tpu.py 2 KB
Newer Older
1
from typing import TYPE_CHECKING
2

3
4
import torch

5
6
7
from vllm.logger import init_logger

from .interface import Platform, PlatformEnum, _Backend
8

9
10
11
12
if TYPE_CHECKING:
    from vllm.config import VllmConfig
else:
    VllmConfig = None
13

14
15
logger = init_logger(__name__)

16
17
18

class TpuPlatform(Platform):
    _enum = PlatformEnum.TPU
19
    device_type: str = "tpu"
20

21
22
23
24
25
26
    @classmethod
    def get_default_attn_backend(cls, selected_backend: _Backend) -> _Backend:
        if selected_backend != _Backend.PALLAS:
            logger.info("Cannot use %s backend on TPU.", selected_backend)
        return _Backend.PALLAS

27
28
29
30
    @classmethod
    def get_device_name(cls, device_id: int = 0) -> str:
        raise NotImplementedError

31
32
33
34
    @classmethod
    def get_device_total_memory(cls, device_id: int = 0) -> int:
        raise NotImplementedError

35
36
    @classmethod
    def inference_mode(cls):
37
        return torch.no_grad()
38
39
40
41
42

    @classmethod
    def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
        from vllm.config import CompilationLevel
        compilation_config = vllm_config.compilation_config
43
44
        if compilation_config.level == CompilationLevel.NO_COMPILATION:
            # TPU does not support NO_COMPILATION
45
46
47
            compilation_config.level = CompilationLevel.DYNAMO_ONCE
        assert compilation_config.level < CompilationLevel.PIECEWISE,\
            "TPU does not support Inductor."
48
49
50

        if compilation_config.backend == "":
            compilation_config.backend = "openxla"
51
52
53
54
55
56
57
58
59
60
61
62

        assert vllm_config.speculative_config is None, \
            "TPU does not support speculative decoding"

        parallel_config = vllm_config.parallel_config
        scheduler_config = vllm_config.scheduler_config
        if parallel_config.worker_cls == "auto":
            if scheduler_config.is_multi_step:
                parallel_config.worker_cls = \
                    "vllm.worker.multi_step_tpu_worker.MultiStepTPUWorker"
            else:
                parallel_config.worker_cls = "vllm.worker.tpu_worker.TPUWorker"