// BOS tokens will be added for each chunk before eval
2023-11-17 03:14:37 +01:00
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
2023-08-23 22:08:04 +02:00
fprintf ( stderr , " %s: tokenizing the input .. \n " , __func__ ) ;
2023-08-28 17:59:39 +02:00
std : : vector < llama_token > tokens = : : llama_tokenize ( ctx , params . prompt , add_bos ) ;
2023-08-29 22:55:45 +02:00
2023-09-28 21:42:38 +02:00
const int n_ctx = llama_n_ctx ( ctx ) ;
if ( int ( tokens . size ( ) ) < 2 * n_ctx ) {
fprintf ( stderr , " %s: you need at least %d tokens to evaluate perplexity with a context of %d \n " , __func__ , 2 * n_ctx ,
n_ctx ) ;
2023-08-29 22:55:45 +02:00
fprintf ( stderr , " %s: the data file you provided tokenizes to only %zu tokens \n " , __func__ , tokens . size ( ) ) ;
return { std : : move ( tokens ) , 0. , { } , { } } ;
}
2023-09-28 18:04:36 +02:00
std : : vector < float > logit_history ;
std : : vector < float > prob_history ;
2023-08-28 17:59:39 +02:00
logit_history . resize ( tokens . size ( ) ) ;
prob_history . resize ( tokens . size ( ) ) ;
if ( params . ppl_stride < = 0 ) {
fprintf ( stderr , " %s: stride is %d but must be greater than zero! \n " , __func__ , params . ppl_stride ) ;
return { tokens , - 1 , logit_history , prob_history } ;
}
2023-08-23 11:56:42 +02:00
2023-09-28 21:42:38 +02:00
const int calc_chunk = n_ctx ;
2023-08-23 11:56:42 +02:00
fprintf ( stderr , " %s: have %zu tokens. Calculation chunk = %d \n " , __func__ , tokens . size ( ) , calc_chunk ) ;
if ( int ( tokens . size ( ) ) < = calc_chunk ) {
fprintf ( stderr , " %s: there are only %zu tokens, this is not enough for a context size of %d and stride %d \n " , __func__ ,
2023-09-28 21:42:38 +02:00
tokens . size ( ) , n_ctx , params . ppl_stride ) ;
2023-08-28 17:59:39 +02:00
return { tokens , - 1 , logit_history , prob_history } ;
2023-08-23 11:56:42 +02:00
}
const int n_chunk_max = ( tokens . size ( ) - calc_chunk + params . ppl_stride - 1 ) / params . ppl_stride ;
const int n_chunk = params . n_chunks < 0 ? n_chunk_max : std : : min ( params . n_chunks , n_chunk_max ) ;
2023-09-28 21:42:38 +02:00
const int n_vocab = llama_n_vocab ( llama_get_model ( ctx ) ) ;
2023-08-23 11:56:42 +02:00
const int n_batch = params . n_batch ;
int count = 0 ;
double nll = 0.0 ;
fprintf ( stderr , " %s: calculating perplexity over %d chunks, batch_size=%d \n " , __func__ , n_chunk , n_batch ) ;
for ( int i = 0 ; i < n_chunk ; + + i ) {
const int start = i * params . ppl_stride ;
const int end = start + calc_chunk ;
const int num_batches = ( calc_chunk + n_batch - 1 ) / n_batch ;
//fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
std : : vector < float > logits ;
const auto t_start = std : : chrono : : high_resolution_clock : : now ( ) ;
2023-09-28 18:04:36 +02:00
// clear the KV cache
2023-10-29 18:31:40 +01:00
llama_kv_cache_clear ( ctx ) ;
2023-09-28 18:04:36 +02:00
2023-08-23 11:56:42 +02:00
for ( int j = 0 ; j < num_batches ; + + j ) {
const int batch_start = start + j * n_batch ;
const int batch_size = std : : min ( end - batch_start , n_batch ) ;
//fprintf(stderr, " Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
2023-09-28 21:42:38 +02:00
if ( llama_decode ( ctx , llama_batch_get_one ( tokens . data ( ) + batch_start , batch_size , j * n_batch , 0 ) ) ) {
2023-08-23 11:56:42 +02:00
//fprintf(stderr, "%s : failed to eval\n", __func__);
2023-08-28 17:59:39 +02:00
return { tokens , - 1 , logit_history , prob_history } ;
2023-08-23 11:56:42 +02:00
}
// save original token and restore it after eval
const auto token_org = tokens [ batch_start ] ;
// add BOS token for the first batch of each chunk
2023-08-23 22:08:04 +02:00
if ( add_bos & & j = = 0 ) {
2023-10-23 21:40:03 +02:00
tokens [ batch_start ] = llama_token_bos ( llama_get_model ( ctx ) ) ;
2023-08-23 11:56:42 +02:00
}
const auto batch_logits = llama_get_logits ( ctx ) ;
logits . insert ( logits . end ( ) , batch_logits , batch_logits + batch_size * n_vocab ) ;
if ( j = = 0 ) {
tokens [ batch_start ] = token_org ;
}
}
const auto t_end = std : : chrono : : high_resolution_clock : : now ( ) ;
if ( i = = 0 ) {
const float t_total = std : : chrono : : duration < float > ( t_end - t_start ) . count ( ) ;
fprintf ( stderr , " %s: %.2f seconds per pass - ETA " , __func__ , t_total ) ;
int total_seconds = ( int ) ( t_total * n_chunk ) ;
if ( total_seconds > = 60 * 60 ) {
fprintf ( stderr , " %d hours " , total_seconds / ( 60 * 60 ) ) ;
total_seconds = total_seconds % ( 60 * 60 ) ;
}
fprintf ( stderr , " %.2f minutes \n " , total_seconds / 60.0 ) ;
}
//fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
2023-09-28 21:42:38 +02:00
for ( int j = n_ctx - params . ppl_stride - 1 ; j < n_ctx - 1 ; + + j ) {
2023-08-23 11:56:42 +02:00
// Calculate probability of next token, given the previous ones.
const std : : vector < float > tok_logits (
logits . begin ( ) + ( j + 0 ) * n_vocab ,
logits . begin ( ) + ( j + 1 ) * n_vocab ) ;
const float prob = softmax ( tok_logits ) [ tokens [ start + j + 1 ] ] ;
2023-08-28 17:59:39 +02:00
logit_history [ start + j + 1 ] = tok_logits [ tokens [ start + j + 1 ] ] ;
prob_history [ start + j + 1 ] = prob ;
2023-08-23 11:56:42 +02:00
nll + = - std : : log ( prob ) ;
+ + count ;
}
// perplexity is e^(average negative log-likelihood)
if ( params . ppl_output_type = = 0 ) {
printf ( " [%d]%.4lf, " , i + 1 , std : : exp ( nll / count ) ) ;
} else {
printf ( " %8d %.4lf \n " , i * params . ppl_stride , std : : exp ( nll / count ) ) ;
}
fflush ( stdout ) ;
}
printf ( " \n " ) ;
2023-08-28 17:59:39 +02:00
return { tokens , std : : exp ( nll / count ) , logit_history , prob_history } ;
2023-08-23 11:56:42 +02:00
}
2024-03-09 19:55:54 +01:00
static results_perplexity perplexity ( llama_context * ctx , const gpt_params & params , const int32_t n_ctx ) {
2023-08-23 11:56:42 +02:00
if ( params . ppl_stride > 0 ) {
2023-08-28 17:59:39 +02:00
return perplexity_v2 ( ctx , params ) ;
2023-08-23 11:56:42 +02:00
}
2024-02-18 21:39:30 +01:00
// Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
2023-03-26 15:14:01 +02:00
// Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
2023-03-25 19:26:40 +01:00
// Output: `perplexity: 13.5106 [114/114]`
2023-05-08 16:41:54 +02:00
// BOS tokens will be added for each chunk before eval
2023-08-23 22:08:04 +02:00
2023-11-17 03:14:37 +01:00
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
2023-08-23 22:08:04 +02:00
2024-01-22 15:10:14 +01:00
std : : ofstream logits_stream ;
if ( ! params . logits_file . empty ( ) ) {
2024-02-02 15:15:30 +01:00
logits_stream . open ( params . logits_file . c_str ( ) , std : : ios : : binary ) ;
2024-01-22 15:10:14 +01:00
if ( ! logits_stream . is_open ( ) ) {
fprintf ( stderr , " %s: failed to open %s for writing \n " , __func__ , params . logits_file . c_str ( ) ) ;
return { } ;
}
fprintf ( stderr , " %s: saving all logits to %s \n " , __func__ , params . logits_file . c_str ( ) ) ;
logits_stream . write ( " _logits_ " , 8 ) ;
2024-02-02 15:15:30 +01:00
logits_stream . write ( reinterpret_cast < const char * > ( & n_ctx ) , sizeof ( n_ctx ) ) ;
2024-01-22 15:10:14 +01:00
}
2023-08-27 15:50:33 +02:00
auto tim1 = std : : chrono : : high_resolution_clock : : now ( ) ;
2023-08-23 22:08:04 +02:00
fprintf ( stderr , " %s: tokenizing the input .. \n " , __func__ ) ;
2023-08-28 17:59:39 +02:00
std : : vector < llama_token > tokens = : : llama_tokenize ( ctx , params . prompt , add_bos ) ;
2023-03-25 19:26:40 +01:00
2023-08-27 15:50:33 +02:00
auto tim2 = std : : chrono : : high_resolution_clock : : now ( ) ;
fprintf ( stderr , " %s: tokenization took %g ms \n " , __func__ , 1e-3 * std : : chrono : : duration_cast < std : : chrono : : microseconds > ( tim2 - tim1 ) . count ( ) ) ;
2023-09-28 21:42:38 +02:00
if ( int ( tokens . size ( ) ) < 2 * n_ctx ) {
fprintf ( stderr , " %s: you need at least %d tokens to evaluate perplexity with a context of %d \n " , __func__ , 2 * n_ctx ,
n_ctx ) ;
2023-08-29 22:55:45 +02:00
fprintf ( stderr , " %s: the data file you provided tokenizes to only %zu tokens \n " , __func__ , tokens . size ( ) ) ;
return { std : : move ( tokens ) , 0. , { } , { } } ;
}
2023-08-28 17:59:39 +02:00
std : : vector < float > logit_history ;
logit_history . resize ( tokens . size ( ) ) ;
std : : vector < float > prob_history ;
prob_history . resize ( tokens . size ( ) ) ;
2023-09-28 21:42:38 +02:00
const int n_chunk_max = tokens . size ( ) / n_ctx ;
2023-05-08 16:41:54 +02:00
2023-07-18 13:24:43 +02:00
const int n_chunk = params . n_chunks < 0 ? n_chunk_max : std : : min ( params . n_chunks , n_chunk_max ) ;
2023-09-28 21:42:38 +02:00
const int n_vocab = llama_n_vocab ( llama_get_model ( ctx ) ) ;
2023-05-08 16:41:54 +02:00
const int n_batch = params . n_batch ;
2023-03-25 19:26:40 +01:00
2023-07-18 13:24:43 +02:00
int count = 0 ;
2023-03-28 18:48:20 +02:00
double nll = 0.0 ;
2023-08-25 18:05:02 +02:00
double nll2 = 0.0 ;
2023-07-18 13:24:43 +02:00
2024-01-19 10:02:39 +01:00
const int num_batches = ( n_ctx + n_batch - 1 ) / n_batch ;
2024-03-09 19:55:54 +01:00
const int n_seq = std : : max ( 1 , n_batch / n_ctx ) ;
GGML_ASSERT ( n_batch < n_ctx | | n_batch % n_ctx = = 0 ) ;
GGML_ASSERT ( params . n_ctx = = n_seq * n_ctx ) ;
llama_batch batch = llama_batch_init ( std : : min ( n_batch , n_ctx * n_seq ) , 0 , 1 ) ;
2024-01-19 10:02:39 +01:00
std : : vector < float > logits ;
if ( num_batches > 1 ) {
logits . reserve ( ( size_t ) n_ctx * n_vocab ) ;
}
2024-03-09 19:55:54 +01:00
fprintf ( stderr , " %s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d \n " , __func__ , n_chunk , n_ctx , n_batch , n_seq ) ;
2023-05-08 16:41:54 +02:00
2023-08-25 18:05:02 +02:00
std : : vector < std : : thread > workers ( std : : thread : : hardware_concurrency ( ) - 1 ) ;
2024-01-22 15:10:14 +01:00
std : : vector < uint16_t > log_probs ;
if ( ! params . logits_file . empty ( ) ) {
logits_stream . write ( ( const char * ) & n_vocab , sizeof ( n_vocab ) ) ;
logits_stream . write ( ( const char * ) & n_chunk , sizeof ( n_chunk ) ) ;
logits_stream . write ( ( const char * ) tokens . data ( ) , n_chunk * n_ctx * sizeof ( tokens [ 0 ] ) ) ;
const int nv = 2 * ( ( n_vocab + 1 ) / 2 ) + 4 ;
log_probs . resize ( n_ctx * nv ) ;
}
2024-03-09 19:55:54 +01:00
// We get the logits for all the tokens in the context window (params.n_ctx)
// from llama_eval above. Now, based on https://huggingface.co/docs/transformers/perplexity,
// calculate the perplexity over the last half of the window (so the model always has
// some context to predict the token).
