globals.py 1.42 KB
Newer Older
1
import torch
Nicolas Patry's avatar
Nicolas Patry committed
2
import os
3
from loguru import logger
drbh's avatar
drbh committed
4
from typing import Dict
5

6
MEM_POOL = torch.cuda.graph_pool_handle() if torch.cuda.is_available() else None
Nicolas Patry's avatar
Nicolas Patry committed
7
# This is overridden by the cli
8
9
10
11
12
13
FLASH_DECODING = os.getenv("FLASH_DECODING") in {"1", "true", "True"}
BLOCK_SIZE: int = 256 if FLASH_DECODING else 16
if FLASH_DECODING:
    logger.info("Using FLASH_DECODING")


14
cuda_graphs = os.getenv("CUDA_GRAPHS")
Nicolas Patry's avatar
Nicolas Patry committed
15
if cuda_graphs is not None:
16
17
18
19
20
21
    try:
        cuda_graphs = [int(item) for item in cuda_graphs.split(",")]
    except Exception as e:
        raise RuntimeError(
            f"Could not parse cuda graphs {cuda_graphs}, expected comma separated list for batch sizes to run on: {e}"
        )
22
23
else:
    cuda_graphs = None
24
25
26
27
28
29
# sorting the cuda graphs in descending order helps reduce the
# memory impact and results in less memory usage
if cuda_graphs is not None:
    cuda_graphs.sort(reverse=True)


30
CUDA_GRAPHS = cuda_graphs
fxmarty's avatar
fxmarty committed
31
32
33
34
35
36
37
38
39

# This is overridden at model loading.
global MODEL_ID
MODEL_ID = None


def set_model_id(model_id: str):
    global MODEL_ID
    MODEL_ID = model_id
drbh's avatar
drbh committed
40
41
42
43
44
45
46
47
48
49
50


# NOTE: eventually we should move this into the router and pass back the
# index in all cases.
global ADAPTER_TO_INDEX
ADAPTER_TO_INDEX: Dict[str, int] = None


def set_adapter_to_index(adapter_to_index: Dict[str, int]):
    global ADAPTER_TO_INDEX
    ADAPTER_TO_INDEX = adapter_to_index
Nicolas Patry's avatar
Nicolas Patry committed
51
52
53
54
55


def get_adapter_to_index():
    global ADAPTER_TO_INDEX
    return ADAPTER_TO_INDEX