# SGLang on RTX 5090 (sm_120), Qwen3.8-27B, 2026-08-20
# Image: lmsysorg/sglang:qwen38-27b (0.0.0.dev0+qwen38.27b.g561c8f3)
# Model: gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 (ModelOpt NVFP4 W4A4, FP8 KV)
# Draft: gittensor-model-hub/Qwen3.8-27B-DSpark-NVFP4 (DSPARK, block-size 7, modelopt_fp4)
# concbench.py, 800 max_tokens, merge_intervals prompt

# --- SGLang + DSpark (speculative; build clamps max_running_requests to 2) ---
N=1  agg=227.3  per-req=227.3  p50=2.83s   (643 tok/2.83s)   <-- single-stream champion
N=2  agg=505.0  per-req=252.6  p50=2.51s   (1270 tok/2.51s)  <-- N=2 champion (spec + batch compound)
N=4  agg=520.5  per-req=173.5  p50=5.34s   (queues past the 2-slot DSpark cap)
# coherence (thinking off): "lambda x: x**2" -> clean. Strict acceptance = target's exact output.

# --- SGLang no-spec (continuous batching, max-running-requests 8, 32K ctx) ---
N=1  agg=84.7   per-req=84.7   p50=7.70s
N=2  agg=167.5  per-req=83.8   p50=9.29s
N=4  agg=326.8  per-req=81.7   p50=8.05s
N=8  agg=638.3  per-req=83.9   p50=8.50s   (flat per-req = textbook continuous batching, no spec)

# Boot ~40-54s (FlashInfer SM120 FP4 autotune cache shipped in image). ~29 GiB resident.
# DSpark is single-stream/low-concurrency by design: the qwen38-27b build caps DSpark at bs=2.