//
// We rely on the fact that attention in the forward pass only looks at previous
// tokens here, so the logits returned for each token are an accurate representation
// of what the model would have predicted at that point.
//
// Example, we have a context window of 512, we will compute perplexity for each of the
// last 256 tokens. Then, we split the input up into context window size chunks to
// process the entire prompt.
const int first = n_ctx / 2 ;
for ( int i = 0 ; i < n_chunk ; i + = n_seq ) {
2023-09-28 21:42:38 +02:00
const int start = i * n_ctx ;
const int end = start + n_ctx ;
2023-03-25 19:26:40 +01:00
2024-03-09 19:55:54 +01:00
const int n_seq_batch = std : : min ( n_seq , n_chunk - i ) ;
2023-05-08 16:41:54 +02:00
const auto t_start = std : : chrono : : high_resolution_clock : : now ( ) ;
2023-09-28 18:04:36 +02:00
// clear the KV cache
2023-10-29 18:31:40 +01:00
llama_kv_cache_clear ( ctx ) ;
2023-09-28 18:04:36 +02:00
2023-04-13 23:50:42 +02:00
for ( int j = 0 ; j < num_batches ; + + j ) {
2023-05-08 16:41:54 +02:00
const int batch_start = start + j * n_batch ;
const int batch_size = std : : min ( end - batch_start , n_batch ) ;
2024-03-09 19:55:54 +01:00
batch . n_tokens = 0 ;
for ( int seq = 0 ; seq < n_seq_batch ; seq + + ) {
int seq_start = batch_start + seq * n_ctx ;
2023-05-08 16:41:54 +02:00
2024-03-09 19:55:54 +01:00
// save original token and restore it after eval
const auto token_org = tokens [ seq_start ] ;
// add BOS token for the first batch of each chunk
if ( add_bos & & j = = 0 ) {
tokens [ seq_start ] = llama_token_bos ( llama_get_model ( ctx ) ) ;
}
for ( int k = 0 ; k < batch_size ; + + k ) {
const int idx = seq * n_ctx + k ;
batch . token [ idx ] = tokens [ seq_start + k ] ;
batch . pos [ idx ] = j * n_batch + k ;
batch . n_seq_id [ idx ] = 1 ;
batch . seq_id [ idx ] [ 0 ] = seq ;
batch . logits [ idx ] = batch . pos [ idx ] > = first ? 1 : 0 ;
}
batch . n_tokens + = batch_size ;
// restore the original token in case it was set to BOS
tokens [ seq_start ] = token_org ;
2023-05-08 16:41:54 +02:00
}
2024-03-09 19:55:54 +01:00
if ( llama_decode ( ctx , batch ) ) {
2023-04-13 23:50:42 +02:00
fprintf ( stderr , " %s : failed to eval \n " , __func__ ) ;
2023-08-28 17:59:39 +02:00
return { tokens , - 1 , logit_history , prob_history } ;
2023-04-13 23:50:42 +02:00
}
2023-05-08 16:41:54 +02:00
2024-01-19 10:02:39 +01:00
if ( num_batches > 1 ) {
const auto * batch_logits = llama_get_logits ( ctx ) ;
logits . insert ( logits . end ( ) , batch_logits , batch_logits + batch_size * n_vocab ) ;
}
2023-03-25 19:26:40 +01:00
}
2023-05-08 16:41:54 +02:00
2023-03-25 19:26:40 +01:00
if ( i = = 0 ) {
2024-03-13 18:54:21 +01:00
llama_synchronize ( ctx ) ;
const auto t_end = std : : chrono : : high_resolution_clock : : now ( ) ;
2023-05-08 16:41:54 +02:00
const float t_total = std : : chrono : : duration < float > ( t_end - t_start ) . count ( ) ;
fprintf ( stderr , " %s: %.2f seconds per pass - ETA " , __func__ , t_total ) ;
2024-03-09 19:55:54 +01:00
int total_seconds = ( int ) ( t_total * n_chunk / n_seq ) ;
2023-04-21 14:57:57 +02:00
if ( total_seconds > = 60 * 60 ) {
2023-05-08 16:41:54 +02:00
fprintf ( stderr , " %d hours " , total_seconds / ( 60 * 60 ) ) ;
2023-04-21 14:57:57 +02:00
total_seconds = total_seconds % ( 60 * 60 ) ;
}
2023-08-18 11:48:55 +02:00
fprintf ( stderr , " %.2f minutes \n " , total_seconds / 60.0 ) ;
2023-03-25 19:26:40 +01:00
}
2023-05-08 16:41:54 +02:00
2024-03-09 19:55:54 +01:00
for ( int seq = 0 ; seq < n_seq_batch ; seq + + ) {
const float * all_logits = num_batches > 1 ? logits . data ( ) : llama_get_logits_ith ( ctx , seq * n_ctx ) ;
llama_token * tokens_data = tokens . data ( ) + start + seq * n_ctx + first ;
if ( ! params . logits_file . empty ( ) ) {
process_logits ( logits_stream , n_vocab , all_logits + first * n_vocab ,
tokens_data , n_ctx - 1 - first ,
workers , log_probs , nll , nll2 ) ;
} else {
process_logits ( n_vocab , all_logits + first * n_vocab ,
tokens_data , n_ctx - 1 - first ,
workers , nll , nll2 ,
logit_history . data ( ) + start + seq * n_ctx + first ,
prob_history . data ( ) + start + seq * n_ctx + first ) ;
}
count + = n_ctx - first - 1 ;
// perplexity is e^(average negative log-likelihood)
if ( params . ppl_output_type = = 0 ) {
printf ( " [%d]%.4lf, " , i + seq + 1 , std : : exp ( nll / count ) ) ;
} else {
double av = nll / count ;
double av2 = nll2 / count - av * av ;
if ( av2 > 0 ) av2 = sqrt ( av2 / ( count - 1 ) ) ;
printf ( " %8d %.4lf %4lf %4lf \n " , i * n_ctx , std : : exp ( nll / count ) , av , av2 ) ;
}
2023-08-23 11:56:42 +02:00
}
2023-03-25 19:26:40 +01:00
fflush ( stdout ) ;
2024-01-19 10:02:39 +01:00
logits . clear ( ) ;
2023-03-25 19:26:40 +01:00
}
printf ( " \n " ) ;
2023-08-28 17:59:39 +02:00
2023-08-25 18:05:02 +02:00
nll2 / = count ;
nll / = count ;
2023-08-28 17:59:39 +02:00
const double ppl = exp ( nll ) ;
2023-08-25 18:05:02 +02:00
nll2 - = nll * nll ;
if ( nll2 > 0 ) {
nll2 = sqrt ( nll2 / ( count - 1 ) ) ;
printf ( " Final estimate: PPL = %.4lf +/- %.5lf \n " , ppl , nll2 * ppl ) ;
} else {
printf ( " Unexpected negative standard deviation of log(prob) \n " ) ;
}
2023-08-28 17:59:39 +02:00
2024-03-09 19:55:54 +01:00
llama_batch_free ( batch ) ;
2023-08-28 17:59:39 +02:00
return { tokens , ppl , logit_history , prob_history } ;
2023-03-25 19:26:40 +01:00
}
2024-01-19 09:45:06 +01:00
static bool decode_helper ( llama_context * ctx , llama_batch & batch , std : : vector < float > & batch_logits , int32_t n_batch , int32_t n_vocab ) {
for ( int32_t i = 0 ; i < ( int32_t ) batch . n_tokens ; i + = n_batch ) {
const int32_t n_tokens = std : : min ( n_batch , ( int32_t ) ( batch . n_tokens - i ) ) ;
llama_batch batch_view = {
n_tokens ,
batch . token + i ,
nullptr ,
batch . pos + i ,
batch . n_seq_id + i ,
batch . seq_id + i ,
batch . logits + i ,
0 , 0 , 0 , // unused
} ;
const int ret = llama_decode ( ctx , batch_view ) ;
if ( ret ! = 0 ) {
LOG_TEE ( " failed to decode the batch, n_batch = %d, ret = %d \n " , n_batch , ret ) ;
return false ;
2023-08-21 10:11:31 +02:00
}
2024-01-19 09:45:06 +01:00
memcpy ( batch_logits . data ( ) + i * n_vocab , llama_get_logits ( ctx ) , n_tokens * n_vocab * sizeof ( float ) ) ;
2023-08-21 10:11:31 +02:00
}
2024-01-19 09:45:06 +01:00
return true ;
2023-08-21 10:11:31 +02:00
}
2024-01-20 16:08:08 +01:00
# define K_TOKEN_CHUNK 4
2024-01-19 10:39:11 +01:00
static void compute_logprobs ( const float * batch_logits , int n_vocab , std : : vector < std : : thread > & workers ,
2024-01-18 18:18:21 +01:00
const std : : vector < std : : pair < size_t , llama_token > > & eval_pairs , std : : vector < float > & eval_results ) {
if ( eval_results . size ( ) ! = eval_pairs . size ( ) ) {
eval_results . resize ( eval_pairs . size ( ) ) ;
}
if ( eval_pairs . empty ( ) ) return ;
2024-01-20 16:08:08 +01:00
size_t max_threads = std : : min ( ( eval_pairs . size ( ) + K_TOKEN_CHUNK - 1 ) / K_TOKEN_CHUNK , workers . size ( ) ) ;
2024-01-18 18:18:21 +01:00
std : : atomic < int > counter ( 0 ) ;
auto compute = [ & counter , & eval_pairs , & eval_results , batch_logits , n_vocab ] ( ) {
2024-01-20 16:08:08 +01:00
float local_logprobs [ K_TOKEN_CHUNK ] ;
2024-01-18 18:18:21 +01:00
while ( true ) {
2024-01-20 16:08:08 +01:00
size_t first = counter . fetch_add ( K_TOKEN_CHUNK , std : : memory_order_relaxed ) ;
2024-01-18 18:18:21 +01:00
if ( first > = eval_results . size ( ) ) break ;
2024-01-20 16:08:08 +01:00
size_t last = std : : min ( first + K_TOKEN_CHUNK , eval_results . size ( ) ) ;
2024-01-18 18:18:21 +01:00
for ( size_t i = first ; i < last ; + + i ) {
auto logits = batch_logits + eval_pairs [ i ] . first * n_vocab ;
float max_logit = logits [ 0 ] ;
for ( int j = 1 ; j < n_vocab ; + + j ) {
max_logit = std : : max ( max_logit , logits [ j ] ) ;
}
float sum_p = 0.f ;
for ( int j = 0 ; j < n_vocab ; + + j ) {
sum_p + = expf ( logits [ j ] - max_logit ) ;
}
local_logprobs [ i - first ] = logits [ eval_pairs [ i ] . second ] - max_logit - std : : log ( sum_p ) ;
}
std : : memcpy ( eval_results . data ( ) + first , local_logprobs , ( last - first ) * sizeof ( float ) ) ;
}
} ;
for ( size_t it = 0 ; it < max_threads ; + + it ) {
workers [ it ] = std : : thread ( compute ) ;
}
for ( size_t it = 0 ; it < max_threads ; + + it ) {
workers [ it ] . join ( ) ;
}
}
2023-09-15 21:38:27 +02:00
static void hellaswag_score ( llama_context * ctx , const gpt_params & params ) {
2023-07-28 20:25:36 +02:00
// Calculates hellaswag score (acc_norm) from prompt
//
// Data extracted from the HellaSwag validation dataset (MIT license) https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
// All used data fields are preprocessed as in https://github.com/EleutherAI/lm-evaluation-harness/blob/df3da98c5405deafd519c2ddca52bb7c3fe36bef/lm_eval/tasks/hellaswag.py#L62-L68
//
// All 10042 tasks should be extracted to keep the results standardized like other implementations.
