fix: make the package importable on Windows

[JoaoZaokk]

Two unrelated defects, both of which stop import deepcompressor — and therefore deepcompressor.app.diffusion.ptq — before any work begins. Neither is Windows-specific by design; both are places where a platform default got assumed.

1. csrc/load.py builds with GCC flags

MSVC ignores them with D9002 warnings, which alone would be cosmetic. But CUDA ≥ 13 ships CCCL, and CCCL rejects MSVC's traditional preprocessor outright:

cccl/cuda/std/__cccl/preprocessor.h(23): fatal error C1189: #error:  MSVC/cl.exe with
traditional preprocessor is used. This may lead to unexpected compilation errors. Please
switch to the standard conforming preprocessor by passing `/Zc:preprocessor` to cl.exe.

This is not an opt-in build step a user can skip. The extension is JIT-built at import time along deepcompressor/data/__init__.py → dtype.py → codebook.py → csrc/load.py, so the package does not import at all.

Fixed by passing /Zc:preprocessor to cl.exe directly (for pybind.cpp) and through nvcc via -Xcompiler (for quantize.cu), with MSVC equivalents substituted for the GCC flags. -fopenmp is part of that substitution on purpose: being silently ignored means the C++ side builds without OpenMP, which is quiet rather than harmless.

2. longbench/eval.py reads a UTF-8 file with the locale default

task2prompt.json contains CJK text. Without an explicit encoding the read uses cp1252 on Windows and raises at import:

UnicodeDecodeError: 'charmap' codec can't decode byte 0x9d in position 1522: character maps to <undefined>

The sibling write on line 123 already passes encoding="utf-8"; only this read was missed.

Worth noting the blast radius: an LLM benchmark file that no diffusion workflow will ever read prevents app.diffusion.ptq from importing, because the diffusion PTQ entry point imports the LLM one.

Verification

Windows 11, Python 3.13, torch 2.13.0+cu130, CUDA 13.2, MSVC 14.44 (VS 2022 Build Tools), RTX 3090 (sm_86):

	before	after
JIT build	fatal error C1189	compiles for sm_86
from deepcompressor.app.diffusion import ptq	UnicodeDecodeError	succeeds
python -m deepcompressor.app.diffusion.ptq --help	never reached	prints usage

Beyond importing, the pipeline was exercised end to end on SANA 1.6B: calibration collected 128/128 samples, and PTQ ran into the smoothing stage before being stopped for time on this hardware. So these two changes are what stood between the package and a working run on Windows, not the only thing that would have to work afterwards.

Linux behaviour is unchanged — both changes sit inside a sys.platform guard or add an argument that was already the implied default.

🤖 Generated with Claude Code