Skip to content

Commit a5a11bf

Browse files
ggerganovslaren
andauthored
Update tests/test-sampling.cpp
Co-authored-by: slaren <[email protected]>
1 parent 3cb33a8 commit a5a11bf

File tree

1 file changed

+19
-19
lines changed

1 file changed

+19
-19
lines changed

tests/test-sampling.cpp

Lines changed: 19 additions & 19 deletions
Original file line numberDiff line numberDiff line change
@@ -248,25 +248,25 @@ static void test_sampler_queue(const size_t n_vocab, const std::string & sampler
248248
samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p);
249249
}
250250

251-
#define BENCH(__cnstr, __data, __n_iter) do { \
252-
auto * cnstr = (__cnstr); \
253-
std::vector<llama_token_data> cur((__data).size()); \
254-
std::copy((__data).begin(), (__data).end(), cur.begin()); \
255-
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }; \
256-
llama_sampler_apply(cnstr, &cur_p); \
257-
llama_sampler_reset(cnstr); \
258-
const int64_t t_start = ggml_time_us(); \
259-
const int n_iter = (__n_iter); \
260-
for (int i = 0; i < n_iter; i++) { \
261-
std::copy((__data).begin(), (__data).end(), cur.begin()); \
262-
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false }; \
263-
llama_sampler_apply(cnstr, &cur_p); \
264-
llama_sampler_reset(cnstr); \
265-
} \
266-
const int64_t t_end = ggml_time_us(); \
267-
llama_sampler_free(cnstr); \
268-
printf("%-42s: %8.3f us/iter\n", #__cnstr, (t_end - t_start) / (float)n_iter); \
269-
} while(0)
251+
static void bench(llama_sampler * cnstr, const char * cnstr_name, const std::vector<llama_token_data> & data, int n_iter) {
252+
std::vector<llama_token_data> cur(data.size());
253+
std::copy(data.begin(), data.end(), cur.begin());
254+
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
255+
llama_sampler_apply(cnstr, &cur_p);
256+
llama_sampler_reset(cnstr);
257+
const int64_t t_start = ggml_time_us();
258+
for (int i = 0; i < n_iter; i++) {
259+
std::copy(data.begin(), data.end(), cur.begin());
260+
llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
261+
llama_sampler_apply(cnstr, &cur_p);
262+
llama_sampler_reset(cnstr);
263+
}
264+
const int64_t t_end = ggml_time_us();
265+
llama_sampler_free(cnstr);
266+
printf("%-42s: %8.3f us/iter\n", cnstr_name, (t_end - t_start) / (float)n_iter);
267+
}
268+
269+
#define BENCH(__cnstr, __data, __n_iter) bench((__cnstr), #__cnstr, (__data), (__n_iter))
270270

271271
static void test_perf() {
272272
const int n_vocab = 1 << 17;

0 commit comments

Comments
 (0)