voice_audio_multi/docs/reports/ace_step_xl_vulkan.md

XL ACE-Step Vulkan validation

RTX 3090 (24 GiB), Windows Release, NVIDIA Vulkan device 1, eight CPU threads, one session. Candidate: PR #737 at 48156adc. The control uses the same PR server/framework/build configuration, with only the three ACE-Step model files restored to upstream ed96b730; the optimized model files were restored byte-for-byte afterward. No compilation or other GPU workload overlapped these measurements.

Both XL Turbo and XL SFT Q8-DiT packages use bf16 text/planner/VAE. Same supplied long English caption, seed 1234, five-second direct generation and eight diffusion steps. Warm medians are requests 2?8. Each run adds six mixed-history requests: changed caption/lyrics/duration/seed, planner generation, return to the original caption, and guidance=7 for the longer SFT requests. Turbo ignores guidance by design.

Results

56/56 original/optimized response artifacts and WAV SHA-256 comparisons are byte-identical, across both packages and memory saver on/off. All 112 server requests returned HTTP 200. Two direct Vulkan conditioning probes passed ten histories each (20 total): repeated/changed reference bytes, frame growth/shrinkage, graph release, multi-reference bypass and invalid-order rejection. Both CLI/server and the conditioning probe build with Vulkan.

Model	Memory saver	Original warm (s)	PR warm (s)	Less time	Original extra warm VRAM (GiB)	PR extra warm VRAM (GiB)	Sequence device peak original / PR (GiB)
XL SFT Q8 DiT	true	0.721	0.620	14.1%	2.373	0.820	13.935 / 12.624
XL SFT Q8 DiT	false	0.620	0.588	5.1%	0.105	0.105	15.706 / 14.425
XL Turbo Q8 DiT	true	0.744	0.618	16.9%	2.373	0.820	12.773 / 12.731
XL Turbo Q8 DiT	false	0.604	0.586	3.1%	0.105	0.105	15.814 / 14.425

Extra warm VRAM is the request's sampled device peak minus its immediately preceding loaded-idle reading, median of requests 2?8. It excludes the loaded-idle state (weights, backend state and retained caches), not literally every non-weight byte. NVML sampled every 10 ms. Memory saver on saves 65.4% of incremental warm-request peak for both variants. Memory saver off retains graphs: incremental warm memory stays the same, while full mixed-sequence device peak falls by about 1.3?1.4 GiB. Whole-sequence peaks include loaded weights, cold requests and larger graphs; they are not the warm-overhead figure. Timing percentages describe this measured workload, not every prompt or duration.

This validates Vulkan one-session generation/conditioning on these Q8-DiT packages. It does not validate parallel slots, editing routes, longer songs, CPU/Metal or the full repository test suite.

Local evidence

• ace_step_xl_vulkan.json (original local name xl-vulkan-summary.json): timings, memory, comparison counts and exact executable SHA-256s.
• vulkan-original-model-manifest.json: original model source revision and executable hash.
• test_xl_vulkan.py: full server matrix and probe invocations.
• probe-vulkan-xl-turbo-q8dit.log, probe-vulkan-xl-sft-q8dit.log: all 20 conditioning histories.
• Raw configs, requests, traces, results and WAVs in the local test workspace: outputs/ace-graph-memory-20260929/xl-vulkan-{original,pr}-{mem,nomem}/.
• Builds: build_pr_vulkan.cmd, build_vulkan_control.py and associated logs. The latter restores candidate source in a finally block and rebuilds it after making the original-model control.

Run the test matrix from the workspace root after building both executables:

python outputs/ace-next-20260929/test_xl_vulkan.py

The test harness starts hidden server subprocesses and stops its own server above a 23.5 GiB device-use guard. No guards fired; no server was left running after the tests. Cleanup is process termination, not proof of graceful server shutdown.