TPS optimization

State

Completed · 64 rollouts

Report date

13-Nov-2026

Configuration

32 lanes · No speculation; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks × eight rollouts

Hardware

One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement

Stack

Gemma 4 31B W4A16 QAT · rank-16 LoRA · vLLM 0.19.1 · temperature 1. Exceptions are stated above.

Control

Baseline row

Window

895 s server-busy measurement window; not whole-job wall time

Outcome

36/64 solved (56%); 83% hit cap; 251.7 rollouts/hour

Limits

Real Docker sandboxes and grading. Not a golden evaluation or promotion.

Replay pads turns to logged lengths; four-turn scaling cells have shorter prompts. Combined stack = ngram k=3, thinking-budget patch, sliding-window prefix-hit fix and tuned LoRA kernels. Aligned-engine compatibility is not established by this measurement. Report dates do not establish exact start timestamps.

State

Completed · 64 rollouts

Report date

13-Nov-2026

Configuration

32 lanes · ngram k=3 + thinking budget patch; no adapter; budget_aware_v2_plus_31b; temperature 1; 7.5-minute cap; eight dev-v1 tasks × eight rollouts

Hardware

One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement

Stack

Gemma 4 31B W4A16 QAT · rank-16 LoRA · vLLM 0.19.1 · temperature 1. Exceptions are stated above.

Control

Same-box paired task arms

Window

895 s server-busy measurement window; not whole-job wall time

Outcome

33/64 solved (52%); 77% hit cap; 251.7 rollouts/hour

Limits

Real Docker sandboxes and grading. Not a golden evaluation or promotion.

Replay pads turns to logged lengths; four-turn scaling cells have shorter prompts. Combined stack = ngram k=3, thinking-budget patch, sliding-window prefix-hit fix and tuned LoRA kernels. Aligned-engine compatibility is not established by this measurement. Report dates do not establish exact start timestamps.

State

Preempted · 207/480 rollouts

Report date

13-Nov-2026

Configuration

32 lanes · Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks × eight rollouts planned

Hardware

One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement

Stack

Gemma 4 31B W4A16 QAT · rank-16 LoRA · vLLM 0.19.1 · temperature 1. Exceptions are stated above.

Control

Historical / not matched: partial 207-rollout attempt vs earlier 64-rollout arm

Window

Not recorded per cell

Outcome

58.5% resolved in this partial attempt; roughly 253 rollouts/hour

Limits

Partial attempt; do not mix with restarted complete run

Replay pads turns to logged lengths; four-turn scaling cells have shorter prompts. Combined stack = ngram k=3, thinking-budget patch, sliding-window prefix-hit fix and tuned LoRA kernels. Aligned-engine compatibility is not established by this measurement. Report dates do not establish exact start timestamps.

State

Completed · 480 rollouts

Report date

13-Nov-2026 · Saved run metadata: 13-Nov 22:19 to 14-Nov 00:07, America/New_York (EST)

Configuration

32 lanes · Combined stack; zero-output r16 policy LoRA; 60 dev-v1 tasks × eight; temperature 1; 7.5-minute cap

Hardware

One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement

Stack

Gemma 4 31B W4A16 QAT · rank-16 LoRA · vLLM 0.19.1 · temperature 1. Exceptions are stated above.

Control

Historical / not matched: complete 480-rollout run vs earlier 64-rollout arm

Window

108.5 min whole rollout attempt; first 256-rollout update ready at 61.1 min

Outcome

260/479 solved (54%); 265 rollouts/hour overall, 254/hour steady; 70% capped; KV peak 91%; no preemptions

Limits

479 graded outcomes, one infra drop. Zero-output LoRA pays adapter compute but is not a trained policy. Aligned-engine compatibility remains unmeasured. Linked saved-run browser reports 261/480 solved, while the pinned throughput report states 260/479 after excluding one infra drop. Counts differ and have not been reconciled; retain both rather than treat them as identical scoring.

Replay pads turns to logged lengths; four-turn scaling cells have shorter prompts. Combined stack = ngram k=3, thinking-budget patch, sliding-window prefix-hit fix and tuned LoRA kernels. Aligned-engine compatibility is not established by this measurement. Report dates do not establish exact start timestamps.

State

Cancelled · 289/480 rollouts

Report date

14-Nov-2026

Configuration

48 lanes · Same 480 planned rollouts; combined stack + async fix + latest-match drafting; bf16 KV

Hardware

One NVIDIA RTX PRO 6000 Blackwell (96 GB) per measurement

Stack

Gemma 4 31B W4A16 QAT · rank-16 LoRA · vLLM 0.19.1 · temperature 1. Exceptions are stated above.

Control

Different whole-run busy windows; no TPS improvement inferred

Window

51 min before cancellation; 36 groups of eight complete

Outcome

99% capped; 82% empty patches; KV 100%; 173 preemptions; prefix hits 49.7%; 367 rollouts/hour over first 2838 s

Limits

Do not compare whole-window server metrics as if time-matched. Faster rollouts/hour here means capped, truncated trajectories, not better useful output.

Replay pads turns to logged lengths; four-turn scaling cells have shorter prompts. Combined stack = ngram k=3, thinking-budget patch, sliding-window prefix-hit fix and tuned LoRA kernels. Aligned-engine compatibility is not established by this measurement. Report dates do not establish exact start timestamps.