[0xShug0]

Summary

Use gallocr unconditionally for the shared WavLM encoder instead of allocating storage for every context tensor. This lets non-overlapping intermediates reuse memory without changing model math or introducing an opt-in setting.

• Preserve persistent position bias and all requested hidden-layer outputs across graph executions.
• Create and upload masks only for the normalization/attention paths that consume them.
• Release captured CUDA graph resources before freeing an allocator.
• Keep existing session-owned graph caching and reuse.

One framework source file changes. No model-local, public-option, spec, kernel, or build changes.

A/B Results

Baseline: main 977dc40a. Debug builds (-O3 -g), RTX 5090 / Ryzen 9 7900X, eight threads. Existing MioCodec/MioTTS original GGUFs and MeanVC2 FP32. Sequential runs using the existing CLI path-test tool, conda qwen3-tts, --log, and 20 ms process memory sampling.

Backend	Model	Peak Memory Before (MiB)	Peak Memory After (MiB)	Warm RTF Before	Warm RTF After
CUDA	MioCodec	4,108	1,748	0.00605	0.00546
CUDA	MioTTS	6,598	5,660	0.10596	0.10341
CUDA	MeanVC2	4,776	2,192	0.03692	0.03662
Vulkan	MioCodec	3,466	1,168	0.01741	0.01126
Vulkan	MioTTS	5,973	5,034	0.11277	0.10976
Vulkan	MeanVC2	4,182	1,599	0.05697	0.05065
CPU	MioCodec	3,797	1,493	0.17578	0.11774
CPU	MioTTS	6,494	5,556	1.86485	1.83905
CPU	MeanVC2	4,462	1,881	0.21633	0.15403

GPU memory is sampled process peak VRAM; CPU memory is sampled peak RSS. Peaks cover the entire session. Warm RTF aggregates inference wall time/output duration after excluding the first warmup request. No measured performance regression; small timing differences are not statistical speedup claims.

Correctness And Lifecycle

Each model was tested in one session with a short warmup, a normal request, an exact repeat, a changed shorter reference/input, and a longer reference/input. MioTTS additionally changes the reference recording without changing its length, exercising a reused WavLM graph on a reference-cache miss. Audio lengths span 1.6, 3.2, and 5.95 seconds.

• CPU and Vulkan: all 32 output WAVs are byte-identical to baseline.
• CUDA MeanVC2: all five output WAVs are byte-identical.
• CUDA MioCodec/MioTTS: output lengths match; worst waveform RMSE is 2.632e-7, minimum cosine similarity is 0.99999999999399. This is not byte-exact parity. Baseline CUDA repeated requests already show comparable tiny waveform variation (RMSE 2.437e-7 / 1.611e-7).
• No non-finite samples, missing outputs, duration changes, or new repeat/changing-length failures in the final candidate.

Plapre is not included in this validation. These are focused C++ allocation A/B checks, not new Python parity or ASR evaluations.