PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
fix sparge on torch2.13 (#2130)
Yi Liud8a300596a07
diff --git a/auto_round_extension/ark/auto_round_kernel/sparge_preprocess_triton.py b/auto_round_extension/ark/auto_round_kernel/sparge_preprocess_triton.pyindex 5306cf4e8..d94263c47 100644--- a/auto_round_extension/ark/auto_round_kernel/sparge_preprocess_triton.py+++ b/auto_round_extension/ark/auto_round_kernel/sparge_preprocess_triton.py@@ -12,6 +12,35 @@ import triton.language as tllogger = logging.getLogger(__name__)+_APPLIED_TRITON_PREDICATED_CHECK = False+++def _apply_xpu_triton_workarounds() -> None:+ """Patch triton's Intel compiler so JIT kernels avoid the rejected SPIR-V extension."""+ global _APPLIED_TRITON_PREDICATED_CHECK+ if _APPLIED_TRITON_PREDICATED_CHECK:+ return+ try:+ import triton.backends.intel.compiler as _intel_compiler+ except ImportError:+ # No intel triton backend (e.g. stock triton); nothing to patch.+ return++ _orig_parse_target = _intel_compiler.XPUBackend.parse_target++ def _patched_parse_target(self, tgt_prop):+ dev_prop = _orig_parse_target(self, tgt_prop)+ dev_prop["has_predicated_io"] = False+ return dev_prop++ _intel_compiler.XPUBackend.parse_target = _patched_parse_target+ _APPLIED_TRITON_PREDICATED_CHECK = True+ logger.info("Applied XPU triton workaround: has_predicated_io forced off (SPV_INTEL_predicated_io)")+++_apply_xpu_triton_workarounds()++_TRITON_FALLBACK_WARNING_LOGGED = False