已合并
fix: adapt Inductor compile mode for TA3.6 #45315
luqichao创建于 23 天前
fix: adapt Inductor compile mode for TA3.6 #45315
已合并
共 2 个文件变更+7-3
| @@ -34,7 +34,7 @@ class NPUKernelType(Enum): | |||
| 34 | 34 | ||
| 35 | def compile_mode(self): | 35 | def compile_mode(self): |
| 36 | if self == NPUKernelType.SIMT_TEMPLATE: | 36 | if self == NPUKernelType.SIMT_TEMPLATE: |
| 37 | - return "unstructured_in_simt" | 37 | + return "simd_simt_template" |
| 38 | return str(self) | 38 | return str(self) |
| 39 | 39 | ||
| 40 | 40 | ||
| @@ -825,7 +825,9 @@ class NPUCachingAutotuner(CachingAutotuner): | |||
| 825 | and not self.inductor_meta.get("is_hip", False) | 825 | and not self.inductor_meta.get("is_hip", False) |
| 826 | ) | 826 | ) |
| 827 | 827 | ||
| 828 | - compile_meta['compile_mode'] = cfg_kwargs.get('compile_mode') | 828 | + compile_meta['compile_mode'] = ( |
| 829 | + cfg_kwargs.get('compile_mode') or NPUKernelType.SIMT_TEMPLATE.compile_mode() | ||
| 830 | + ) | ||
| 829 | 831 | ||
| 830 | # device type will be "hip" rather than "cuda" here | 832 | # device type will be "hip" rather than "cuda" here |
| 831 | compile_meta["device_type"] = self.device_props.type | 833 | compile_meta["device_type"] = self.device_props.type |
| @@ -1261,7 +1263,9 @@ class NPUCachingAutotuner(CachingAutotuner): | |||
| 1261 | and not self.inductor_meta.get("is_hip", False) | 1263 | and not self.inductor_meta.get("is_hip", False) |
| 1262 | ) | 1264 | ) |
| 1263 | 1265 | ||
| 1264 | - compile_meta['compile_mode'] = cfg_kwargs.get('compile_mode') | 1266 | + compile_meta['compile_mode'] = ( |
| 1267 | + cfg_kwargs.get('compile_mode') or NPUKernelType.SIMT_TEMPLATE.compile_mode() | ||
| 1268 | + ) | ||
| 1265 | 1269 | ||
| 1266 | # device type will be "hip" rather than "cuda" here | 1270 | # device type will be "hip" rather than "cuda" here |
| 1267 | compile_meta["device_type"] = self.device_props.type | 1271 | compile_meta["device_type"] = self.device_props.type |