| """Process environment the worker's numerics depend on; imported by the package before any exllamav3 call. | |
| The Spark serve runs with EXL3_INT8_GEMV=0 (exllamav3 1.6.0 otherwise sends one-row mul1 GEMVs through a plain-int8 | |
| activation path, ~0.7 % output error measured on these tensors, offload/results/o2/d1_exl3_identity.json) and | |
| EXL3_MOE_COOP_WIDE=1 (the wide coop tile; already exllamav3's default below Blackwell). exllamav3 reads both with | |
| getenv on first use, so they are forced here, at import, before the extension runs anything. | |
| """ | |
| import logging | |
| import os | |
| FORCED = {"EXL3_INT8_GEMV": "0", "EXL3_MOE_COOP_WIDE": "1"} | |
| log = logging.getLogger("dspark_worker") | |
| for _k, _v in FORCED.items(): | |
| _old = os.environ.get(_k) | |
| if _old not in (None, _v): | |
| log.warning("%s=%s overridden to %s (the Spark's setting)", _k, _old, _v) | |
| os.environ[_k] = _v | |