//
// Datafile layout:
// ['??'] denotes json fields
// 6 lines per task:
// ['activity_label'] + ": " +['ctx'] - The first part of the query, the context
// ['label'] - The index the best common sense ending aka gold ending
// ['endings'][0] - Endings added to the first part of the query
// ['endings'][1]
// ['endings'][2]
// ['endings'][3]
2023-07-22 14:21:24 +02:00
std : : vector < std : : string > prompt_lines ;
std : : istringstream strstream ( params . prompt ) ;
std : : string line ;
while ( std : : getline ( strstream , line , ' \n ' ) ) {
prompt_lines . push_back ( line ) ;
}
2024-01-18 14:33:01 +01:00
if ( prompt_lines . size ( ) % 6 ! = 0 ) {
2023-07-28 20:25:36 +02:00
fprintf ( stderr , " %s : number of lines in prompt not a multiple of 6. \n " , __func__ ) ;
return ;
}
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
size_t hs_task_count = prompt_lines . size ( ) / 6 ;
2023-08-04 12:07:21 +02:00
fprintf ( stderr , " %s : loaded %zu tasks from prompt. \n " , __func__ , hs_task_count ) ;
2023-07-22 14:21:24 +02:00
2023-09-28 21:42:38 +02:00
const bool is_spm = llama_vocab_type ( llama_get_model ( ctx ) ) = = LLAMA_VOCAB_TYPE_SPM ;
2023-08-26 15:48:53 +02:00
fprintf ( stderr , " ================================= is_spm = %d \n " , is_spm ) ;
2023-08-23 22:08:04 +02:00
2023-07-28 20:25:36 +02:00
// This is needed as usual for LLaMA models
2023-11-17 03:14:37 +01:00
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
2023-07-28 20:25:36 +02:00
2024-01-21 13:42:44 +01:00
// The tasks should be randomized so the score stabilizes quickly.
bool randomize_tasks = true ;
2023-07-28 20:25:36 +02:00
// Number of tasks to use when computing the score
2024-01-18 14:33:01 +01:00
if ( params . hellaswag_tasks < hs_task_count ) {
2023-07-28 20:25:36 +02:00
hs_task_count = params . hellaswag_tasks ;
}
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
// The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
std : : mt19937 rng ( 1 ) ;
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
// Dataholder for hellaswag tasks
struct hs_data_t {
std : : string context ;
size_t gold_ending_idx ;
std : : string ending [ 4 ] ;
size_t ending_logprob_count [ 4 ] ;
double ending_logprob [ 4 ] ;
2024-01-18 14:33:01 +01:00
size_t i_batch ; // starting index in the llama_batch
size_t common_prefix ; // max number of initial tokens that are the same in all sentences
size_t required_tokens ; // needed number of tokens to evaluate all 4 endings
std : : vector < llama_token > seq_tokens [ 4 ] ;
2023-07-28 20:25:36 +02:00
} ;
2023-07-22 14:21:24 +02:00
2023-08-04 12:07:21 +02:00
fprintf ( stderr , " %s : selecting %zu %s tasks. \n " , __func__ , hs_task_count , ( randomize_tasks ? " randomized " : " the first " ) ) ;
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
// Select and read data from prompt lines
2024-01-18 14:33:01 +01:00
std : : vector < hs_data_t > hs_data ( hs_task_count ) ;
for ( size_t i = 0 ; i < hs_task_count ; i + + ) {
2023-07-28 20:25:36 +02:00
size_t idx = i ;
2023-07-22 14:21:24 +02:00
2024-01-18 14:33:01 +01:00
auto & hs_cur = hs_data [ i ] ;
2023-07-28 20:25:36 +02:00
// Select a random example of those left in the prompt
if ( randomize_tasks ) {
std : : uniform_int_distribution < size_t > dist ( 0 , prompt_lines . size ( ) / 6 - 1 ) ;
idx = dist ( rng ) ;
2023-07-22 14:21:24 +02:00
}
2024-01-18 14:33:01 +01:00
hs_cur . context = prompt_lines [ idx * 6 ] ;
hs_cur . gold_ending_idx = std : : stoi ( prompt_lines [ idx * 6 + 1 ] ) ;
for ( size_t j = 0 ; j < 4 ; j + + ) {
hs_cur . ending [ j ] = prompt_lines [ idx * 6 + 2 + j ] ;
hs_cur . seq_tokens [ j ] = : : llama_tokenize ( ctx , hs_cur . context + " " + hs_cur . ending [ j ] , add_bos ) ;
2023-07-22 14:21:24 +02:00
}
2024-01-18 14:33:01 +01:00
// determine the common prefix of the endings
hs_cur . common_prefix = 0 ;
for ( size_t k = 0 ; k < hs_cur . seq_tokens [ 0 ] . size ( ) ; k + + ) {
if ( hs_cur . seq_tokens [ 0 ] [ k ] ! = hs_cur . seq_tokens [ 1 ] [ k ] | |
hs_cur . seq_tokens [ 0 ] [ k ] ! = hs_cur . seq_tokens [ 2 ] [ k ] | |
hs_cur . seq_tokens [ 0 ] [ k ] ! = hs_cur . seq_tokens [ 3 ] [ k ] ) {
break ;
}
hs_cur . common_prefix + + ;
}
hs_cur . required_tokens = hs_cur . common_prefix +
hs_cur . seq_tokens [ 0 ] . size ( ) - hs_cur . common_prefix +
hs_cur . seq_tokens [ 1 ] . size ( ) - hs_cur . common_prefix +
hs_cur . seq_tokens [ 2 ] . size ( ) - hs_cur . common_prefix +
hs_cur . seq_tokens [ 3 ] . size ( ) - hs_cur . common_prefix ;
//GGML_ASSERT(hs_cur.common_prefix >= ::llama_tokenize(ctx, hs_cur.context, add_bos).size());
2023-07-28 20:25:36 +02:00
// Delete the selected random example from the prompt
if ( randomize_tasks ) {
prompt_lines . erase ( std : : next ( prompt_lines . begin ( ) , idx * 6 ) , std : : next ( prompt_lines . begin ( ) , idx * 6 + 6 ) ) ;
}
}
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
fprintf ( stderr , " %s : calculating hellaswag score over selected tasks. \n " , __func__ ) ;
2024-01-18 14:33:01 +01:00
2023-07-28 20:25:36 +02:00
printf ( " \n task \t acc_norm \n " ) ;
2023-07-22 14:21:24 +02:00
2023-07-28 20:25:36 +02:00
double acc = 0.0f ;
2024-01-18 14:33:01 +01:00
2023-09-28 21:42:38 +02:00
const int n_vocab = llama_n_vocab ( llama_get_model ( ctx ) ) ;
2024-01-18 14:33:01 +01:00
const int n_ctx = llama_n_ctx ( ctx ) ;
const int n_batch = params . n_batch ;
2023-07-28 20:25:36 +02:00
2024-01-19 09:45:06 +01:00
const int max_tasks_per_batch = 32 ;
2024-03-11 16:49:47 +01:00
const int max_seq = std : : min ( 4 * max_tasks_per_batch , ( int ) llama_n_seq_max ( ctx ) ) ;
2023-08-26 15:48:53 +02:00
2024-01-18 14:33:01 +01:00
llama_batch batch = llama_batch_init ( n_ctx , 0 , max_seq ) ;
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
std : : vector < float > tok_logits ( n_vocab ) ;
2024-01-19 09:45:06 +01:00
std : : vector < float > batch_logits ( n_vocab * n_ctx ) ;
2024-01-18 14:33:01 +01:00
2024-01-18 18:18:21 +01:00
std : : vector < std : : pair < size_t , llama_token > > eval_pairs ;
std : : vector < float > eval_results ;
std : : vector < std : : thread > workers ( std : : thread : : hardware_concurrency ( ) ) ;
2024-01-18 14:33:01 +01:00
for ( size_t i0 = 0 ; i0 < hs_task_count ; i0 + + ) {
int n_cur = 0 ;
2023-09-28 18:04:36 +02:00
2024-01-18 14:33:01 +01:00
size_t i1 = i0 ;
size_t i_batch = 0 ; // this tells us where in `llama_batch` we are currently
llama_batch_clear ( batch ) ;
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
// batch as much tasks as possible into the available context
// each task has 4 unique seuqnce ids - one for each ending
// the common prefix is shared among the 4 sequences to save tokens
// we extract logits only from the last common token and from all ending tokens of each sequence
while ( n_cur + ( int ) hs_data [ i1 ] . required_tokens < = n_ctx ) {
auto & hs_cur = hs_data [ i1 ] ;
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
const int s0 = 4 * ( i1 - i0 ) ;
if ( s0 + 4 > max_seq ) {
break ;
}
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
for ( size_t i = 0 ; i < hs_cur . common_prefix ; + + i ) {
llama_batch_add ( batch , hs_cur . seq_tokens [ 0 ] [ i ] , i , { s0 + 0 , s0 + 1 , s0 + 2 , s0 + 3 } , false ) ;
}
batch . logits [ batch . n_tokens - 1 ] = true ; // we need logits for the last token of the common prefix
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
for ( int s = 0 ; s < 4 ; + + s ) {
for ( size_t i = hs_cur . common_prefix ; i < hs_cur . seq_tokens [ s ] . size ( ) ; + + i ) {
llama_batch_add ( batch , hs_cur . seq_tokens [ s ] [ i ] , i , { s0 + s } , true ) ;
}
}
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
hs_cur . i_batch = i_batch ;
i_batch + = hs_cur . required_tokens ;
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
n_cur + = hs_data [ i1 ] . required_tokens ;
if ( + + i1 = = hs_task_count ) {
break ;
}
2023-08-20 15:44:46 +02:00
}
2024-01-18 14:33:01 +01:00
if ( i0 = = i1 ) {
fprintf ( stderr , " %s : task %zu does not fit in the context window \n " , __func__ , i0 ) ;
return ;
}
2023-08-20 15:44:46 +02:00
2024-01-18 14:33:01 +01:00
llama_kv_cache_clear ( ctx ) ;
2023-07-28 20:25:36 +02:00
2024-01-18 14:33:01 +01:00
// decode all tasks [i0, i1)
2024-01-19 09:45:06 +01:00
if ( ! decode_helper ( ctx , batch , batch_logits , n_batch , n_vocab ) ) {
2024-01-18 14:33:01 +01:00
fprintf ( stderr , " %s: llama_decode() failed \n " , __func__ ) ;
return ;
}
2023-07-28 20:25:36 +02:00
2024-01-18 18:18:21 +01:00
// Compute log-probs in parallel
// First we collect all tasks
eval_pairs . clear ( ) ;
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & hs_cur = hs_data [ i ] ;
size_t li = hs_cur . common_prefix ;
for ( int s = 0 ; s < 4 ; + + s ) {
