return moe method based on device platform
(layer=None)
| 43 | |
| 44 | |
| 45 | def get_moe_method(layer=None): |
| 46 | """ |
| 47 | return moe method based on device platform |
| 48 | """ |
| 49 | |
| 50 | if current_platform.is_cuda() or current_platform.is_iluvatar(): |
| 51 | from .fused_moe_cutlass_backend import CutlassMoEMethod |
| 52 | |
| 53 | return CutlassMoEMethod(None) |
| 54 | elif current_platform.is_xpu(): |
| 55 | from fastdeploy.model_executor.layers.backends import XPUMoEMethod |
| 56 | |
| 57 | return XPUMoEMethod(None, layer) |
| 58 | elif current_platform.is_gcu(): |
| 59 | from fastdeploy.model_executor.layers.backends import GCUFusedMoeMethod |
| 60 | |
| 61 | return GCUFusedMoeMethod(None) |
| 62 | |
| 63 | elif current_platform.is_intel_hpu(): |
| 64 | from fastdeploy.model_executor.layers.backends import HpuMoEMethod |
| 65 | |
| 66 | return HpuMoEMethod(None) |
| 67 | |
| 68 | elif current_platform.is_maca(): |
| 69 | from fastdeploy.model_executor.layers.backends import ( |
| 70 | MetaxCutlassUnquantizedFusedMoEMethod, |
| 71 | ) |
| 72 | |
| 73 | return MetaxCutlassUnquantizedFusedMoEMethod(None) |
| 74 | return None |
| 75 | |
| 76 | |
| 77 | def get_moe_scores( |
no test coverage detected