feat: add multimodal benchmark history and observability

[ActivePeter]

Summary

• add reusable stream-world profiling for TeleFuser and SGLang targets
• add active process-tree, host, container, GPU, Ethernet, and RDMA telemetry
• add required GreptimeDB history ingestion and a Vue 3 plus TypeScript comparison dashboard
• group the UI by canonical design dimensions with multi-run left and right comparison
• document the multimodal contracts, metrics, history service, and target integration boundary

Validation

• 13,640 unit tests passed, 98 skipped
• full pre-commit suite passed
• Vue TypeScript production build passed with no generated-artifact drift
• all stream configs and contracts loaded through their real typed parsers
• live 2xH100 TeleFuser profile persisted 4,273 merged statistic and resource points to GreptimeDB with zero dropped samples

Integration

The companion TeleFuser PR fixes the AIPerf checkout at <TeleFuser>/benchmarks/aiperf and contains only thin target launchers, contracts, and raw runtime facts.

Companion TeleFuser PR: Tele-AI/TeleFuser#13

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive benchmark history service backed by GreptimeDB, featuring a new CLI for data ingestion and a Vue-based dashboard for visualization. It also implements multimodal benchmark extensions and active resource telemetry collection for monitoring process trees, cgroups, and GPU usage. The reviewer provided actionable feedback, including suggestions to prevent silent data overwrites in the video generation endpoint, improve type safety in the payload merging logic, and add logging for potential PID truncation in the cgroup collector.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

If both an image and a video are provided in the turn, the video data will overwrite the image data in the payload because both branches can set reference_url and input_reference. If the API doesn't support both simultaneously, it might be better to raise an error or log a warning to avoid this silent overwrite and make the behavior more explicit to the user.

[gemini-code-assist]

For better type safety and code clarity, the items parameter should be typed more specifically as Iterable[tuple[str, Any]] instead of Any. This more accurately reflects its usage with both a list of tuples and dictionary items.

You'll need to add from collections.abc import Iterable at the top of the file.

Suggested change

	def _merge_with_reserved_filter(
	self,
	payload: dict[str, Any],
	items: Any,
	label: str,
	) -> None:
	def _merge_with_reserved_filter(
	self,
	payload: dict[str, Any],
	items: Iterable[tuple[str, Any]],
	label: str,
	) -> None:

[gemini-code-assist]

If the number of process IDs found in the cgroup subtree exceeds _MAX_CGROUP_PROCESSES, the list of PIDs is truncated. This could lead to incomplete resource attribution for the container without any warning. Consider logging a warning when this truncation occurs to alert the user that the collected container resource metrics might be incomplete.

[ActivePeter]

Follow-up hardening landed on teleai in f5e936e9, with the fresh-checkout artifact-root documentation correction in e977ffbb: runtime setup guidance, image/video reference validation, and bounded cgroup PID traversal that disables partial container-GPU attribution when scan limits are exceeded. Targeted tests, pre-commit, ergonomics checks, clean-room History startup, real-artifact ingest, and wheel asset verification passed.