for ( size_t j = hs_cur . common_prefix ; j < hs_cur . seq_tokens [ s ] . size ( ) - 1 ; j + + ) {
2024-02-03 12:23:37 +01:00
eval_pairs . emplace_back ( hs_cur . i_batch + li + + , hs_cur . seq_tokens [ s ] [ j + 1 ] ) ;
2024-01-18 18:18:21 +01:00
}
+ + li ;
}
}
// Then we do the actual calculation
2024-01-19 10:39:11 +01:00
compute_logprobs ( batch_logits . data ( ) , n_vocab , workers , eval_pairs , eval_results ) ;
2024-01-18 18:18:21 +01:00
size_t ir = 0 ;
2024-01-18 14:33:01 +01:00
// compute the logprobs for each ending of the decoded tasks
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & hs_cur = hs_data [ i ] ;
2023-07-22 14:21:24 +02:00
2024-01-18 14:33:01 +01:00
std : : memcpy ( tok_logits . data ( ) , batch_logits . data ( ) + n_vocab * ( hs_cur . i_batch + hs_cur . common_prefix - 1 ) , n_vocab * sizeof ( float ) ) ;
2023-07-28 20:25:36 +02:00
2024-01-18 14:33:01 +01:00
const auto first_probs = softmax ( tok_logits ) ;
2023-07-28 20:25:36 +02:00
2024-01-18 14:33:01 +01:00
for ( int s = 0 ; s < 4 ; + + s ) {
hs_cur . ending_logprob_count [ s ] = 1 ;
hs_cur . ending_logprob [ s ] = std : : log ( first_probs [ hs_cur . seq_tokens [ s ] [ hs_cur . common_prefix ] ] ) ;
for ( size_t j = hs_cur . common_prefix ; j < hs_cur . seq_tokens [ s ] . size ( ) - 1 ; j + + ) {
2024-01-18 18:18:21 +01:00
hs_cur . ending_logprob [ s ] + = eval_results [ ir + + ] ;
2024-01-18 14:33:01 +01:00
hs_cur . ending_logprob_count [ s ] + + ;
}
hs_cur . ending_logprob [ s ] / = hs_cur . ending_logprob_count [ s ] ;
}
2023-07-22 14:21:24 +02:00
2024-01-18 14:33:01 +01:00
// Find the ending with maximum logprob
size_t ending_logprob_max_idx = 0 ;
double ending_logprob_max_val = hs_cur . ending_logprob [ 0 ] ;
for ( size_t s = 1 ; s < 4 ; s + + ) {
if ( hs_cur . ending_logprob [ s ] > ending_logprob_max_val ) {
ending_logprob_max_idx = s ;
ending_logprob_max_val = hs_cur . ending_logprob [ s ] ;
}
2023-07-28 20:25:36 +02:00
}
2023-07-22 14:21:24 +02:00
2024-01-18 14:33:01 +01:00
//printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_cur.gold_ending_idx);
2023-07-22 14:21:24 +02:00
2024-01-18 14:33:01 +01:00
// If the gold ending got the maximum logprobe add one accuracy point
if ( ending_logprob_max_idx = = hs_cur . gold_ending_idx ) {
acc + = 1.0 ;
}
// Print the accumulated accuracy mean x 100
printf ( " %zu \t %.8lf \n " , i + 1 , acc / double ( i + 1 ) * 100.0 ) ;
fflush ( stdout ) ;
2023-07-28 20:25:36 +02:00
}
2024-01-18 14:33:01 +01:00
i0 = i1 - 1 ;
2023-07-22 14:21:24 +02:00
}
2024-01-18 14:33:01 +01:00
llama_batch_free ( batch ) ;
2023-07-28 20:25:36 +02:00
2023-07-22 14:21:24 +02:00
printf ( " \n " ) ;
}
2024-01-18 12:46:27 +01:00
struct winogrande_entry {
std : : string first ;
std : : string second ;
std : : array < std : : string , 2 > choices ;
int answer ;
2024-01-19 09:45:06 +01:00
size_t i_batch ;
size_t common_prefix ;
size_t required_tokens ;
size_t n_base1 ; // number of tokens for context + choice 1
size_t n_base2 ; // number of tokens for context + choice 2
std : : vector < llama_token > seq_tokens [ 2 ] ;
2024-01-18 12:46:27 +01:00
} ;
static std : : vector < winogrande_entry > load_winogrande_from_csv ( const std : : string & prompt ) {
std : : vector < winogrande_entry > result ;
std : : istringstream in ( prompt ) ;
std : : string line ;
std : : array < int , 4 > comma_pos ;
while ( true ) {
std : : getline ( in , line ) ;
if ( in . fail ( ) | | in . eof ( ) ) break ;
int ipos = 0 ;
bool quote_open = false ;
for ( int i = 0 ; i < int ( line . size ( ) ) ; + + i ) {
if ( ! quote_open ) {
if ( line [ i ] = = ' , ' ) {
comma_pos [ ipos + + ] = i ;
if ( ipos = = 4 ) break ;
}
else if ( line [ i ] = = ' " ' ) {
quote_open = true ;
}
}
else {
if ( line [ i ] = = ' " ' ) {
quote_open = false ;
}
}
}
if ( ipos ! = 4 ) {
printf ( " %s: failed to find comma separators in <%s> \n " , __func__ , line . c_str ( ) ) ;
continue ;
}
auto sentence = line [ comma_pos [ 0 ] + 1 ] = = ' " ' ? line . substr ( comma_pos [ 0 ] + 2 , comma_pos [ 1 ] - comma_pos [ 0 ] - 3 )
: line . substr ( comma_pos [ 0 ] + 1 , comma_pos [ 1 ] - comma_pos [ 0 ] - 1 ) ;
auto choice1 = line . substr ( comma_pos [ 1 ] + 1 , comma_pos [ 2 ] - comma_pos [ 1 ] - 1 ) ;
auto choice2 = line . substr ( comma_pos [ 2 ] + 1 , comma_pos [ 3 ] - comma_pos [ 2 ] - 1 ) ;
auto answer = line . substr ( comma_pos [ 3 ] + 1 , line . size ( ) - comma_pos [ 3 ] - 1 ) ;
auto index = line . substr ( 0 , comma_pos [ 0 ] ) ;
int where = 0 ;
for ( ; where < int ( sentence . size ( ) ) ; + + where ) {
if ( sentence [ where ] = = ' _ ' ) break ;
}
if ( where = = int ( sentence . size ( ) ) ) {
printf ( " %s: no _ in <%s> \n " , __func__ , sentence . c_str ( ) ) ;
continue ;
}
std : : istringstream stream ( answer . c_str ( ) ) ;
int i_answer ; stream > > i_answer ;
if ( stream . fail ( ) | | i_answer < 1 | | i_answer > 2 ) {
printf ( " %s: failed to parse answer <%s> \n " , __func__ , answer . c_str ( ) ) ;
continue ;
}
result . emplace_back ( ) ;
auto & wg = result . back ( ) ;
wg . first = sentence . substr ( 0 , where ) ;
wg . second = sentence . substr ( where + 1 , sentence . size ( ) - where - 1 ) ;
wg . choices [ 0 ] = std : : move ( choice1 ) ;
wg . choices [ 1 ] = std : : move ( choice2 ) ;
wg . answer = i_answer ;
}
return result ;
}
/*
* Evaluates the Winogrande score .
* Uses a CSV containing task index , dentence , choice 1 , choice 2 , answer ( 1 or 2 )
* You can get one such dataset from e . g . https : //huggingface.co/datasets/ikawrakow/winogrande-eval-for-llama.cpp
* As an example , the 1 st row in the above dataset is
*
* 0 , Sarah was a much better surgeon than Maria so _ always got the easier cases . , Sarah , Maria , 2
*
*/
static void winogrande_score ( llama_context * ctx , const gpt_params & params ) {
constexpr int k_min_trailing_ctx = 3 ;
auto data = load_winogrande_from_csv ( params . prompt ) ;
if ( data . empty ( ) ) {
fprintf ( stderr , " %s: no tasks \n " , __func__ ) ;
return ;
}
fprintf ( stderr , " %s : loaded %zu tasks from prompt. \n " , __func__ , data . size ( ) ) ;
if ( params . winogrande_tasks > 0 & & params . winogrande_tasks < data . size ( ) ) {
fprintf ( stderr , " %s : selecting %zu random tasks \n " , __func__ , params . winogrande_tasks ) ;
std : : mt19937 rng ( 1 ) ;
std : : vector < int > aux ( data . size ( ) ) ;
for ( int i = 0 ; i < int ( data . size ( ) ) ; + + i ) {
aux [ i ] = i ;
}
float scale = 1 / ( 1.f + ( float ) rng . max ( ) ) ;
std : : vector < winogrande_entry > selected ;
2024-01-18 19:49:00 +01:00
selected . resize ( params . winogrande_tasks ) ;
2024-01-18 12:46:27 +01:00
for ( int i = 0 ; i < int ( params . winogrande_tasks ) ; + + i ) {
int j = int ( scale * rng ( ) * aux . size ( ) ) ;
selected [ i ] = std : : move ( data [ aux [ j ] ] ) ;
aux [ j ] = aux . back ( ) ;
aux . pop_back ( ) ;
}
data = std : : move ( selected ) ;
}
2024-01-19 09:45:06 +01:00
fprintf ( stderr , " %s : tokenizing selected tasks \n " , __func__ ) ;
2024-01-18 12:46:27 +01:00
// This is needed as usual for LLaMA models
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
2024-01-19 09:45:06 +01:00
for ( auto & task : data ) {
task . seq_tokens [ 0 ] = : : llama_tokenize ( ctx , task . first + task . choices [ 0 ] + task . second , add_bos ) ;
task . seq_tokens [ 1 ] = : : llama_tokenize ( ctx , task . first + task . choices [ 1 ] + task . second , add_bos ) ;
task . common_prefix = 0 ;
for ( size_t k = 0 ; k < task . seq_tokens [ 0 ] . size ( ) ; k + + ) {
if ( task . seq_tokens [ 0 ] [ k ] ! = task . seq_tokens [ 1 ] [ k ] ) {
break ;
}
task . common_prefix + + ;
}
task . required_tokens = task . common_prefix +
task . seq_tokens [ 0 ] . size ( ) - task . common_prefix +
task . seq_tokens [ 1 ] . size ( ) - task . common_prefix ;
task . n_base1 = : : llama_tokenize ( ctx , task . first + task . choices [ 0 ] , add_bos ) . size ( ) ;
task . n_base2 = : : llama_tokenize ( ctx , task . first + task . choices [ 1 ] , add_bos ) . size ( ) ;
}
2024-01-18 12:46:27 +01:00
fprintf ( stderr , " %s : calculating winogrande score over selected tasks. \n " , __func__ ) ;
const int n_vocab = llama_n_vocab ( llama_get_model ( ctx ) ) ;
2024-01-19 09:45:06 +01:00
const int n_ctx = llama_n_ctx ( ctx ) ;
const int n_batch = params . n_batch ;
const int max_tasks_per_batch = 128 ;
2024-03-11 16:49:47 +01:00
const int max_seq = std : : min ( 2 * max_tasks_per_batch , ( int ) llama_n_seq_max ( ctx ) ) ;
2024-01-19 09:45:06 +01:00
llama_batch batch = llama_batch_init ( n_ctx , 0 , max_seq ) ;
2024-01-18 12:46:27 +01:00
std : : vector < float > tok_logits ( n_vocab ) ;
2024-01-19 09:45:06 +01:00
std : : vector < float > batch_logits ( n_vocab * n_ctx ) ;
