wehub-resource-sync
33 lines
1021 B
Python
33 lines
1021 B
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import torch
|
|
|
|
from vllm.logger import init_logger
|
|
from vllm.platforms.cpu import CpuPlatform
|
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
class ZenCpuPlatform(CpuPlatform):
|
|
"""CPU platform with AMD Zen (ZenDNN/zentorch) optimizations.
|
|
|
|
Model-load time (dispatch_cpu_unquantized_gemm in layers/utils.py):
|
|
- Routes linear ops to zentorch_linear_unary.
|
|
- When VLLM_ZENTORCH_WEIGHT_PREPACK=1 (default), eagerly prepacks
|
|
weights via zentorch_weight_prepack_for_linear.
|
|
"""
|
|
|
|
device_name: str = "cpu"
|
|
device_type: str = "cpu"
|
|
|
|
def is_zen_cpu(self) -> bool:
|
|
# is_cpu() also returns True for this platform (inherited from CpuPlatform).
|
|
return True
|
|
|
|
# Currently, AMD CPUs do not support float16 compute.
|
|
# Hence explicitly return bfloat16 and float32.
|
|
@property
|
|
def supported_dtypes(self) -> list[torch.dtype]:
|
|
return [torch.bfloat16, torch.float32]
|