@@ -475,13 +475,7 @@ std::optional<NPUDesc> extract_npu_descriptor(ov::Core& core) {
475
475
}
476
476
const auto arch = core.get_property (" NPU" , ov::device::architecture);
477
477
const auto max_tiles = core.get_property (" NPU" , ov::intel_npu::max_tiles);
478
-
479
- bool compiler_dq = false ;
480
- const auto device_caps = core.get_property (" NPU" , ov::device::capabilities);
481
- if (std::find (device_caps.begin (), device_caps.end (),
482
- " COMPILER_DYNAMIC_QUANTIZATION" ) != device_caps.end ()) {
483
- compiler_dq = true ;
484
- }
478
+ const auto compiler_dq = core.get_property (" NPU" , ov::intel_npu::compiler_dynamic_quantization);
485
479
return std::make_optional (NPUDesc{arch, max_tiles, compiler_dq});
486
480
}
487
481
@@ -526,6 +520,7 @@ ov::AnyMap get_default_prefill_config(const std::shared_ptr<ov::Model>& model,
526
520
}
527
521
if (npudesc.has_value () && npudesc->compiler_dq ) {
528
522
config.emplace (" NPUW_DQ_FULL" , " NO" );
523
+ config.emplace (" NPU_COMPILATION_MODE_PARAMS" , " enable-weights-dynamic-dequantization=true" );
529
524
}
530
525
return config;
531
526
}
@@ -547,6 +542,7 @@ ov::AnyMap get_default_generate_config(const std::shared_ptr<ov::Model>& model,
547
542
}
548
543
if (npudesc.has_value () && npudesc->compiler_dq ) {
549
544
config.emplace (" NPUW_DQ_FULL" , " NO" );
545
+ config.emplace (" NPU_COMPILATION_MODE_PARAMS" , " enable-weights-dynamic-dequantization=true" );
550
546
}
551
547
return config;
552
548
}
0 commit comments