nemo_gym.telemetry.memory

Linux process-tree memory sampling for Gym-managed servers.

Module Contents

Classes

Name	Description
`HostMemory`	Current host memory capacity and pressure.
`MemoryProfiler`	Periodically report host memory and Gym-managed server process trees.
`ProcessTreeMemory`	One process tree’s current memory.
`ServerMemoryTarget`	One configured Gym server and the root process Gym spawned for it.

Functions

Name	Description
`_descendant_pids`	Return a snapshot of a Linux process tree, including its root.
`_get_instruments`	-
`_read_process_memory`	Return `(rss, pss)` bytes for one process, or `None` after an exit race.
`collect_host_memory`	Measure current host memory using psutil’s cross-platform accounting.
`collect_process_tree_memory`	Measure current RSS and PSS for a root process and its descendants.
`process_tree_memory_supported`	Return whether this host exposes Linux procfs memory accounting.
`record_host_memory`	Record current host memory and its sampled peak through Lens.
`record_server_memory`	Record current and sampled-peak process-tree memory through Lens.

Data

_INSTRUMENTS

logger

API

class nemo_gym.telemetry.memory.HostMemory(
    used_bytes: int,
    available_bytes: int,
    total_bytes: int
)

Dataclass

Current host memory capacity and pressure.

available_bytes

int

total_bytes

int

used_bytes

int

class nemo_gym.telemetry.memory.MemoryProfiler(
    targets: list[nemo_gym.telemetry.memory.ServerMemoryTarget],
    interval_seconds: float,
    collector: collections.abc.Callable[[int], nemo_gym.telemetry.memory.ProcessTreeMemory] = collect_process_tree_memory,
    host_collector: collections.abc.Callable[[], nemo_gym.telemetry.memory.HostMemory] = collect_host_memory
)

Periodically report host memory and Gym-managed server process trees.

_host_peaks

dict[str, int] = {}

_peaks

dict[tuple[str, str], int] = {}

_stop

= threading.Event()

_thread

Thread | None = None

nemo_gym.telemetry.memory.MemoryProfiler._run() -> None

nemo_gym.telemetry.memory.MemoryProfiler.sample_once() -> None

Record one host sample and one sample for every configured server.

nemo_gym.telemetry.memory.MemoryProfiler.start() -> None

Start background sampling after recording an initial sample.

nemo_gym.telemetry.memory.MemoryProfiler.stop() -> None

Stop sampling and record final live host and process-tree values.

class nemo_gym.telemetry.memory.ProcessTreeMemory(
    rss_bytes: int,
    pss_bytes: int | None,
    process_count: int
)

Dataclass

One process tree’s current memory.

process_count

int

pss_bytes

int | None

rss_bytes

int

class nemo_gym.telemetry.memory.ServerMemoryTarget(
    name: str,
    server_type: str,
    pid: int
)

Dataclass

One configured Gym server and the root process Gym spawned for it.

name

str

pid

int

server_type

str

nemo_gym.telemetry.memory._descendant_pids(
    root_pid: int
) -> set[int]

Return a snapshot of a Linux process tree, including its root.

nemo_gym.telemetry.memory._get_instruments(
    meter
) -> dict

nemo_gym.telemetry.memory._read_process_memory(
    pid: int
) -> tuple[int, int | None] | None

Return (rss, pss) bytes for one process, or None after an exit race.

nemo_gym.telemetry.memory.collect_host_memory() -> nemo_gym.telemetry.memory.HostMemory

Measure current host memory using psutil’s cross-platform accounting.

nemo_gym.telemetry.memory.collect_process_tree_memory(
    root_pid: int
) -> nemo_gym.telemetry.memory.ProcessTreeMemory

Measure current RSS and PSS for a root process and its descendants.

nemo_gym.telemetry.memory.process_tree_memory_supported() -> bool

Return whether this host exposes Linux procfs memory accounting.

nemo_gym.telemetry.memory.record_host_memory(
    memory: nemo_gym.telemetry.memory.HostMemory,
    peaks: dict[str, int]
) -> None

Record current host memory and its sampled peak through Lens.

nemo_gym.telemetry.memory.record_server_memory(
    target: nemo_gym.telemetry.memory.ServerMemoryTarget,
    memory: nemo_gym.telemetry.memory.ProcessTreeMemory,
    peaks: dict[tuple[str, str], int]
) -> None

Record current and sampled-peak process-tree memory through Lens.

nemo_gym.telemetry.memory._INSTRUMENTS: WeakKeyDictionary = weakref.WeakKeyDictionary()

nemo_gym.telemetry.memory.logger = logging.getLogger(__name__)