2024-01-18 12:46:27 +01:00
2024-01-19 10:39:11 +01:00
std : : vector < std : : pair < size_t , llama_token > > eval_pairs ;
std : : vector < float > eval_results ;
std : : vector < std : : thread > workers ( std : : thread : : hardware_concurrency ( ) ) ;
2024-01-18 12:46:27 +01:00
int n_correct = 0 ;
int n_done = 0 ;
2024-01-19 09:45:06 +01:00
for ( size_t i0 = 0 ; i0 < data . size ( ) ; i0 + + ) {
int n_cur = 0 ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
size_t i1 = i0 ;
size_t i_batch = 0 ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
llama_batch_clear ( batch ) ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
while ( n_cur + ( int ) data [ i1 ] . required_tokens < = n_ctx ) {
const int s0 = 2 * ( i1 - i0 ) ;
if ( s0 + 2 > max_seq ) {
break ;
}
for ( size_t i = 0 ; i < data [ i1 ] . common_prefix ; + + i ) {
llama_batch_add ( batch , data [ i1 ] . seq_tokens [ 0 ] [ i ] , i , { s0 + 0 , s0 + 1 } , false ) ;
}
batch . logits [ batch . n_tokens - 1 ] = true ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
for ( int s = 0 ; s < 2 ; + + s ) {
for ( size_t i = data [ i1 ] . common_prefix ; i < data [ i1 ] . seq_tokens [ s ] . size ( ) ; + + i ) {
llama_batch_add ( batch , data [ i1 ] . seq_tokens [ s ] [ i ] , i , { s0 + s } , true ) ;
}
}
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
data [ i1 ] . i_batch = i_batch ;
i_batch + = data [ i1 ] . required_tokens ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
n_cur + = data [ i1 ] . required_tokens ;
if ( + + i1 = = data . size ( ) ) {
break ;
}
}
if ( i0 = = i1 ) {
fprintf ( stderr , " %s : task %zu does not fit in the context window \n " , __func__ , i0 ) ;
return ;
}
2024-01-18 12:46:27 +01:00
llama_kv_cache_clear ( ctx ) ;
2024-01-19 09:45:06 +01:00
// decode all tasks [i0, i1)
if ( ! decode_helper ( ctx , batch , batch_logits , n_batch , n_vocab ) ) {
fprintf ( stderr , " %s: llama_decode() failed \n " , __func__ ) ;
2024-01-18 12:46:27 +01:00
return ;
}
2024-01-19 10:39:11 +01:00
eval_pairs . clear ( ) ;
2024-01-19 09:45:06 +01:00
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & task = data [ i ] ;
const bool skip_choice =
task . seq_tokens [ 0 ] . size ( ) - task . common_prefix > k_min_trailing_ctx & &
task . seq_tokens [ 1 ] . size ( ) - task . common_prefix > k_min_trailing_ctx ;
const auto & n_base1 = skip_choice ? task . n_base1 : task . common_prefix ;
const int last_1st = task . seq_tokens [ 0 ] . size ( ) - n_base1 > 1 ? 1 : 0 ;
size_t li = n_base1 - 1 ;
for ( size_t j = n_base1 - 1 ; j < task . seq_tokens [ 0 ] . size ( ) - 1 - last_1st ; + + j ) {
2024-02-03 12:23:37 +01:00
eval_pairs . emplace_back ( task . i_batch + li + + , task . seq_tokens [ 0 ] [ j + 1 ] ) ;
2024-01-18 12:46:27 +01:00
}
2024-01-19 09:45:06 +01:00
const auto & n_base2 = skip_choice ? task . n_base2 : task . common_prefix ;
const int last_2nd = task . seq_tokens [ 1 ] . size ( ) - n_base2 > 1 ? 1 : 0 ;
li = task . seq_tokens [ 0 ] . size ( ) - task . common_prefix + n_base2 - 1 ;
for ( size_t j = n_base2 - 1 ; j < task . seq_tokens [ 1 ] . size ( ) - 1 - last_2nd ; + + j ) {
2024-02-03 12:23:37 +01:00
eval_pairs . emplace_back ( task . i_batch + li + + , task . seq_tokens [ 1 ] [ j + 1 ] ) ;
2024-01-18 12:46:27 +01:00
}
2024-01-19 10:39:11 +01:00
}
compute_logprobs ( batch_logits . data ( ) , n_vocab , workers , eval_pairs , eval_results ) ;
size_t ir = 0 ;
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & task = data [ i ] ;
const bool skip_choice =
task . seq_tokens [ 0 ] . size ( ) - task . common_prefix > k_min_trailing_ctx & &
task . seq_tokens [ 1 ] . size ( ) - task . common_prefix > k_min_trailing_ctx ;
2024-01-18 12:46:27 +01:00
2024-01-19 10:39:11 +01:00
float score_1st = 0 ;
const auto & n_base1 = skip_choice ? task . n_base1 : task . common_prefix ;
const int last_1st = task . seq_tokens [ 0 ] . size ( ) - n_base1 > 1 ? 1 : 0 ;
for ( size_t j = n_base1 - 1 ; j < task . seq_tokens [ 0 ] . size ( ) - 1 - last_1st ; + + j ) {
score_1st + = eval_results [ ir + + ] ;
2024-01-19 09:45:06 +01:00
}
2024-01-19 10:39:11 +01:00
score_1st / = ( task . seq_tokens [ 0 ] . size ( ) - n_base1 - last_1st ) ;
2024-01-18 12:46:27 +01:00
2024-01-19 10:39:11 +01:00
float score_2nd = 0 ;
const auto & n_base2 = skip_choice ? task . n_base2 : task . common_prefix ;
const int last_2nd = task . seq_tokens [ 1 ] . size ( ) - n_base2 > 1 ? 1 : 0 ;
for ( size_t j = n_base2 - 1 ; j < task . seq_tokens [ 1 ] . size ( ) - 1 - last_2nd ; + + j ) {
score_2nd + = eval_results [ ir + + ] ;
2024-01-19 09:45:06 +01:00
}
2024-01-19 10:39:11 +01:00
score_2nd / = ( task . seq_tokens [ 1 ] . size ( ) - n_base2 - last_2nd ) ;
2024-01-18 12:46:27 +01:00
2024-01-19 09:45:06 +01:00
int result = score_1st > score_2nd ? 1 : 2 ;
if ( result = = task . answer ) {
+ + n_correct ;
}
+ + n_done ;
2024-01-18 12:46:27 +01:00
2024-01-19 10:39:11 +01:00
// print the accumulated accuracy mean x 100
2024-01-19 09:45:06 +01:00
printf ( " %zu \t %.4lf \t %10.6f %10.6f %d %d \n " , i + 1 , 100.0 * n_correct / n_done , score_1st , score_2nd , result , task . answer ) ;
fflush ( stdout ) ;
2024-01-18 12:46:27 +01:00
}
2024-01-19 09:45:06 +01:00
i0 = i1 - 1 ;
2024-01-18 12:46:27 +01:00
}
printf ( " \n " ) ;
if ( n_done < 100 ) return ;
const float p = 1.f * n_correct / n_done ;
const float sigma = 100.f * sqrt ( p * ( 1 - p ) / ( n_done - 1 ) ) ;
printf ( " Final Winogrande score(%d tasks): %.4lf +/- %.4lf \n " , n_done , 100 * p , sigma ) ;
}
2024-01-23 13:12:57 +01:00
static bool deserialize_string ( std : : istream & in , std : : string & str ) {
2024-01-21 13:42:44 +01:00
uint32_t size ;
if ( ! in . read ( ( char * ) & size , sizeof ( size ) ) . fail ( ) ) {
str . resize ( size ) ;
2024-01-23 13:12:57 +01:00
if ( ! in . read ( ( char * ) & str [ 0 ] , size ) . fail ( ) ) return true ;
2024-01-21 13:42:44 +01:00
}
return false ;
}
struct multiple_choice_answers {
std : : vector < std : : string > answers ;
std : : vector < int > labels ;
bool deserialize ( std : : istream & in ) {
uint32_t n ;
in . read ( ( char * ) & n , sizeof ( n ) ) ;
if ( in . fail ( ) | | n > 100 ) return false ; // 100 as max. number of answers should be good enough for any practical purpose
answers . resize ( n ) ;
labels . resize ( n ) ;
for ( auto & a : answers ) {
if ( ! deserialize_string ( in , a ) ) return false ;
}
in . read ( ( char * ) labels . data ( ) , n * sizeof ( int ) ) ;
return ! in . fail ( ) ;
}
} ;
struct multiple_choice_task {
std : : string question ; // the question (or context that needs to be continued)
multiple_choice_answers mc1 ; // possible answers (continuations) with a single correct answer
multiple_choice_answers mc2 ; // possible answers (continuations) with multiple correct answers - not handled yet
bool deserialize ( std : : istream & in ) {
if ( ! deserialize_string ( in , question ) ) return false ;
return mc1 . deserialize ( in ) & & mc2 . deserialize ( in ) ;
}
// For evaluation
size_t i_batch ; // starting index in the llama_batch
size_t common_prefix ; // max number of initial tokens that are the same in all sentences
size_t required_tokens ; // needed number of tokens to evaluate all answers
std : : vector < std : : vector < llama_token > > seq_tokens ;
std : : vector < float > log_probs ;
} ;
static bool multiple_choice_prepare_one_task ( llama_context * ctx , bool add_bos , multiple_choice_task & task , bool log_error ) {
if ( task . question . empty ( ) | | task . mc1 . answers . empty ( ) ) {
if ( log_error ) {
printf ( " %s: found bad task with empty question and/or answers \n " , __func__ ) ;
}
return false ;
}
task . seq_tokens . reserve ( task . mc1 . answers . size ( ) ) ;
for ( auto & answer : task . mc1 . answers ) {
if ( answer . empty ( ) ) {
if ( log_error ) {
printf ( " %s: found empty answer \n " , __func__ ) ;
}
return false ;
}
task . seq_tokens . emplace_back ( : : llama_tokenize ( ctx , task . question + " " + answer , add_bos ) ) ;
}
auto min_len = task . seq_tokens . front ( ) . size ( ) ;
for ( auto & seq : task . seq_tokens ) {
min_len = std : : min ( min_len , seq . size ( ) ) ;
}
task . common_prefix = 0 ;
for ( size_t k = 0 ; k < min_len ; + + k ) {
auto token = task . seq_tokens [ 0 ] [ k ] ;
bool all_same = true ;
for ( size_t i = 1 ; i < task . seq_tokens . size ( ) ; + + i ) {
if ( task . seq_tokens [ i ] [ k ] ! = token ) {
all_same = false ;
break ;
}
}
if ( ! all_same ) {
break ;
}
+ + task . common_prefix ;
}
task . required_tokens = task . common_prefix ;
for ( auto & seq : task . seq_tokens ) {
task . required_tokens + = seq . size ( ) - task . common_prefix ;
}
return true ;
}
//
// Calculates score for multiple choice tasks with single correct answer from prompt.
