Environment variables

TensorPlay reads a small set of environment variables to adjust runtime behavior: where downloaded files are cached, whether out-of-tree device backends load at import, how errors are reported, and how the experimental graph translator validates itself.

Unless noted otherwise, set a variable before the process starts (or at least before the module that reads it is imported — several are read once and cached). Values are strings; 1 and 0 enable and disable.

Paths and caching

Variable	Default	Effect
`TENSORPLAY_HOME`	`~/.cache/tensorplay/datasets/vision`	Root directory for weights downloaded by the vision model loaders.

TENSORPLAY_HOME is consulted by tensorplay.vision.load_state_dict_from_url (and the model zoo helpers built on it) when deciding where to cache downloaded checkpoints. The hub cache used by tensorplay.hub is a different tree — it lives under ~/.cache/tensorplay and is relocated through the tensorplay.hub.set_dir API, not through this variable.

Device and CUDA

Variable	Default	Effect
`TENSORPLAY_DEVICE_BACKEND_AUTOLOAD`	`1`	Import out-of-tree device backends at `import tensorplay`.
`TENSORPLAY_NVML_BASED_CUDA_CHECK`	unset	Make `cuda.is_available()` use NVML.
`TENSORPLAY_DEVICE_NAME`	`cpu`	Device key for the sparse operator tuning store.

Setting TENSORPLAY_DEVICE_BACKEND_AUTOLOAD=0 skips the out-of-tree backend scan during import, for startup time or isolation. The current state is visible through the helper the import path itself uses:

import tensorplay
tensorplay._is_device_backend_autoload_enabled()
# True, unless the variable is set to something other than 1

With TENSORPLAY_NVML_BASED_CUDA_CHECK=1, tensorplay.cuda.is_available() probes the driver through NVML instead of initializing a CUDA context — useful because a context created before fork poisons the child. The check is therefore fork-safe under this flag.

TENSORPLAY_DEVICE_NAME selects the device-name key under which the sparse operator tuning metadata is stored and looked up; set it to the accelerator in use (e.g. a GPU name) to read the matching tuning entries.

Debugging

Variable	Default	Effect
`TENSORPLAY_SHOW_CPP_STACKTRACES`	unset	Attach captured native stack traces to errors raised from C++.

When set to 1, errors raised on the C++ side of the library carry the native stack trace captured at the throw site. The flag is read once at first use and cached — set it before the first error occurs, which in practice means before the workload starts.

Serialization

Variable	Default	Effect
`TENSORPLAY_SERIALIZATION_WORKERS`	`min(4, cpu count)`	Upper bound on threads staging MEGA tensor payloads.

When saving a MEGA checkpoint, the tensor payloads are staged by a thread pool whose size is min(TENSORPLAY_SERIALIZATION_WORKERS, number of tensors) — never more threads than tensors, and never more than requested. The default uses at most four workers regardless of core count. An invalid (non-integer) value falls back to the default.

Graph translation (experimental)

Variable	Default	Effect
`TENSORPLAY_TRANSLATION_VALIDATION`	`0`	Validate translated graphs against the original.
`TENSORPLAY_TRANSLATION_VALIDATION_TIMEOUT`	`600000`	Wall-clock budget for validation, in milliseconds.
`TENSORPLAY_TRANSLATION_NO_BISECT`	`0`	Skip bisection when validation fails.

These configure the experimental graph translator’s self-check: when TENSORPLAY_TRANSLATION_VALIDATION=1, each translated program’s guards are checked symbolically (via sympy, and z3 when installed) to admit exactly the executions of the original program. The timeout bounds the symbolic check; when a mismatch is found, a bisection pass pinpoints the first divergent guard unless TENSORPLAY_TRANSLATION_NO_BISECT=1 skips it. All three are read when tensorplay.graph.experimental is imported.