Long-music inference benchmark

Run 20261101T172743Z-89688399, revision bbc19f2-dirty, NVIDIA GeForce RTX 5070 Ti. Image sha256:2c8ed04ae8f7960b8bf46b57c06d9017fbcefde66f8f43f79731b1db9eabb53d. Assessment: complete.

English, original long prompt, 300-second duration ceiling, eight steps and normal 24576 context. Models remain resident; every timed call requires zero weight loads. This suite makes no ASR requests and produces no lyric-quality score.

Warmup mode: short. 5 warm measured long songs. The first long request after a short warmup can include allocation/graph setup for the longer shapes; these costs remain in its timing. Duration is a ceiling; actual output lengths are measured rather than padded or stretched.

Seed Audio s Pipeline s Worker s Client s Audio s / generation s Device peak MiB Loads
831001 286.24 63.15 63.43 67.56 4.53 13142 0
831002 300.00 69.37 69.64 73.72 4.32 13089 0
831003 300.00 70.03 70.32 74.33 4.28 13191 0
831004 300.00 70.58 70.92 75.43 4.25 13418 0
831005 295.92 72.40 72.74 77.16 4.09 13329 0
Summary Result
Pipeline median / p95 70.03 / 72.03 s
Worker median / client median 70.32 / 74.33 s
Aggregate audio s / generation s 4.29
Sampled device peak / peak above idle 13418 / 10821 MiB
Retained weight buffers 3755 MiB
Total suite wall time 430.96 s
Cold/warmup requests 27.48 s
Measured client requests 368.19 s
Other setup/cleanup/audio analysis 35.29 s

Pipeline excludes output encoding/packaging and HTTP transfer. Worker adds encoding/packaging; client includes the complete download. Sampled nvidia-smi peaks are device-wide lower bounds, including desktop/driver allocations, not exact process/allocator peaks. Total ends after audio integrity analysis; report export is excluded.

Five samples are a controlled local reference, not statistical proof or a before/after speedup claim. This profile is separate from the 30-second multilingual suite.