Stage 1: runtime knob sweep (poll, cpu-strict, load mode) and -ub sweep: no gains

README.md

@@ -82,6 +82,13 @@ Interleaved runs on a shared host (see `benchmarks/ab_results.csv`; 12-thread re
   much slower. `ngram-mod` costs little on novel text and helps when output repeats earlier text (editing files, echoing tool
   output). "Warm" means the same text had already been generated once in the session, so it is a best case.
 
+- **Runtime knobs: nothing to gain** (`benchmarks/knobs_decode.csv`, tg64, 12 threads, 3 interleaved rounds, medians):
+  default `--poll 50` 10.37, `--poll 0` 10.16, `--poll 100` 10.39, `--cpu-strict 1` 9.96, `-lm none` 10.20 t/s. All differences are
+  inside the run-to-run noise (single outliers of ~6 t/s appeared once for `--cpu-strict 1` and `-lm none`). `--prio 2` did not run
+  in this container. Transparent hugepages are `madvise` system-wide and were not changed.
+- **`-ub` for prefill** (`benchmarks/ubatch_prefill.csv`, pp512, 12 threads, 2 rounds): 128 -> 33-35 t/s, 256 -> 37, 512 -> 40,
+  1024 -> 39-41 t/s. The default 512 is already the best choice.
+
 Suggested server flags on a quota-limited shared host: `-t 12 -fa on -lm mlock --spec-type ngram-mod`.
 
 ## Accuracy

benchmarks/knobs_decode.csv

@@ -0,0 +1,20 @@
+round,config,tg64_tps
+1,base,10.370985
+1,poll0,10.163646
+1,poll100,10.389921
+1,strict,6.191479
+1,lmnone,6.667170
+1,prio2,
+2,poll0,9.785995
+2,poll100,10.493260
+2,strict,9.958393
+2,lmnone,10.200113
+2,prio2,
+2,base,9.529010
+3,poll100,9.595270
+3,strict,9.990220
+3,lmnone,10.460284
+3,prio2,
+3,base,10.366802
+3,poll0,10.329963
+KNOBS_DONE

benchmarks/ubatch_prefill.csv

@@ -0,0 +1,10 @@
+round,ub,pp512_tps
+1,128,33.298578
+1,256,37.306879
+1,512,40.081132
+1,1024,39.106427
+2,128,34.787931
+2,256,37.228463
+2,512,39.742511
+2,1024,40.676948
+UB_DONE