Nano-vLLM-VoxCPM 2.0.5 Synchronization

This benchmark reviews Nano-vLLM-VoxCPM 2.0.5 plus the merged upstream block-cache fix from PR 89. Performance is compared against the established 2.0.4 baseline from run 20261007-184048. A same-day 2.0.4 run was retained for paired output-equivalence evidence but excluded from performance conclusions because its timing and RAM measurements were clear outliers.

Result

Signal Established 2.0.4 2.0.5 + PR 89 Change
Median RTF 0.2707 0.2707 unchanged
p95 RTF 0.3153 0.3294 +4.5%
Total synthesis time 467.698 s 461.444 s -1.3%
Aggregate realtime speed 3.70x 3.75x +1.4%
Design median RTF 0.2754 0.2769 +0.5%
Clone median RTF 0.2581 0.2510 -2.8%
Clone total synthesis time 199.014 s 192.782 s -3.1%
Peak VRAM delta 8,491 MiB 8,330 MiB -161 MiB
Final VRAM delta 7,533 MiB 7,288 MiB -245 MiB
Errors 0 0 unchanged

The correct performance conclusion is parity. Overall median speed is identical, total synthesis time differs by only 1.3%, and the small clone gain is balanced by a 4.5% p95 regression. These differences are modest enough to be normal run-to-run variation rather than evidence of a broad speed change.

Container RAM is not compared against the established run. The application, dependency set, and image evolved substantially between October 7 and October 22, so its increase from 3,376.5 MiB to 7,870.3 MiB cannot be attributed to Nano-vLLM 2.0.5. In the same-day pair, 2.0.5 measured 398 MiB more peak VRAM and 349.3 MiB more peak container RAM than 2.0.4, but the 2.0.4 run also had outlier timing and RAM behavior. Additional repeated runs are required before claiming a package-level memory change.

Output Equivalence

The same-day runs 20261022-052037 and 20261022-054748 used matching stable seeds for every call. All 312 corresponding WAV files had identical SHA-256 hashes and durations. All 312 ASR similarity and exact-match results were also identical. The synchronization therefore preserved generated output exactly under this controlled concurrency-one workload.

The established baseline used unseeded generation, so its aggregate ASR score must not be compared directly with the seeded October 22 run. The paired same-day result is the valid quality comparison.

Decision

Adopt 2.0.5 plus PR 89. It preserves output, maintains established inference performance, improves worker-failure handling and first-audio internals, and bounds the prefix-cache hash map for long-running servers. The decision does not depend on a claimed speedup.

Controls

  • Hardware: NVIDIA GeForce RTX 5070 Ti, driver 610.88.
  • Model: pinned openbmb/VoxCPM2 revision 32279effe8c19989596f05d353d1447f51d9e915.
  • Workload: 312 calls across 24 common VoxCPM2 and Qwen3-ASR languages.
  • Modes: 240 voice-design calls and 72 reference-clone calls.
  • Runtime: one container, one visible GPU, concurrency one, 10 inference steps, WAV output, no denoising, no text normalization, and no output conversion.
  • The October 22 pair used a stable 32-bit seed derived from each call ID with randomization disabled.
  • The same LAN Qwen3-ASR service judged staged audio after TTS measurements.

Images And Runs

Role Run Image ID Package
Performance reference 20261007-184048 sha256:3a4e1ccda4af76e389521374637905c6462974480dac20c476da9414255923f1 nano-vllm-voxcpm==2.0.4
Paired output reference 20261022-052037 sha256:896881513b15712f99bc5a38bc15c717816fe3b730a90cd5790e23146c60f91a nano-vllm-voxcpm==2.0.4
Candidate 20261022-054748 sha256:62a9f53e11c5c5f5d40a745eadd2ea653f524538ec806269c0019e5591e210aa nano-vllm-voxcpm==2.0.5 plus PR 89

The 2.0.5 wheel was the published pure-Python wheel with SHA-256 ae37a585df30d6d708d56c224e7cd9050868462d1203a9e2e92432c2d4193174. No dependency wheel was compiled. The image retains VoxCPMTTS’s guarded reference-latent allocator cleanup and corrected KV-cache guidance patches. PR 89 bounds the prefix-cache hash map by evicting stale entries when blocks are recycled.

Command

The paired images were run separately with the same command shape:

python benchmarks/run_baseline.py `
  --backends nano `
  --nano-image <preserved-image-tag> `
  --container voxcpmtts_benchmark `
  --port 8811 `
  --gpu-index 0 `
  --cache-volume voxcpmtts_hf_cache `
  --asr-url <qwen3-asr-url> `
  --comment "<run description>"

If another model service occupies the selected GPU, stop it before the run or pass --stop-container <your-container-name> to pause and restore that service.

Raw measurements are retained locally in the git-ignored canonical baseline history. This committed report preserves the reviewed comparison, immutable image IDs, controls, and key metrics without duplicating multi-megabyte JSON across benchmark suites.