| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from vllm.logger import init_logger |
| 7 | from vllm.platforms.cpu import CpuPlatform |
| 8 | |
| 9 | logger = init_logger(__name__) |
| 10 | |
| 11 | |
| 12 | class ZenCpuPlatform(CpuPlatform): |
| 13 | """CPU platform with AMD Zen (ZenDNN/zentorch) optimizations. |
| 14 | |
| 15 | Model-load time (dispatch_cpu_unquantized_gemm in layers/utils.py): |
| 16 | - Routes linear ops to zentorch_linear_unary. |
| 17 | - When VLLM_ZENTORCH_WEIGHT_PREPACK=1 (default), eagerly prepacks |
| 18 | weights via zentorch_weight_prepack_for_linear. |
| 19 | """ |
| 20 | |
| 21 | device_name: str = "cpu" |
| 22 | device_type: str = "cpu" |
| 23 | |
| 24 | def is_zen_cpu(self) -> bool: |
| 25 | # is_cpu() also returns True for this platform (inherited from CpuPlatform). |
| 26 | return True |
| 27 | |
| 28 | # Currently, AMD CPUs do not support float16 compute. |
| 29 | # Hence explicitly return bfloat16 and float32. |
| 30 | @property |
| 31 | def supported_dtypes(self) -> list[torch.dtype]: |
| 32 | return [torch.bfloat16, torch.float32] |