happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

with logging

This commit is contained in:
ddh0 2025-12-14 00:36:59 -06:00
parent d1e5c60442
commit 9613c48172
1 changed files with 16 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
src/llama-sampling.cpp
View File

@ -2362,12 +2362,16 @@ static const char * llama_sampler_power_law_name(const struct llama_sampler * /*
static float llama_sampler_power_law_compute_target(const llama_sampler_power_law * ctx, float decay) { static float llama_sampler_power_law_compute_target(const llama_sampler_power_law * ctx, float decay) {
if (ctx->total_weight == 0.0f) { if (ctx->total_weight == 0.0f) {
// if there is no history, just use base target // if there is no history, just use base target
fprintf(stderr, "power-law: compute_target: total_weight == 0.0 (target fixed at %.3f)\n", ctx->target);
fflush(stderr);
return ctx->target; return ctx->target;
} }
// maintain a running weighted sum with exponential decay // maintain a running weighted sum with exponential decay
float new_total_weight = 1.0f + decay * ctx->total_weight; float new_total_weight = 1.0f + decay * ctx->total_weight;
fprintf(stderr, "power-law: compute_target: new_total_weight = %.3f\n", new_total_weight); fflush(stderr);
float next_value = ctx->target * new_total_weight - decay * ctx->weighted_sum; float next_value = ctx->target * new_total_weight - decay * ctx->weighted_sum;
fprintf(stderr, "power-law: compute_target: next_value = %.3f\n", next_value); fflush(stderr);
// clamp to [0.0, 1.0] // clamp to [0.0, 1.0]
return std::max(0.0f, std::min(next_value, 1.0f)); return std::max(0.0f, std::min(next_value, 1.0f));
@ -2378,6 +2382,7 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
if (ctx->target < 0.0f) { if (ctx->target < 0.0f) {
// no-op: just sample from the distribution as-is // no-op: just sample from the distribution as-is
fprintf(stderr, "power-law: no-op!"); fflush(stderr);
llama_sampler_softmax_impl(cur_p, false); llama_sampler_softmax_impl(cur_p, false);
const int idx = llama_sample_dist(cur_p, ctx->rng); const int idx = llama_sample_dist(cur_p, ctx->rng);
cur_p->selected = idx; cur_p->selected = idx;
@ -2386,6 +2391,7 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
// clamp decay to avoid degenerate case at 1.0 (unbounded accumulation) // clamp decay to avoid degenerate case at 1.0 (unbounded accumulation)
const float decay = std::min(ctx->decay, 0.99f); const float decay = std::min(ctx->decay, 0.99f);
fprintf(stderr, "power-law: decay = %.3f\n", decay); fflush(stderr);
// fixed power law transform parameters // fixed power law transform parameters
const float distribution_width = 0.3f; const float distribution_width = 0.3f;
@ -2403,6 +2409,7 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
} }
float computed_target = llama_sampler_power_law_compute_target(ctx, decay); float computed_target = llama_sampler_power_law_compute_target(ctx, decay);
fprintf(stderr, "power-law: computed_target = %.3f\n", computed_target); fflush(stderr);
// //
// power law transform // power law transform
@ -2410,8 +2417,12 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
for (size_t i = 0; i < cur_p->size; ++i) { for (size_t i = 0; i < cur_p->size; ++i) {
float p = cur_p->data[i].p; float p = cur_p->data[i].p;
float normalized_distance = std::abs(p - computed_target) / distribution_width; fprintf(stderr, "power-law: transform: p = %.3f\n", p); fflush(stderr);
cur_p->data[i].logit = peak_logit_value / (1.0f + std::pow(normalized_distance, tail_heaviness)); float normed_distance = std::abs(p - computed_target) / distribution_width;
fprintf(stderr, "power-law: transform: normed_distance = %.3f\n", normed_distance); fflush(stderr);
float new_p = peak_logit_value / (1.0f + std::pow(normed_distance, tail_heaviness));
fprintf(stderr, "power-law: transform: new_p = %.3f\n", new_p); fflush(stderr);
cur_p->data[i].logit = new_p;
} }
llama_sampler_softmax_impl(cur_p, false); llama_sampler_softmax_impl(cur_p, false);
@ -2419,6 +2430,7 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
// sample from transformed distribution // sample from transformed distribution
const int idx = llama_sample_dist(cur_p, ctx->rng); const int idx = llama_sample_dist(cur_p, ctx->rng);
cur_p->selected = idx; cur_p->selected = idx;
fprintf(stderr, "power-law: selected token %d\n", idx); fflush(stderr);
// update running history with the original probability of the selected token // update running history with the original probability of the selected token
float original_p = original_probs[idx]; float original_p = original_probs[idx];