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

fix compiler warning, add commented-out logging per token

This commit is contained in:
ddh0 2025-12-13 00:23:15 -06:00
parent 0a19a3fd6c
commit 824bb3aa6e
1 changed files with 13 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
src/llama-sampling.cpp
View File

@ -2401,7 +2401,7 @@ static float llama_sampler_power_law_compute_target(
if (sz > 0) {
// Check if window is at capacity (oldest element will be evicted on next push)
// Use the window_size parameter from context, not a capacity() method
const bool window_full = (sz == ctx->window_size);
const bool window_full = (sz == (size_t)ctx->window_size);
// Compute weighted sum with exponential decay
// rat(0) = newest in buffer, gets weight 1
@ -2496,6 +2496,18 @@ static void llama_sampler_power_law_apply(struct llama_sampler * smpl, llama_tok
const int idx = llama_sample_dist(cur_p, ctx->rng);
cur_p->selected = idx;
// uncomment this to log the target values and history window contents for every token
//
// fprintf(stderr, "power_law: window_size=%zu/%d values=[",
// ctx->window.size(), ctx->window_size);
// for (size_t i = 0; i < ctx->window.size(); ++i) {
// fprintf(stderr, "%.1f", ctx->window.rat(i));
// if (i < ctx->window.size() - 1) fprintf(stderr, ",");
// }
// fprintf(stderr, "] computed_target=%.4f selected_token=%d orig_prob=%.4f\n",
// computed_target, cur_p->data[idx].id, original_probs[idx]);
// fflush(stderr);
// add the ORIGINAL probability to the rolling window
float original_p = original_probs[idx];