// Commonly used LLM evaluation metrics of this type are
// * ARC
// * HellaSwag
// * MMLU
// * TruthfulQA
//
// Validation datasets for these 4 tests can be found at
// https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp
// The data for these datasets was extracted from
// git@hf.co:datasets/allenai/ai2_arc
// https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl
// git@hf.co:datasets/Stevross/mmlu
// https://huggingface.co/datasets/truthful_qa
//
static void multiple_choice_score ( llama_context * ctx , const gpt_params & params ) {
std : : istringstream strstream ( params . prompt ) ;
uint32_t n_task ;
strstream . read ( ( char * ) & n_task , sizeof ( n_task ) ) ;
if ( strstream . fail ( ) | | n_task = = 0 ) {
printf ( " %s: no tasks \n " , __func__ ) ;
return ;
}
printf ( " %s: there are %u tasks in prompt \n " , __func__ , n_task ) ;
std : : vector < uint32_t > task_pos ( n_task ) ;
strstream . read ( ( char * ) task_pos . data ( ) , task_pos . size ( ) * sizeof ( uint32_t ) ) ;
if ( strstream . fail ( ) ) {
printf ( " %s: failed to raad task positions from prompt \n " , __func__ ) ;
return ;
}
std : : vector < multiple_choice_task > tasks ;
if ( params . multiple_choice_tasks = = 0 | | params . multiple_choice_tasks > = ( size_t ) n_task ) {
// Use all tasks
tasks . resize ( n_task ) ;
printf ( " %s: reading tasks " , __func__ ) ;
int n_dot = n_task / 100 ;
int i = 0 ;
for ( auto & task : tasks ) {
+ + i ;
if ( ! task . deserialize ( strstream ) ) {
printf ( " %s: failed to read task %d of %u \n " , __func__ , i , n_task ) ;
return ;
}
if ( i % n_dot = = 0 ) printf ( " . " ) ;
}
printf ( " done \n " ) ;
}
else {
printf ( " %s: selecting %zu random tasks from %u tasks available \n " , __func__ , params . multiple_choice_tasks , n_task ) ;
std : : mt19937 rng ( 1 ) ;
std : : vector < int > aux ( n_task ) ;
for ( uint32_t i = 0 ; i < n_task ; + + i ) aux [ i ] = i ;
float scale = 1.f / ( 1.f + ( float ) std : : mt19937 : : max ( ) ) ;
tasks . resize ( params . multiple_choice_tasks ) ;
for ( auto & task : tasks ) {
int j = ( int ) ( scale * rng ( ) * aux . size ( ) ) ;
int idx = aux [ j ] ;
aux [ j ] = aux . back ( ) ;
aux . pop_back ( ) ;
strstream . seekg ( task_pos [ idx ] , std : : ios : : beg ) ;
if ( ! task . deserialize ( strstream ) ) {
printf ( " %s: failed to read task %d at position %u \n " , __func__ , idx , task_pos [ idx ] ) ;
return ;
}
}
n_task = params . multiple_choice_tasks ;
}
// This is needed as usual for LLaMA models
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
printf ( " %s: preparing task data " , __func__ ) ;
fflush ( stdout ) ;
if ( n_task > 500 ) {
printf ( " ... " ) ;
fflush ( stdout ) ;
std : : atomic < int > counter ( 0 ) ;
std : : atomic < int > n_bad ( 0 ) ;
auto prepare = [ & counter , & n_bad , & tasks , ctx , add_bos ] ( ) {
int num_tasks = tasks . size ( ) ;
int n_bad_local = 0 ;
while ( true ) {
int first = counter . fetch_add ( K_TOKEN_CHUNK ) ;
if ( first > = num_tasks ) {
if ( n_bad_local > 0 ) n_bad + = n_bad_local ;
break ;
}
int last = std : : min ( first + K_TOKEN_CHUNK , num_tasks ) ;
for ( int i = first ; i < last ; + + i ) {
if ( ! multiple_choice_prepare_one_task ( ctx , add_bos , tasks [ i ] , false ) ) + + n_bad_local ;
}
}
} ;
size_t max_thread = std : : thread : : hardware_concurrency ( ) ;
max_thread = std : : min ( max_thread , ( tasks . size ( ) + K_TOKEN_CHUNK - 1 ) / K_TOKEN_CHUNK ) ;
std : : vector < std : : thread > workers ( max_thread - 1 ) ;
for ( auto & w : workers ) w = std : : thread ( prepare ) ;
prepare ( ) ;
for ( auto & w : workers ) w . join ( ) ;
printf ( " done \n " ) ;
fflush ( stdout ) ;
int nbad = n_bad ;
if ( nbad > 0 ) {
printf ( " %s: found %d malformed tasks \n " , __func__ , nbad ) ;
return ;
}
} else {
int n_dot = n_task / 100 ;
int i_task = 0 ;
for ( auto & task : tasks ) {
+ + i_task ;
if ( ! multiple_choice_prepare_one_task ( ctx , add_bos , task , true ) ) {
return ;
}
if ( i_task % n_dot = = 0 ) {
printf ( " . " ) ;
fflush ( stdout ) ;
}
}
printf ( " done \n " ) ;
}
printf ( " %s : calculating TruthfulQA score over %zu tasks. \n " , __func__ , tasks . size ( ) ) ;
printf ( " \n task \t acc_norm \n " ) ;
const int n_vocab = llama_n_vocab ( llama_get_model ( ctx ) ) ;
const int n_ctx = llama_n_ctx ( ctx ) ;
const int n_batch = params . n_batch ;
const int max_tasks_per_batch = 32 ;
2024-03-11 16:49:47 +01:00
const int max_seq = std : : min ( 4 * max_tasks_per_batch , ( int ) llama_n_seq_max ( ctx ) ) ;
2024-01-21 13:42:44 +01:00
llama_batch batch = llama_batch_init ( n_ctx , 0 , max_seq ) ;
std : : vector < float > tok_logits ( n_vocab ) ;
std : : vector < float > batch_logits ( n_vocab * n_ctx ) ;
std : : vector < std : : pair < size_t , llama_token > > eval_pairs ;
std : : vector < float > eval_results ;
std : : vector < std : : thread > workers ( std : : thread : : hardware_concurrency ( ) ) ;
std : : vector < int > batch_indeces ;
int n_done = 0 ;
int n_correct = 0 ;
int n_tot_answers = 0 ;
for ( size_t i0 = 0 ; i0 < tasks . size ( ) ; i0 + + ) {
int n_cur = 0 ;
size_t i1 = i0 ;
size_t i_batch = 0 ; // this tells us where in `llama_batch` we are currently
llama_batch_clear ( batch ) ;
// batch as much tasks as possible into the available context
// each task has 4 unique seuqnce ids - one for each ending
// the common prefix is shared among the 4 sequences to save tokens
// we extract logits only from the last common token and from all ending tokens of each sequence
int s0 = 0 ;
while ( n_cur + ( int ) tasks [ i1 ] . required_tokens < = n_ctx ) {
auto & cur_task = tasks [ i1 ] ;
int num_answers = cur_task . seq_tokens . size ( ) ;
if ( s0 + num_answers > max_seq ) {
break ;
}
if ( int ( batch_indeces . size ( ) ) ! = num_answers ) {
batch_indeces . resize ( num_answers ) ;
}
for ( int s = 0 ; s < num_answers ; + + s ) batch_indeces [ s ] = s0 + s ;
for ( size_t i = 0 ; i < cur_task . common_prefix ; + + i ) {
//llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
llama_batch_add ( batch , cur_task . seq_tokens [ 0 ] [ i ] , i , batch_indeces , false ) ;
}
batch . logits [ batch . n_tokens - 1 ] = true ; // we need logits for the last token of the common prefix
for ( int s = 0 ; s < int ( cur_task . seq_tokens . size ( ) ) ; + + s ) {
for ( size_t i = cur_task . common_prefix ; i < cur_task . seq_tokens [ s ] . size ( ) ; + + i ) {
llama_batch_add ( batch , cur_task . seq_tokens [ s ] [ i ] , i , { s0 + s } , true ) ;
}
}
s0 + = num_answers ;
cur_task . i_batch = i_batch ;
i_batch + = cur_task . required_tokens ;
n_cur + = cur_task . required_tokens ;
if ( + + i1 = = tasks . size ( ) ) {
break ;
}
}
if ( i0 = = i1 ) {
fprintf ( stderr , " %s : task %zu does not fit in the context window \n " , __func__ , i0 ) ;
return ;
}
llama_kv_cache_clear ( ctx ) ;
// decode all tasks [i0, i1)
if ( ! decode_helper ( ctx , batch , batch_logits , n_batch , n_vocab ) ) {
fprintf ( stderr , " %s: llama_decode() failed \n " , __func__ ) ;
return ;
}
// Compute log-probs in parallel
// First we collect all tasks
eval_pairs . clear ( ) ;
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & cur_task = tasks [ i ] ;
size_t li = cur_task . common_prefix ;
for ( int s = 0 ; s < int ( cur_task . seq_tokens . size ( ) ) ; + + s ) {
for ( size_t j = cur_task . common_prefix ; j < cur_task . seq_tokens [ s ] . size ( ) - 1 ; j + + ) {
2024-02-03 12:23:37 +01:00
eval_pairs . emplace_back ( cur_task . i_batch + li + + , cur_task . seq_tokens [ s ] [ j + 1 ] ) ;
2024-01-21 13:42:44 +01:00
}
+ + li ;
}
}
// Then we do the actual calculation
compute_logprobs ( batch_logits . data ( ) , n_vocab , workers , eval_pairs , eval_results ) ;
size_t ir = 0 ;
// compute the logprobs for each ending of the decoded tasks
for ( size_t i = i0 ; i < i1 ; + + i ) {
auto & cur_task = tasks [ i ] ;
//printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str());
//for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) {
// if (cur_task.mc1.labels[j] == 1) {
// printf("%d", j+1);
// }
//}
//printf("\n common_prefix: %zu\n", cur_task.common_prefix);
std : : memcpy ( tok_logits . data ( ) , batch_logits . data ( ) + n_vocab * ( cur_task . i_batch + cur_task . common_prefix - 1 ) , n_vocab * sizeof ( float ) ) ;
const auto first_probs = softmax ( tok_logits ) ;
cur_task . log_probs . resize ( cur_task . seq_tokens . size ( ) ) ;
for ( int s = 0 ; s < int ( cur_task . seq_tokens . size ( ) ) ; + + s ) {
size_t count = 1 ;
float log_prob = std : : log ( first_probs [ cur_task . seq_tokens [ s ] [ cur_task . common_prefix ] ] ) ;
for ( size_t j = cur_task . common_prefix ; j < cur_task . seq_tokens [ s ] . size ( ) - 1 ; j + + ) {
//printf(" %zu %g\n", ir, eval_results[ir]);
+ + count ;
log_prob + = eval_results [ ir + + ] ;
}
cur_task . log_probs [ s ] = log_prob / count ;
//printf(" Final: %g\n", log_prob / count);
//printf(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count);
}
// Find the ending with maximum logprob
size_t logprob_max_idx = 0 ;
float logprob_max_val = cur_task . log_probs [ 0 ] ;
for ( size_t s = 1 ; s < cur_task . log_probs . size ( ) ; s + + ) {
if ( cur_task . log_probs [ s ] > logprob_max_val ) {
logprob_max_val = cur_task . log_probs [ s ] ;
logprob_max_idx = s ;
}
}
n_tot_answers + = cur_task . log_probs . size ( ) ;
if ( cur_task . mc1 . labels [ logprob_max_idx ] = = 1 ) {
+ + n_correct ;
}
+ + n_done ;
// Print the accumulated accuracy mean x 100
printf ( " %d \t %.8lf \n " , n_done , 100. * n_correct / n_done ) ;
fflush ( stdout ) ;
}
i0 = i1 - 1 ;
}
llama_batch_free ( batch ) ;
if ( n_done < 100 ) return ;
float p = 1.f * n_correct / n_done ;
float sigma = sqrt ( p * ( 1 - p ) / ( n_done - 1 ) ) ;
printf ( " \n Final result: %.4f +/- %.4f \n " , 100.f * p , 100.f * sigma ) ;
p = 1.f * n_done / n_tot_answers ;
sigma = sqrt ( p * ( 1 - p ) / ( n_done - 1 ) ) ;
printf ( " Random chance: %.4f +/- %.4f \n " , 100.f * p , 100.f * sigma ) ;
printf ( " \n " ) ;
}
2024-01-22 15:10:14 +01:00
static void kl_divergence ( llama_context * ctx , const gpt_params & params ) {
if ( params . logits_file . empty ( ) ) {
fprintf ( stderr , " %s: you must provide a name of a file containing the log probabilities of the base model \n " , __func__ ) ;
return ;
}
std : : ifstream in ( params . logits_file . c_str ( ) , std : : ios : : binary ) ;
if ( ! in ) {
