File size: 3,911 Bytes
bcdf9fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import logging
from typing import Tuple

import torch
import torch.distributed as dist

from verl.utils.logger.aggregate_logger import DecoratorLoggerBase


def _get_current_mem_info(unit: str = "GB", precision: int = 2) -> Tuple[str]:
    """Get current memory usage."""
    assert unit in ["GB", "MB", "KB"]
    divisor = 1024**3 if unit == "GB" else 1024**2 if unit == "MB" else 1024
    mem_allocated = torch.cuda.memory_allocated()
    mem_reserved = torch.cuda.memory_reserved()
    # use torch.cuda.mem_get_info to profile device memory
    # since vllm's sleep mode works below pytorch
    # see https://github.com/vllm-project/vllm/pull/11743#issuecomment-2754338119
    mem_free, mem_total = torch.cuda.mem_get_info()
    mem_used = mem_total - mem_free
    mem_allocated = f"{mem_allocated / divisor:.{precision}f}"
    mem_reserved = f"{mem_reserved / divisor:.{precision}f}"
    mem_used = f"{mem_used / divisor:.{precision}f}"
    mem_total = f"{mem_total / divisor:.{precision}f}"
    return mem_allocated, mem_reserved, mem_used, mem_total


def log_gpu_memory_usage(head: str, logger: logging.Logger = None, level=logging.DEBUG, rank: int = 0):
    if (not dist.is_initialized()) or (rank is None) or (dist.get_rank() == rank):
        mem_allocated, mem_reserved, mem_used, mem_total = _get_current_mem_info()
        message = f"{head}, memory allocated (GB): {mem_allocated}, memory reserved (GB): {mem_reserved}, device memory used/total (GB): {mem_used}/{mem_total}"

        if logger is None:
            print(message)
        else:
            logger.log(msg=message, level=level)


class GPUMemoryLogger(DecoratorLoggerBase):
    """A decorator class to log GPU memory usage.



    Usage:

        For example, in actor function, we initialize a GPUMemoryLogger



        ```

        from verl.utils.debug.performance import GPUMemoryLogger

        @GPUMemoryLogger(role="actor")

        def update_actor(self, batch):

            # do something

            return

        ```



    """

    def __init__(self, role: str, logger: logging.Logger = None, level=logging.DEBUG, log_only_rank_0: bool = True):
        if dist.is_initialized() and dist.get_world_size() > 1:
            rank = dist.get_rank()
        else:
            rank = 0
        super().__init__(role, logger, level, rank, log_only_rank_0)

    def __call__(self, decorated_function: callable):
        def f(*args, **kwargs):
            return self.log(decorated_function, *args, **kwargs)

        return f

    def log(self, func, *args, **kwargs):
        name = func.__name__
        mem_allocated, mem_reserved, mem_used, mem_total = _get_current_mem_info()
        message = f"Before {name}, memory allocated (GB): {mem_allocated}, memory reserved (GB): {mem_reserved}, device memory used/total (GB): {mem_used}/{mem_total}"
        self.logging_function(message)

        output = func(*args, **kwargs)

        mem_allocated, mem_reserved, mem_used, mem_total = _get_current_mem_info()
        message = f"After {name}, memory allocated (GB): {mem_allocated}, memory reserved (GB): {mem_reserved}, device memory used/total (GB): {mem_used}/{mem_total}"

        self.logging_function(message)
        return output