Share deployment topology mapping between benchmarks and evals

[cquil11]

Benchmark and eval ingestion had separate topology calculations. Eval could multiply TP by EP and count mirrored aggregate metadata as two workers, making a four-GPU SGLang eval appear as 32 GPUs.

Route performance results, individual eval artifacts, and collected eval results through one shared deployment-topology resolver. It owns physical GPU counting, worker multiplicity, aggregate role mirroring, and disaggregation. AgentX SGLang TP/EP share devices; physical node placement alone does not imply separate prefill/decode engines. Explicit GPU counts remain authoritative, and historical fixed-sequence count conventions are preserved.

This addresses the eval mappings found in SemiAnalysisAI/InferenceX#3701, SemiAnalysisAI/InferenceX#3782, and SemiAnalysisAI/InferenceX#3744. Existing database rows are not rewritten; already-ingested incorrect configurations require a separate audited reconciliation.

---

Note

Medium Risk
Changes ETL config identity (GPU counts, disagg flags, aggregate mirroring) for benchmarks and evals; incorrect mapping could mis-key configs until reconciliation, though behavior is heavily test-covered.

Overview
Introduces resolveConfigTopology() in packages/db/src/etl/topology.ts and routes benchmark, individual eval, and agg eval mappers through it, replacing duplicated resolveParallelism / buildEvalConfig logic.

Config identity fixes: legacy role GPU fallbacks now multiply per-worker TP×EP (or AgentX TP×PP×PCP) by worker replicas on both paths; SGLang/Dynamo-SGLang AgentX treats EP/DCP as sharing TP devices; single-node SGLang rows with mirrored 1P/1D role fields collapse to one aggregate config (zero role workers, same GPU count on both sides). Disaggregation is driven by framework aliases and non-zero decode worker pools, not is_multinode or prefill workers alone.

Docs in docs/data-pipeline.md describe the shared resolver and note that fixes apply on future ingest only—existing rows need audited reconciliation, not blind re-ingest.

^{Reviewed by Cursor Bugbot for commit eb61c96. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Oct 9, 2026 5:13pm UTC

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit eb61c96. Configure here.}