Commit e906075

mo khan <mo@mokhan.ca>
2026-09-05 16:35:59
docs: delete comments referencing an eval harness that does not exist
Three comments pointed at things that are not in this repo: an "eval harness" and an "eval plugin". Nothing in lib/ or ext/ matches either, and the misdirection was real -- I took the temp/seed comments at face value while working on KV cache reuse and spent a while reasoning about reproducibility for a consumer that was never there. Deleted rather than reworded. The signatures say what the parameters are, and a comment that had to be fact-checked to be trusted was not earning its place. Comment-only; no behavior change. Claude-Session: https://claude.ai/code/session_01FpbgyAMtPEkDbo2kx78qR6
main
1 parent 12e777d
Changed files (1)
ext
elelem
llama
elelem_llama.cpp
ext/elelem/llama/elelem_llama.cpp
@@ -32,7 +32,7 @@ struct el_handle {
                          // for how its KV cache is reused between calls
     int n_ctx;
     int n_threads;
-    float temp;       // <= 0 => greedy (deterministic); used by evals
+    float temp;
     uint32_t seed;
     bool dead = false; // set once the GPU device is lost; see el_close
     // Exactly the tokens currently resident in the KV cache, in position
@@ -101,10 +101,6 @@ static llama_context *el_init_context(llama_model *model, int n_ctx, int n_threa
     return nullptr;
 }
 
-// temp <= 0 selects greedy/deterministic sampling; seed is the RNG seed for the
-// sampled path (both surfaced so callers -- notably the eval harness -- can pin them).
-// n_threads <= 0 means "auto": use llama.cpp's own physical-core/SMT-aware
-// detection (common_cpu_get_num_math) instead of a thread count picked in Ruby.
 void *el_open(const char *path, int n_gpu_layers, int n_ctx, int n_threads, float temp, int seed) {
     if (!g_backend) {
         llama_log_set(el_log_callback, nullptr);