stormlog.infer.vllm_analysis

Explain a run’s queueing, cache and token signals from vLLM’s own telemetry.

Everything here is aggregate evidence: a scrape describes every request the engine served, from every client, and nothing in it says which request used which GPU time. Deltas are taken between the scrape just before a phase’s first send and the scrape after its drain, and only when both succeeded on one engine epoch. A counter that went backwards, an engine that restarted, or a series one scrape lacks leaves that field unresolved with the reason, never a zero.

Functions

joined_span_attributes(records[, external_spans])	The attributes of each measured request's span, conflicts quarantined.
load_external_spans(records, paths)	Spans someone else collected, stamped with this artifact's identity.
vllm_case_lines(block)	Per-case lines for the text report.
vllm_lines(block)	Report-level lines for the text report.
vllm_report(records[, external_spans])	The telemetry.vllm block of an analysis report.

Classes

JoinedSpans(by_request, quarantined)

Each measured request's single trusted span, by x_request_id.

stormlog.infer.vllm_analysis.load_external_spans(records, paths)

Spans someone else collected, stamped with this artifact’s identity.

Parameters:

• records (list[dict[str, Any]])

• paths (Iterable[str | Path])

Return type:

list[VllmSpanRecord]

stormlog.infer.vllm_analysis.vllm_report(records, external_spans=())

The telemetry.vllm block of an analysis report.

Parameters:

• records (list[dict[str, Any]])

• external_spans (Iterable[VllmSpanRecord])

Return type:

dict[str, Any]

class stormlog.infer.vllm_analysis.JoinedSpans(by_request, quarantined)

Bases: object

Each measured request’s single trusted span, by x_request_id.

quarantined maps the requests left out to why: conflicting_spans when a span arrived again with different content, multiple_spans when a request has more than one span.

Parameters:

• by_request (Mapping[str, Mapping[str, Any]])

• quarantined (Mapping[str, str])

by_request: Mapping[str, Mapping[str, Any]]

quarantined: Mapping[str, str]

stormlog.infer.vllm_analysis.joined_span_attributes(records, external_spans=())

The attributes of each measured request’s span, conflicts quarantined.

Parameters:

• records (list[dict[str, Any]])

• external_spans (Iterable[VllmSpanRecord])

Return type:

JoinedSpans

stormlog.infer.vllm_analysis.vllm_lines(block)

Report-level lines for the text report.

Parameters:

block (Any)

Return type:

list[str]

stormlog.infer.vllm_analysis.vllm_case_lines(block)

Per-case lines for the text report.

Parameters:

block (Any)

Return type:

list[str]