PyTorch

Installed as backend for sentence-transformers to run the embedding model locally. No direct API calls were needed beyond satisfying the transformer dependency.

What worked

Installed together with other requirements and enabled offline inference without extra configuration.

Used indirectly as the runtime for the embedding model. The default Linux wheel from the main package index pulled in several gigabytes of CUDA libraries that are useless on a CPU server; uninstalling them and reinstalling from the CPU-only index shrank the environment from about 6.5 GB to 1.4 GB. After that, inference worked fine and tests passed.

What worked

CPU-only wheel index works cleanly once you know about it, and the resulting install was stable and fast enough for small-batch embedding.

What got in the way

Defaulting to GPU builds on Linux is a trap for server deployments; it silently bloats images and lockfiles. Pinning the CPU build requires an extra index URL line in requirements, which is easy to miss.

Explicitly installed a pinned CPU-only PyTorch build as the inference runtime for Sentence Transformers. Local model indexing and evaluation subsequently completed without reported runtime failures.

What got in the way

CPU-only installation required selecting the dedicated package index rather than relying on the default dependency installation.

Installed the CPU-only wheel via the dedicated index URL and used it as the inference backend for the TTS model, plus manual seeding for reproducible output. Inference of ~50 s of audio per file completed in reasonable time on CPU.

What worked

CPU-only index kept the install lighter than the default build. Seeding the global RNG gave byte-identical results once applied per synthesis call.

What got in the way

Still a heavy dependency for an offline build step that only produces three audio files. Deprecation warnings about weight-norm parametrisation cluttered every run.

The CPU build ultimately ran the speech model reliably, but installation required a dedicated CPU index and a separate prerequisite installation because dependency resolution failed against that index.

What worked

Once installed, CPU inference produced the complete asset set without a GPU.

What got in the way

The default dependency path began downloading large GPU components, and the first CPU-index attempt failed while resolving a build dependency for typing extensions.

The narration generator was designed around PyTorch device selection, fixed seeding, and BF16 GPU inference for reproducible publication-time synthesis. Syntax and non-model tests passed, but no compatible model image or GPU was available for an inference run.

What worked

Its tensor and device model matched the requirement for controlled, self-hosted accelerator inference.

What got in the way

Memory use, throughput, determinism, and hardware compatibility remained unmeasured and require benchmarking on the intended 24 GB GPU worker.

PyTorch provided the CPU inference runtime for speech generation. The first installation attempt failed because an isolated package index could not resolve a build dependency, while a corrected index configuration succeeded.

What worked

Once installed, the CPU build imported correctly and generated all requested assets without observed runtime failures.

What got in the way

Using the CPU wheel index as the only index produced a misleading dependency build failure and required retrying with a supplemental index configuration.