Virtual-LPU / lpu /compat.py
iparaco's picture
Upload 118 files
e59bd9d verified
Raw History Blame Contribute Delete
3.72 kB
"""Optional-dependency imports.
LPU is split into two runtime footprints:
1. The background controller / process discovery / monitoring stack runs with
NO torch at all (psutil + PyYAML + nvidia-ml-py only, ~20 MB venv). This is
a deliberate architectural decision so LPU never installs or upgrades
anything inside a user's AI environment.
2. The GPU/training/benchmark stack imports torch lazily and only inside the
*target project's* interpreter, where torch already exists.
Every torch-touching module must use these helpers instead of a bare
``import torch`` so the controller footprint stays torch-free.
"""
from __future__ import annotations
import importlib
import importlib.util
import sys
from types import ModuleType
from typing import Any, Optional
# --- lazy torch ------------------------------------------------------------
# TORCH_AVAILABLE is a CHEAP spec probe (no import). `compat.torch` is resolved
# lazily on first attribute access via module __getattr__, so importing lpu
# never loads torch: the controller stays ~15 MB and every hooked process
# avoids seconds of torch DLL load at interpreter startup.
def _torch_available() -> bool:
try:
return importlib.util.find_spec("torch") is not None
except (ImportError, ValueError, ModuleNotFoundError):
return False
TORCH_AVAILABLE = _torch_available()
def __getattr__(name: str):
if name == "torch":
t = importlib.import_module("torch")
globals()["torch"] = t # cache: subsequent lookups skip __getattr__
return t
raise AttributeError(f"module {__name__!r} has no attribute {name!r}")
def optional_import(name: str) -> Optional[ModuleType]:
"""Import a module, returning None when unavailable. Never raises."""
try:
return importlib.import_module(name)
except Exception: # noqa: BLE001 - optional by design
return None
def require_torch() -> ModuleType:
"""Return the torch module or raise an actionable error.
Used by GPU/training/benchmark entry points. The error message tells the
operator exactly which interpreter needs torch (the target project's
interpreter, never an LPU-owned venv).
"""
try:
return importlib.import_module("torch")
except ImportError:
raise RuntimeError(
"PyTorch is required for this LPU component but is not importable "
"in the current interpreter. LPU never installs packages into your "
"environment. Run this component with the SAME Python interpreter "
"your AI project uses (the one where torch is already installed), "
"for example:\n"
" C:/path/to/your/project/venv/Scripts/python.exe -m lpu doctor\n"
"or pass it explicitly:\n"
" lpu-run --python C:/path/to/venv/Scripts/python.exe train.py"
) from None
def torch_version() -> Optional[str]:
"""Return the torch version string, or None when torch is absent."""
try:
return importlib.import_module("torch").__version__
except Exception: # noqa: BLE001
return None
def pynvml():
"""Return an initialized pynvml module, or None when NVML is unavailable.
nvidia-ml-py provides the ``pynvml`` import name. NVML comes from the
NVIDIA display driver, so a healthy driver is a prerequisite. All callers
must handle None and report the metric as unavailable - never as zero.
"""
nv = optional_import("pynvml")
if nv is None:
return None
try:
nv.nvmlInit()
return nv
except Exception: # noqa: BLE001 - driver may be absent
return None
def is_windows() -> bool:
return sys.platform.startswith("win")