fprintf ( stderr , " %s: failed to open %s \n " , __func__ , params . logits_file . c_str ( ) ) ;
return ;
}
{
char check [ 9 ] ; check [ 8 ] = 0 ;
in . read ( check , 8 ) ;
if ( in . fail ( ) | | strncmp ( " _logits_ " , check , 8 ) ! = 0 ) {
fprintf ( stderr , " %s: %s does not look like a file containing log-probabilities \n " , __func__ , params . logits_file . c_str ( ) ) ;
return ;
}
}
uint32_t n_ctx ;
in . read ( ( char * ) & n_ctx , sizeof ( n_ctx ) ) ;
if ( n_ctx > llama_n_ctx ( ctx ) ) {
2024-02-18 17:20:12 +01:00
fprintf ( stderr , " %s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry \n " ,
2024-01-22 15:10:14 +01:00
__func__ , params . logits_file . c_str ( ) , n_ctx , params . n_ctx ) ;
}
int n_vocab , n_chunk ;
in . read ( ( char * ) & n_vocab , sizeof ( n_vocab ) ) ;
in . read ( ( char * ) & n_chunk , sizeof ( n_chunk ) ) ;
if ( in . fail ( ) ) {
2024-01-23 14:17:20 +01:00
fprintf ( stderr , " %s: failed reading n_vocab, n_chunk from %s \n " , __func__ , params . logits_file . c_str ( ) ) ;
2024-01-22 15:10:14 +01:00
return ;
}
if ( n_vocab ! = llama_n_vocab ( llama_get_model ( ctx ) ) ) {
fprintf ( stderr , " %s: inconsistent vocabulary (%d vs %d) \n " , __func__ , n_vocab , llama_n_vocab ( llama_get_model ( ctx ) ) ) ;
}
std : : vector < llama_token > tokens ( n_ctx * n_chunk ) ;
if ( in . read ( ( char * ) tokens . data ( ) , tokens . size ( ) * sizeof ( tokens [ 0 ] ) ) . fail ( ) ) {
fprintf ( stderr , " %s: failed reading evaluation tokens from %s \n " , __func__ , params . logits_file . c_str ( ) ) ;
return ;
}
const int n_batch = params . n_batch ;
const int num_batches = ( n_ctx + n_batch - 1 ) / n_batch ;
const int nv = 2 * ( ( n_vocab + 1 ) / 2 ) + 4 ;
const bool add_bos = llama_should_add_bos_token ( llama_get_model ( ctx ) ) ;
std : : vector < uint16_t > log_probs_uint16 ( size_t ( n_ctx - 1 - n_ctx / 2 ) * nv ) ;
2024-01-23 14:17:20 +01:00
std : : vector < float > kld_values ( size_t ( n_ctx - 1 - n_ctx / 2 ) * n_chunk ) ;
2024-01-22 15:10:14 +01:00
std : : vector < float > logits ;
if ( num_batches > 1 ) {
logits . reserve ( n_ctx * n_vocab ) ;
}
std : : vector < std : : thread > workers ( std : : thread : : hardware_concurrency ( ) - 1 ) ;
auto mean_and_uncertainty = [ ] ( double sum , double sum2 , size_t count ) {
if ( count < 1 ) {
return std : : make_pair ( 0. , 0. ) ;
}
double f = sum / count ;
double df = sum2 / count - f * f ;
df = df > 0 & & count > 10 ? sqrt ( df / ( count - 1 ) ) : 0. ;
return std : : make_pair ( f , df ) ;
} ;
kl_divergence_result kld ;
2024-01-23 14:17:20 +01:00
auto kld_ptr = kld_values . data ( ) ;
2024-01-22 15:10:14 +01:00
for ( int i = 0 ; i < n_chunk ; + + i ) {
const int start = i * n_ctx ;
const int end = start + n_ctx ;
const auto t_start = std : : chrono : : high_resolution_clock : : now ( ) ;
if ( in . read ( ( char * ) log_probs_uint16 . data ( ) , log_probs_uint16 . size ( ) * sizeof ( uint16_t ) ) . fail ( ) ) {
fprintf ( stderr , " %s: failed reading log-probs for chunk %d \n " , __func__ , i ) ;
return ;
}
// clear the KV cache
llama_kv_cache_clear ( ctx ) ;
for ( int j = 0 ; j < num_batches ; + + j ) {
const int batch_start = start + j * n_batch ;
const int batch_size = std : : min ( end - batch_start , n_batch ) ;
// save original token and restore it after eval
const auto token_org = tokens [ batch_start ] ;
// add BOS token for the first batch of each chunk
if ( add_bos & & j = = 0 ) {
tokens [ batch_start ] = llama_token_bos ( llama_get_model ( ctx ) ) ;
}
if ( llama_decode ( ctx , llama_batch_get_one ( tokens . data ( ) + batch_start , batch_size , j * n_batch , 0 ) ) ) {
fprintf ( stderr , " %s : failed to eval \n " , __func__ ) ;
return ;
}
// restore the original token in case it was set to BOS
tokens [ batch_start ] = token_org ;
if ( num_batches > 1 ) {
const auto * batch_logits = llama_get_logits ( ctx ) ;
logits . insert ( logits . end ( ) , batch_logits , batch_logits + batch_size * n_vocab ) ;
}
}
const auto t_end = std : : chrono : : high_resolution_clock : : now ( ) ;
if ( i = = 0 ) {
const float t_total = std : : chrono : : duration < float > ( t_end - t_start ) . count ( ) ;
fprintf ( stderr , " %s: %.2f seconds per pass - ETA " , __func__ , t_total ) ;
int total_seconds = ( int ) ( t_total * n_chunk ) ;
if ( total_seconds > = 60 * 60 ) {
fprintf ( stderr , " %d hours " , total_seconds / ( 60 * 60 ) ) ;
total_seconds = total_seconds % ( 60 * 60 ) ;
}
fprintf ( stderr , " %.2f minutes \n " , total_seconds / 60.0 ) ;
2024-01-23 14:17:20 +01:00
printf ( " \n chunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence Same top \n " ) ;
2024-01-22 15:10:14 +01:00
}
const int first = n_ctx / 2 ;
const float * all_logits = num_batches > 1 ? logits . data ( ) : llama_get_logits ( ctx ) ;
process_logits ( n_vocab , all_logits + first * n_vocab , tokens . data ( ) + start + first , n_ctx - 1 - first ,
2024-01-23 14:17:20 +01:00
workers , log_probs_uint16 , kld , kld_ptr ) ;
kld_ptr + = n_ctx - 1 - first ;
2024-01-22 15:10:14 +01:00
auto ppl = mean_and_uncertainty ( kld . sum_nll , kld . sum_nll2 , kld . count ) ;
auto log_ppl_ratio = mean_and_uncertainty ( kld . sum_nll_diff , kld . sum_nll_diff2 , kld . count ) ;
auto kl_div = mean_and_uncertainty ( kld . sum_kld , kld . sum_kld2 , kld . count ) ;
2024-01-23 14:17:20 +01:00
auto p_top = 1. * kld . n_same_top / kld . count ;
auto d_p_top = sqrt ( p_top * ( 1 - p_top ) / ( kld . count - 1 ) ) ;
2024-01-22 15:10:14 +01:00
2024-01-23 14:17:20 +01:00
printf ( " %4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf %.5f ± %.5f \n " , i + 1 , exp ( ppl . first ) ,
log_ppl_ratio . first , log_ppl_ratio . second , kl_div . first , kl_div . second ,
p_top , d_p_top ) ;
2024-01-22 15:10:14 +01:00
fflush ( stdout ) ;
logits . clear ( ) ;
}
printf ( " \n " ) ;
2024-01-23 14:17:20 +01:00
if ( kld . count < 100 ) return ; // we do not wish to do statistics on so few values
std : : sort ( kld_values . begin ( ) , kld_values . end ( ) ) ;
printf ( " ===== KL-divergence statistics \n " ) ;
auto kl_div = mean_and_uncertainty ( kld . sum_kld , kld . sum_kld2 , kld . count ) ;
printf ( " Average: %10.6f ±%10.6lf \n " , kl_div . first , kl_div . second ) ;
auto kld_median = kld_values . size ( ) % 2 = = 0 ? 0.5f * ( kld_values [ kld_values . size ( ) / 2 ] + kld_values [ kld_values . size ( ) / 2 - 1 ] )
: kld_values [ kld_values . size ( ) / 2 ] ;
printf ( " Median : %10.6f \n " , kld_median ) ;
auto percentile = [ & kld_values ] ( float fraction ) {
if ( fraction < = 0 ) return kld_values . front ( ) ;
if ( fraction > = 1 ) return kld_values . back ( ) ;
float p = fraction * ( kld_values . size ( ) - 1 ) ;
size_t ip = size_t ( p ) ; p - = ip ;
return ( 1 - p ) * kld_values [ ip ] + p * kld_values [ std : : min ( ip + 1 , kld_values . size ( ) - 1 ) ] ;
} ;
printf ( " Maximum: %10.6f \n " , kld_values . back ( ) ) ;
printf ( " KLD_99 : %10.6f \n " , percentile ( 0.99f ) ) ;
printf ( " KLD_95 : %10.6f \n " , percentile ( 0.95f ) ) ;
printf ( " KLD_90 : %10.6f \n " , percentile ( 0.90f ) ) ;
printf ( " Minimum: %10.6f \n " , kld_values . front ( ) ) ;
printf ( " KLD_01 : %10.6f \n " , percentile ( 0.01f ) ) ;
printf ( " KLD_05 : %10.6f \n " , percentile ( 0.05f ) ) ;
printf ( " KLD_10 : %10.6f \n " , percentile ( 0.10f ) ) ;
2024-01-22 15:10:14 +01:00
}
2024-01-18 12:46:27 +01:00
2023-03-25 19:26:40 +01:00
int main ( int argc , char * * argv ) {
gpt_params params ;
2023-09-07 19:22:29 +02:00
if ( ! gpt_params_parse ( argc , argv , params ) ) {
2023-03-25 19:26:40 +01:00
return 1 ;
}
2023-09-28 18:04:36 +02:00
params . logits_all = true ;
2024-03-09 19:55:54 +01:00
const int32_t n_ctx = params . n_ctx ;
const bool ppl = ! params . hellaswag & & ! params . winogrande & & ! params . multiple_choice & & ! params . kl_divergence ;
if ( ppl ) {
int n_seq = std : : max ( 1 , params . n_batch / n_ctx ) ;
int32_t n_kv = n_seq * n_ctx ;
params . n_parallel = n_seq ;
params . n_ctx = n_kv ;
params . n_batch = std : : min ( params . n_batch , n_kv ) ;
} else {
params . n_batch = std : : min ( params . n_batch , params . n_ctx ) ;
}
2023-03-25 19:26:40 +01:00
2023-08-23 11:56:42 +02:00
if ( params . ppl_stride > 0 ) {
fprintf ( stderr , " Will perform strided perplexity calculation -> adjusting context size from %d to %d \n " ,
params . n_ctx , params . n_ctx + params . ppl_stride / 2 ) ;
params . n_ctx + = params . ppl_stride / 2 ;
}
2023-09-15 22:59:49 +02:00
print_build_info ( ) ;
2023-05-01 18:23:47 +02:00
2023-06-29 15:15:15 +02:00
if ( params . seed = = LLAMA_DEFAULT_SEED ) {
2023-03-25 19:26:40 +01:00
params . seed = time ( NULL ) ;
}
2023-06-29 15:15:15 +02:00
fprintf ( stderr , " %s: seed = %u \n " , __func__ , params . seed ) ;
2023-03-25 19:26:40 +01:00
std : : mt19937 rng ( params . seed ) ;
if ( params . random_prompt ) {
params . prompt = gpt_random_prompt ( rng ) ;
}
2024-02-16 10:31:07 +01:00
llama_backend_init ( ) ;
llama_numa_init ( params . numa ) ;
2023-05-20 10:06:11 +02:00
2023-06-24 10:47:58 +02:00
llama_model * model ;
2023-03-25 19:26:40 +01:00
llama_context * ctx ;
llama : support Mamba Selective State Space Models (#5328)
* mamba : begin working on support for Mamba SSM
* mamba : begin figuring out how to (ab)use the kv cache for Mamba
* mamba : recurrent inference almost works, but incoherent
* mamba : recurrent inference WORKS!!!
* convert : optionally use d_conv and d_state from config.json for Mamba
* mamba : refactor recurrent conv, resulting in 20% perf increase
It's still slower than I'd like, but I did not really optimize `ggml_exp` yet.
I also refactored `ggml_exp` to work with tensors with more than 2 dimensions.
* ggml : parallelize ggml_exp
This results in 8% faster token generation for Mamba-130M.
* mamba : simplify the conv step with a self-overlapping view
Turns out the conv_state can be made smaller by one column.
Note that this breaks existing GGUFs of Mamba,
because the key_value_length field is tied to the conv_state size.
Convolution with a self-overlapping view is cool!
And it's much simpler than what I initially thought would be necessary
to make the convolution step work with more than 1 token at a time.
Next step is to make the SSM step work on batches of tokens too,
and thus I need to figure out a way to make a parallel selective scan
which will keep the ssm_state small and won't make it bigger
by a factor of (n_layer * batch_size).
