stormlog.infer.vllm_execution_devices

Map a profiler trace to the GPU it ran on, from the vLLM hook’s worker hellos.

Each worker epoch’s hello names the worker’s host, pid, CUDA ordinal and GPU UUID (docs/vllm_execution.md). A trace names its host and launching pids and spans a wall-clock window, so a worker epoch that was alive on that host, with that pid, across that window is the process that wrote it: its UUID is the trace’s device. A pid that no epoch covers, or that several do, is left for --device-uuid.

Functions

trace_wall_window(trace)

The wall-clock span of a trace's launches and GPU events, if any.

Classes

TraceBinding(status[, uuids, workers, ...])	The GPU UUIDs a trace's processes map to, and what could not be mapped.
WorkerEpoch(epoch, host, boot_id, pid, ...)	One worker process lifetime, as its hello and liveness records say.
WorkerIndex(workers[, directory])	Every worker epoch of a hook directory, ready to bind traces.

class stormlog.infer.vllm_execution_devices.TraceBinding(status, uuids=<factory>, workers=<factory>, ambiguous=<factory>, unmatched=<factory>)

Bases: object

The GPU UUIDs a trace’s processes map to, and what could not be mapped.

Parameters:

• status (str)

• uuids (dict[int, dict[int, str]])

• workers (dict[int, str])

• ambiguous (dict[int, list[str]])

• unmatched (list[int])

status: str

uuids: dict[int, dict[int, str]]

workers: dict[int, str]

ambiguous: dict[int, list[str]]

unmatched: list[int]

summary()

Return type:

dict[str, Any]

class stormlog.infer.vllm_execution_devices.WorkerEpoch(epoch, host, boot_id, pid, engine_producer, cuda_ordinal, device_uuid, local_rank, rank, trace_rank_suffix, enabled, refused, started_wall_ns, last_seen_wall_ns, ended)

Bases: object

One worker process lifetime, as its hello and liveness records say.

Parameters:

• epoch (str)

• host (str)

• boot_id (str | None)

• pid (int)

• engine_producer (str | None)

• cuda_ordinal (int | None)

• device_uuid (str | None)

• local_rank (int | None)

• rank (dict[str, int])

• trace_rank_suffix (str | None)

• enabled (bool)

• refused (str | None)

• started_wall_ns (int | None)

• last_seen_wall_ns (int | None)

• ended (bool)

epoch: str

host: str

boot_id: str | None

pid: int

engine_producer: str | None

cuda_ordinal: int | None

device_uuid: str | None

local_rank: int | None

rank: dict[str, int]

trace_rank_suffix: str | None

enabled: bool

refused: str | None

started_wall_ns: int | None

last_seen_wall_ns: int | None

ended: bool

covers(start_wall_ns, end_wall_ns)

Whether this process was alive across the whole window.

Parameters:

• start_wall_ns (int)

• end_wall_ns (int)

Return type:

bool

class stormlog.infer.vllm_execution_devices.WorkerIndex(workers, directory=None)

Bases: object

Every worker epoch of a hook directory, ready to bind traces.

Parameters:

• workers (Sequence[WorkerEpoch])

• directory (Path | None)

classmethod from_log(read)

Parameters:

read (LogRead)

Return type:

WorkerIndex

classmethod from_directory(directory, *, importer=None)

Parameters:

• directory (str | Path)

• importer (Importer | None)

Return type:

WorkerIndex

bind_trace(trace)

Bind a loaded trace by its host, launching pids and time window.

Parameters:

trace (KinetoTrace)

Return type:

TraceBinding

bind(*, host, pids, start_wall_ns, end_wall_ns)

The one worker epoch per pid alive on host across the window.

Parameters:

• host (str | None)

• pids (Iterable[int])

• start_wall_ns (int)

• end_wall_ns (int)

Return type:

TraceBinding

summary()

Return type:

dict[str, Any]

stormlog.infer.vllm_execution_devices.trace_wall_window(trace)

The wall-clock span of a trace’s launches and GPU events, if any.

Parameters:

trace (KinetoTrace)

Return type:

tuple[int, int] | None