stormlog.infer.vllm_execution_report

Coverage of an imported vLLM execution log, as separate dimensions.

The report answers how much of the measured GPU time the execution log explains, and what it does not, without mixing the answers:

1. linkage: busy time with an iteration link against without, by reason;

2. membership: linked time whose iteration has complete, incomplete or no membership;

3. ownership: linked time in steps that ran only this run’s requests, only other clients’, both, or unresolved IDs;

4. measurement: measured activity (a device UUID and a device clock) against unmeasured, which is counted but never added to a device’s time;

5. capture loss: what the hook dropped, what still waits, and the calls that ran without a range.

Every GPU figure is a union of busy intervals per device and clock scope, never a sum of per-iteration unions. A case’s figure covers every step one of its requests shared, so case figures that share a batch are labelled non-additive. Nothing here estimates per-request GPU cost.

Functions

execution_lines(block)	Text lines for the coverage block; nothing when it was not imported.
execution_report(records)	The coverage block for an artifact's raw records.

stormlog.infer.vllm_execution_report.execution_lines(block)

Text lines for the coverage block; nothing when it was not imported.

Parameters:

block (Any)

Return type:

list[str]

stormlog.infer.vllm_execution_report.execution_report(records)

The coverage block for an artifact’s raw records.

Parameters:

records (list[dict[str, Any]])

Return type:

dict[str, Any]