Record LavaSR torch.compile measurements

[Tinnci]

Summary

• add a Colab-only T4 eager/torch.compile benchmark
• record cold/warm latency, peak CUDA allocation, determinism, and numerical drift
• keep torch-eager as the only selectable provider based on measured evidence
• complete P4 without adding an unsupported optimization surface

Colab evidence

Two fresh Tesla T4 sessions, torch 2.11.0+cu128, identical one-second fixture:

• warm speedup contradicted across sessions: 0.96x and 1.82x
• eager first inference: about 2.3-2.4 s
• compile first inference: about 34-35 s
• peak CUDA allocation: 72,180,224 vs 80,844,288 bytes
• compiled repeats were not bit-exact
• eager/compile max absolute drift: about 5.2e-8 to 6.1e-8
• TorchInductor warned that complex operators lack code generation support

Decision: no public torch-compile provider. Require repeatable representative-length benefit before reopening. All real-weight, compile, GPU, and audio computation ran through Colab CLI.

Local verification

• git diff --check
• pixi run lint
• pixi run format-check
• pixi run typecheck
• pixi run test (163 passed, 6 skipped)

[Copilot]

Pull request overview

This PR adds recorded evidence around a Colab T4 torch.compile experiment for LavaSR and documents the resulting decision to keep torch-eager as the only selectable runtime provider.

Changes:

• Add a Colab-only benchmark script that measures eager vs torch.compile latency, peak CUDA allocation, determinism, and numerical drift.
• Document the T4 measurement results and the evidence-based decision not to expose torch-compile as a provider.
• Minor formatting tweak in a LavaSR compatibility test.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_lavasr_compat.py	Reformats a monkeypatch call for readability.
ROADMAP.md	Marks the “Optimize Only After Measurement” priority as completed and records the compile findings.
examples/colab_lavasr_compile_benchmark.py	Adds a Colab T4 benchmark runner that records eager vs torch.compile evidence.
docs/ACCELERATORS.md	Updates accelerator guidance with the compile trial results and reproduction steps.
CHANGELOG.md	Notes the addition of the compile benchmark and the decision to keep eager as the only provider.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.