card: real step sweep, serve line without ignored hf-overrides

README.md

@@ -42,7 +42,7 @@ oui-1 are committed to the benchmark repo so the number can be rescored offline.
 
 | Model | Active params | Complete runs |
 |---|---|---|
-| **oui-1** (this model, 64 denoising steps) | 4B | **132/184 (71.7%)** |
+| **oui-1** (this model, 48 denoising steps) | 4B | **132/184 (71.7%)** |
 | Gemma 4 31B | 31B | 86/184 (46.7%) |
 | Phi-4 14B | 14B | 81/184 (44.0%) |
 | Gemma 4 26B-A4B (autoregressive twin) | 4B | 55/184 (29.9%) |
@@ -60,16 +60,17 @@ channel in 167 of 184 outputs and the ones that think complete at the same rate.
 On 100 fresh briefs written after training in the same numbered register (never seen by the
 model, one generation each, internal scorer): 69/100.
 
-Speed, 64 denoising steps, single request on one A100 80GB (vLLM 0.24 FP8, entropy-bound sampler,
+Speed, 48 denoising steps, single request on one A100 80GB (vLLM 0.24 FP8, entropy-bound sampler,
 wall-clock from request to last byte including the ~5k-token prompt), all 46 benchmark briefs,
 median per band: lightest screens (~9 statements) 1.0 s, light (~15) 1.5 s, medium (~30) 2.8 s,
 dense (~46) 4.6 s, densest (~64) 5.7 s; across the 46, median 2.6 s, p90 5.7 s, about 206
 completion tokens per second of wall-clock (250 to 360 tokens per second after the first canvas).
 The sampler stops early on its own (15 to 23 denoising steps per 256-token canvas on average), so
-a 32-step cap changes little: 133/184 (72.3%) on the public protocol against 132/184 at 64 steps,
-and every band within 0.2 s per screen. Four concurrent requests give about 54 screens a minute
-at 170 aggregate tokens per second; on short screens the prompt prefill dominates and does not
-batch well.
+the step cap matters little above 32: on the public protocol 32 steps score 134/184, 48 (the
+checkpoint default, used for every number here) 132/184, 64 steps 126/184; 16 steps collapse to
+59/184 because the forced commit breaks the wiring. Four concurrent requests give about 54 screens
+a minute at 170 aggregate tokens per second; on short screens the prompt prefill dominates and
+does not batch well.
 
 ## What it was trained on
 
@@ -84,13 +85,15 @@ teacher-written answer to them. The training set is not released.
 
 ## How to use
 
-Serve with vLLM (diffusion sampler settings matter; these are the ones every number above was
-measured with):
+Serve with vLLM 0.24 or newer. The sampler settings come from the checkpoint's own
+`generation_config.json` (entropy-bound sampler, entropy bound 0.1, 48 denoising steps, 256-token
+canvas); every number above was measured with them. To change the step cap use
+`--diffusion-config '{"canvas_length":256,"max_denoising_steps":N}'`; note that `--hf-overrides`
+keys such as `diffusion_max_denoising_steps` are not read by vLLM.
 
 ```bash
 vllm serve thesysdev/OUI-1 --trust-remote-code --max-model-len 16384 --quantization fp8 \
-  --served-model-name oui-1 \
-  --hf-overrides '{"diffusion_sampler":"entropy_bound","diffusion_entropy_bound":0.1,"diffusion_max_denoising_steps":64}' \
+  --served-model-name oui-1 --max-num-seqs 4 \
   --enable-auto-tool-choice --tool-call-parser gemma4
 ```