onnx_diagnostic.helpers.memory_peak

class onnx_diagnostic.helpers.memory_peak.MemorySpy(pid: int, delay: float = 0.01, cuda: bool = False)

Information about the spy. It class method start. Method stop can be called to end the measure.

Parameters:

• pid – process id of the process to spy on

• delay – spy on every delay seconds

• cuda – enable cuda monitoring

start() → MemorySpy

Starts another process and tells it to spy.

stop()

Stops spying on.

onnx_diagnostic.helpers.memory_peak.get_memory_rss(pid: int) → int

Returns the physical memory used by a process.

Parameters:

pid – process id, current one is os.getpid()

Returns:

physical memory

It relies on the module psutil.

onnx_diagnostic.helpers.memory_peak.start_spying_on(pid: int | None = None, delay: float = 0.01, cuda: bool = False) → MemorySpy

Starts the memory spy. The function starts another process spying on the one sent as an argument.

Parameters:

• pid – process id to spy or the the current one.

• delay – delay between two measures.

• cuda – True or False to get memory for cuda devices

Example:

from onnx_diagnostic.helpers.memory_peak import start_spying_on, flatten

p = start_spying_on()
# ...
# code to measure
# ...
stat = p.stop()
print(stat)
print(flatten(stat))