root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 11:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

sampling : fix off-by-one in tail-free sampling

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov 2024-09-23 11:44:55 +03:00
parent 37f8c7b4c9
commit 114ab6347e
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
3 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
common/arg.cpp
View File

@ -963,7 +963,7 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
} }
).set_sparam()); ).set_sparam());
add_opt(llama_arg( add_opt(llama_arg(
{"--tfs"}, "N", {"--tfs", "--tfs-z"}, "Z",
format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z), format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z),
[](gpt_params & params, const std::string & value) { [](gpt_params & params, const std::string & value) {
params.sparams.tfs_z = std::stof(value); params.sparams.tfs_z = std::stof(value);

8
src/llama-sampling.cpp
View File

@ -756,20 +756,22 @@ static void llama_sampler_tail_free_apply(struct llama_sampler * smpl, llama_tok
} }
} }
assert(cur_p->size > 0); // guaranteed earlier
size_t last_idx = cur_p->size - 1;
float cum_sum = 0.0f; float cum_sum = 0.0f;
size_t last_idx = cur_p->size;
for (size_t i = 0; i < second_derivatives.size(); ++i) { for (size_t i = 0; i < second_derivatives.size(); ++i) {
cum_sum += second_derivatives[i]; cum_sum += second_derivatives[i];
// Check if the running sum is greater than z or if we have kept at least min_keep tokens // Check if the running sum is greater than z or if we have kept at least min_keep tokens
if (cum_sum > ctx->z && i >= ctx->min_keep) { if (cum_sum > ctx->z && (i + 1) >= ctx->min_keep) {
last_idx = i; last_idx = i;
break; break;
} }
} }
// Resize the output vector to keep only the tokens above the tail location // Resize the output vector to keep only the tokens above the tail location
cur_p->size = last_idx; cur_p->size = last_idx + 1;
} }
static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) { static struct llama_sampler * llama_sampler_tail_free_clone(const struct llama_sampler * smpl) {

4
tests/test-sampling.cpp
View File

@ -272,8 +272,8 @@ int main(void) {
test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f); test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.50f);
test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f, 0.20f}, 0.80f);
test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f); test_typical({0.97f, 0.01f, 0.01f, 0.01f}, {0.97f}, 0.5f);
test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f); test_typical({0.4f, 0.2f, 0.2f, 0.2f}, {0.2f, 0.2f, 0.2f}, 0.5f);