* llama : fix Mamba KV self size wrongly displaying as f16 instead of f32
Relatedly, I also tried to see if other types than f32 worked for the states,
but they don't, because of the operators used.
It's probably better anyway to keep lots of precision there,
since the states are small anyway.
* mamba : fix self-overlapping view depth stride
* mamba : handle batches of more than 1 token
This means running Mamba no longer crashes when using the default settings!
And probably also slightly faster prompt processing.
Both batched and non-batched processing yield the same output.
Previously, the state was not cleared when starting a sequence.
Next step is to make the KV cache API work as expected for Mamba models.
* ggml: add ggml_ssm_scan to help with parallel selective scan
If the selective scan was implemented without a custom operator,
there would be waaay too many nodes in the graph. For example,
for Mamba-130M, with a batch size of 512 (the default),
a naive selective scan could add at least 24*512=12288 nodes,
which is more than LLAMA_MAX_NODES (8192),
and that's only for the smallest Mamba model.
So it's much cleaner with a custom operator.
Not sure about the name, though.
* ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation
This will help with performance on CPU if ggml_vec_mul_f32
and ggml_vec_add_f32 are ever optimized with SIMD.
* mamba : very basic quantization support
Mostly works, but there is currently no difference
between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same).
Most of the SSM-specific weights can be kept in f32 without affecting
the size that much, since they are relatively small.
(the linear projection weights are responsible for most of Mamba's size)
Too much quantization seems to make the state degrade quite fast, and
the model begins to output gibberish.
It seems to affect bigger models to a lesser extent than small models,
but I'm not sure by how much.
Experimentation will be needed to figure out which weights are more important
for the _M (and _L?) variants of k-quants for Mamba.
* convert : fix wrong name for layer norm weight of offical Mamba models
I was using Q-bert/Mamba-* models before, which have a slighlty different
naming scheme for the weights.
(they start with "model.layers" instead of "backbone.layers")
* mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator
This increases performance on CPU by around 30% for prompt processing,
and by around 20% for text generation.
However, it also makes the ggml_exp and ggml_soft_plus operators unused.
Whether or not they should be kept will be decided later.
* convert : for Mamba, also consider the "MambaLMHeadModel" arch name
It's the name of the class of the official implementation,
though they don't use it (yet) in the "architectures" field of config.json
* mamba : fix vocab size problems with official models
The perplexity was waaaay to high for models with a non-round vocab size.
Not sure why, but it needed to be fixed in the metadata.
Note that this breaks existing GGUF-converted Mamba models,
but **only if** the vocab size was not already rounded.
* ggml : remove ggml_exp and ggml_soft_plus
They did not exist anyway outside of this branch,
and since ggml_ssm_scan fused operations together, they are unused.
It's always possible to bring them back if needed.
* mamba : remove some useless comments
No code change.
* convert : fix flake8 linter errors
* mamba : apply suggestions from code review
* mamba : remove unecessary branch for row-wise ssm_state and C multiplication
It was previously done to avoid permuting when only one token is processed
at a time (like when generating text), but permuting is cheap,
and dynamically changing the compute graph is not future-proof.
* ggml : in ggml_ssm_scan, use more appropriate asserts
* ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32
* mamba : multiple sequences, but one at a time
This is a step towards making this Mamba implementation usable
with the server example (the way the system prompt is kept when clearing
the client slots will need to be changed before this can work, though).
The KV cache size for this kind of model is tied to the maximum number
of sequences kept at any single time.
For now, this number is obtained from n_parallel (plus one,
to have an extra sequence to dedicate to the system prompt),
but there might be a better way to do this which won't also
make the main example use 2 cells even if only 1 is really used.
(for this specific case, --parallel 0 helps)
Simultaneous sequence processing will probably require changes to
ggml_ssm_scan, and possibly a new operator for the conv step.
* mamba : support llama_kv_cache_seq_cp
This (mis)uses the logic around K shifts, because tokens in a state
can't be shifted anyway, and because inp_K_shift has the right shape and type.
Using ggml_get_rows is a nice way to do copies, but copy chains can't work.
Fortunately, copy chains don't really seem to be used in the examples.
Each KV cell is dedicated to the sequence ID corresponding to its own index.
* mamba : use a state mask
It's cleaner than the previous heuristic of
checking for the pos of the first token in the batch.
inp_KQ_mask could not be re-used for this, because it has the wrong shape
and because it seems more suited to the next step of
simultaneous sequence processing (helping with the problem of
remembering which token belongs to which sequence(s)/state(s)).
* llama : replace the usage of n_ctx with kv_self.size in many places
* mamba : use n_tokens directly instead of n_tok
* mamba : in comments, properly refer to KV cells instead of slots
* mamba : reduce memory usage of ggml_ssm_scan
From 290.37 MiB to 140.68 MiB of CPU compute buffer size
with Mamba 3B with a batch size of 512.
The result tensor of ggml_ssm_scan was previously a big part
of the CPU compute buffer size. To make it smaller,
it does not contain the intermediate ssm states anymore.
Both y and the last ssm state are combined in the result tensor,
because it seems only a single tensor can be returned by an operator
with the way the graph is built.
* mamba : simultaneous sequence processing
A batch can now contain tokens from multiple sequences.
This is necessary for at least the parallel example, the server example,
and the HellaSwag test in the perplexity example.
However, for this to be useful, uses of llama_kv_cache_seq_rm/cp
will need to be changed to work on whole sequences.
* ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba
This operator makes it possible to use and update the correct states
for each token of the batch in the same way as ggml_ssm_scan.
Other solutions which use existing operators would need loops which would
add too many nodes to the graph (at least the ones I thought of).
Using this operator further reduces the size of the CPU compute buffer
from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512.
And (at least on CPU), it's a bit faster than before.
Note that "ggml_ssm_conv" is probably not the most appropriate name,
and it could be changed if a better one is found.
* llama : add inp_s_seq as a new input tensor
The most convenient implementation to select the correct state (for Mamba)
for each token is to directly get the correct index from a tensor.
This is why inp_s_seq is storing int32_t and not floats.
The other, less convenient way to select the correct state would be
to have inp_KQ_mask contain 1.0f for each state used by a token
and 0.0f otherwise. This complicates quickly fetching the first used
state of a token, and is also less efficient because a whole row
of the mask would always need to be read for each token.
Using indexes makes it easy to stop searching when there are
no more sequences for a token, and the first sequence assigned
is always very quickly available (it's the first element of each row).
* mamba : support llama_kv_cache_seq_cp copy chains
* mamba : support shifting and dividing the kv cache pos
* mamba : make the server and parallel examples work with whole sequences
A seq_id is dedicated to the system prompt in both cases.
* llama : make llama_kv_cache_seq_rm return whether it succeeded or not
* mamba : dedicate an input tensor for state copy indices
This is cleaner and makes it easier to adapt when/if token positions
(and by extension, inp_K_shift) are no longer integers.
* mamba : adapt perplexity, batched, and batched-bench examples
* perplexity : limit the max number of sequences
This adapts to what the loaded model can provide.
* llama : add llama_n_max_seq to get the upper limit for seq_ids
Used by the perplexity example.
* batched : pass n_parallel to the model's context params
This should have been there already, but it wasn't.
* batched-bench : reserve sequences to support Mamba
* batched-bench : fix tokens being put in wrong sequences
Generation quality isn't what's measured in there anyway,
but at least using the correct sequences avoids using non-consecutive
token positions.
* mamba : stop abusing attention metadata
This breaks existing converted-to-GGUF Mamba models,
but will allow supporting mixed architectures like MambaFormer
without needing to break Mamba models.
This will also allow changing the size of Mamba's states
without having to reconvert models in the future.
(e.g. using something else than d_conv - 1 columns for the conv_states
will not require breaking existing converted Mamba models again)
* gguf-py : add new KV metadata key-value pairs for Mamba
* llama : add new metadata key-value pairs for Mamba
* llama : guard against divisions by zero when n_head is 0
* mamba : rename "unlimited" KV cache property to "recurrent"
* mamba : more correctly update the "used" field of the KV cache
* ggml : in ggml_ssm_scan, use a threshold for soft_plus
This is how the official Mamba implementation does it,
and it's also what torch.nn.Softplus does.
* convert : for Mamba, fallback to internal NeoX tokenizer
The resulting models are exactly the same
as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there.
* mamba : support state saving and restoring
* ggml : implicitly pass src tensors through dst for Mamba-related ops
* mamba : clarify some comments
* server : fix cache_tokens not getting correctly resized
Otherwise, when the "we have to evaluate at least 1 token" special case
was triggered, an extra token was kept in cache_tokens even if it was
removed from the KV cache.
For Mamba, this caused useless prompt reprocessing when the previous
request triggered the above case.
* convert-hf : support new metadata keys for Mamba
For the models available at
https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406
* mamba : rename metadata to be more similar to transformers library
This breaks existing converted-to-GGUF models,
but the metadata names are more "standard".
* mamba : support mamba-*-hf models
These models share their token_embd.weight with their output.weight
* mamba : add missing spaces
This is purely a formatting change.
* convert-hf : omit output.weight when identical with token_embd.weight
Only for Mamba for now, but it might be relevant for other models eventually.
Most Mamba models actually share these two tensors, albeit implicitly.
* readme : add Mamba to supported models, and add recent API changes
* mamba : move state_seq and state_mask views outside layer loop
A few tensors were also missing `struct` in front of `ggml_tensor`.
2024-03-08 23:31:00 +01:00
// ensure there's at least enough seq_ids for HellaSwag
params . n_parallel = std : : max ( 4 , params . n_parallel ) ;
2023-05-02 22:39:51 +02:00
// load the model and apply lora adapter, if any
2023-06-24 10:47:58 +02:00
std : : tie ( model , ctx ) = llama_init_from_gpt_params ( params ) ;
if ( model = = NULL ) {
2023-05-02 22:39:51 +02:00
fprintf ( stderr , " %s: error: unable to load model \n " , __func__ ) ;
return 1 ;
2023-04-17 17:28:55 +02:00
}
2023-09-28 21:42:38 +02:00
const int n_ctx_train = llama_n_ctx_train ( model ) ;
2023-09-08 17:43:35 +02:00
if ( params . n_ctx > n_ctx_train ) {
fprintf ( stderr , " %s: warning: model was trained on only %d context tokens (%d specified) \n " ,
__func__ , n_ctx_train , params . n_ctx ) ;
2023-09-04 21:26:24 +02:00
}
2023-03-25 19:26:40 +01:00
// print system information
{
fprintf ( stderr , " \n " ) ;
2023-09-28 21:42:38 +02:00
fprintf ( stderr , " %s \n " , get_system_info ( params ) . c_str ( ) ) ;
2023-03-25 19:26:40 +01:00
}
2023-08-28 17:59:39 +02:00
struct results_perplexity results ;
2023-07-28 20:25:36 +02:00
if ( params . hellaswag ) {
hellaswag_score ( ctx , params ) ;
2024-01-18 12:46:27 +01:00
} else if ( params . winogrande ) {
winogrande_score ( ctx , params ) ;
2024-01-21 13:42:44 +01:00
} else if ( params . multiple_choice ) {
multiple_choice_score ( ctx , params ) ;
2024-01-22 15:10:14 +01:00
} else if ( params . kl_divergence ) {
kl_divergence ( ctx , params ) ;
2023-07-22 14:21:24 +02:00
} else {
2024-03-09 19:55:54 +01:00
results = perplexity ( ctx , params , n_ctx ) ;
2023-07-22 14:21:24 +02:00
}
2023-03-25 19:26:40 +01:00
llama_print_timings ( ctx ) ;
2023-08-28 17:59:39 +02:00
write_logfile ( ctx , params , model , results ) ;
2023-03-25 19:26:40 +01:00
llama_free ( ctx ) ;
2023-06-24 10:47:58 +02:00
llama_free_model ( model ) ;
2023-03-25 19:26:40 +01:00
2023-07-10 17:49:56 +02:00
llama_backend_free ( ) ;
2023-03-25 19:26:40 +01:00
return 0 ;
}