feat(unofficial): support 950dt + huawei-mindie for unofficial runs

[Oseltamivir]

Summary

Unblocks loading the Mock-ascend run 25014782858 via the unofficial-run viewer (?unofficialrun=25014782858).

The mock artifacts use:

• hw: 950dt — new GPU, Huawei Ascend 950DT (new vendor)
• framework: huawei-mindie — new framework
• model: deepseek-ai/DeepSeek-V4-Pro — already supported (dsv4)
• precision: fp4, sequences 8K/1K, 1K/1K — already supported

Without these mappings, hwToGpuKey('950dt') returns null and rows are silently skipped by the ETL → the unofficial-run viewer shows nothing.

Changes

• packages/constants/src/gpu-keys.ts
  • HW_REGISTRY['950dt'] — vendor Huawei, arch Ascend, label Ascend 950DT, sort 9. TDP/power/cost set to placeholder 9.99 — Huawei figures are not yet finalized; bump to real values when available.
  • New huawei color zone in VENDOR_OKLCH_ZONES (amber 60–90°) and VENDOR_HSL_ZONES (amber 30–60°). To make room in the HSL wheel, AMD's lower band shrinks from 0–60 to 0–30 and NVIDIA's band shrinks from 60–195 to 90–195. AMD/NVIDIA still get >90° each.
• packages/constants/src/framework-aliases.ts
  • FW_REGISTRY['huawei-mindie'] → label Huawei MindIE.
• packages/app/src/lib/dynamic-colors.ts
  • Vendor type now includes 'huawei'; getVendor() recognizes the Huawei vendor string.
• packages/app/src/lib/chart-utils.ts
  • BANNED_HUE_TEST / PREFERRED_ZONE huawei entries — avoids AMD-red (320–40°) and NVIDIA-green (120–195°) in CIELab; prefers amber/yellow (50–95°).

Notes for review

• This is added solely so the unofficial-run viewer can render the mock data — it is not intended to ingest mock benchmarks into the production DB. The mock artifacts use tput_per_gpu: 510.2041, median_ttft: 0.0, etc. — fake values.
• Cost/power/TDP are placeholders; once real Huawei pricing/specs are known, replace 9.99 and tune the chart sort order.

Test plan

• pnpm typecheck
• pnpm lint
• pnpm test:unit (1896 tests pass across all packages)
• On Vercel preview, load ?unofficialrun=25014782858 — confirm 950dt overlay points render under DeepSeek-V4-Pro at 1K/1K, 8K/1K, etc., with conc 16/64/128.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	May 7, 2026 9:10pm

[Oseltamivir]

@claude

Make it able to render 8192 input tokens, 256 output tokens, like from runs/25457884421

[github-actions]

Claude finished @Oseltamivir's task in 7m 56s —— View job • claude/pr-257-20260506-2045 • Create PR ➔

---

Done — registered the 8K/256 sequence (8192/256)

Run 25457884421 is the InfiniteBench DSv4 sweep that emits isl=8192, osl=256. The unofficial-run path was silently dropping every row because islOslToSequence(8192, 256) returned null (and getModelAndSequenceFromArtifact would mis-map it to 8k8k due to a binary 1k-vs-8k branch).

What changed

• packages/constants/src/models.ts — registered '8k/256' ↔ {isl: 8192, osl: 256} in both sequenceToIslOsl and islOslToSequence.
• packages/app/src/lib/data-mappings.ts — added Sequence.EightK_256 = '8k/256' and a SEQUENCE_CONFIG entry (label '8K / 256', compact '8k256', category 'default'). Replaced the binary 1k-vs-8k ternary in getModelAndSequenceFromArtifact with a lookup against islOslToSequence, so future ISL/OSL pairs registered in constants are picked up automatically rather than silently mis-mapped.
• Tests — added round-trip + label coverage in packages/constants/src/models.test.ts and a new getModelAndSequenceFromArtifact case for dsv4 + 8192/256 in packages/app/src/lib/data-mappings.test.ts.

The flow that now works for ?unofficialrun=25457884421:

1. Artifact JSON ships isl: 8192, osl: 256 → mapBenchmarkRow passes through unchanged.
2. buildChartData in unofficial-run-provider.tsx:113 calls islOslToSequence(8192, 256) → returns '8k/256' (was null).
3. GlobalFilterContext availableSequences picks up '8k/256' from unofficialAvailable, the dropdown surfaces it, and useChartData filters rows by {isl: 8192, osl: 256}.

Checks

• pnpm typecheck ✅
• pnpm lint ✅ (0 warnings, 0 errors)
• pnpm test:unit — app: 1732/1732 ✅, constants: 22/22 ✅. The 3 failures in packages/db/src/connection.test.ts (postgres TLS for localhost) are pre-existing on master, unrelated to this change — verified by git stash && pnpm --filter db test:unit.

Not done

• No browser verification — I don't have GitHub artifact access from this sandbox to actually load ?unofficialrun=25457884421 against the dev server, so I can't visually confirm the overlay points render. Once the Vercel preview is up, please load that URL and confirm overlay points appear under DeepSeek-V4-Pro at the new 8K / 256 sequence option. AGENTS.md flags this as a mandatory step for inference-path features.

Todos

• Find where ISL/OSL sequences are registered
• Inspect run 25457884421 to confirm ISL/OSL
• Add 8192/256 to sequenceToIslOsl / islOslToSequence
• Add Sequence.EightK_256 + SEQUENCE_CONFIG entry
• Generalize getModelAndSequenceFromArtifact
• Update tests
• typecheck / lint / unit
• Commit and push

· branch: claude/pr-257-20260506-2045

[Oseltamivir]

@claude

Make it able to render 8192 input tokens, 256 output tokens, like from runs/25457884421

Use this PR/branch, don't create a new branch

[github-actions]

Claude finished @Oseltamivir's task in 4m 23s —— View job

---

Done — registered 8K/256 sequence (8192/256) on this PR branch

Pushed commit f1549ed to claude/huawei-950dt-mindie-unofficial. Sweep run 25457884421 ships isl=8192, osl=256; islOslToSequence previously returned null and the rows were silently dropped by the unofficial-run viewer.

Changes

• packages/constants/src/models.ts — added '8k/256' ↔ {isl: 8192, osl: 256} in both sequenceToIslOsl and islOslToSequence.
• packages/app/src/lib/data-mappings.ts — added Sequence.EightK_256 = '8k/256' and a SEQUENCE_CONFIG entry (label '8K / 256', compact '8k256', category 'default'). Replaced the binary 1k-vs-8k ternary in getModelAndSequenceFromArtifact with islOslToSequence, so future ISL/OSL pairs registered in constants are picked up automatically rather than silently mis-mapped (the old ternary would have mapped osl=256 → 8k, producing a wrong 8k1k sequence key).
• Tests — added round-trip + label coverage in packages/constants/src/models.test.ts and packages/app/src/lib/models-mapping.test.ts, plus a new getModelAndSequenceFromArtifact case for dsv4 + 8192/256 in packages/app/src/lib/data-mappings.test.ts.

The flow that now works for ?unofficialrun=25457884421:

1. Artifact JSON ships isl: 8192, osl: 256 → mapBenchmarkRow passes through.
2. unofficial-run-provider.tsx:113 calls islOslToSequence(8192, 256) → returns '8k/256' (was null).
3. GlobalFilterContext availableSequences picks up '8k/256' from unofficialAvailable, the sequence dropdown surfaces it, and useChartData filters rows by {isl: 8192, osl: 256}.

Checks

• pnpm typecheck — clean
• pnpm --filter @semianalysisai/inferencex-constants test:unit — 22/22 passed
• pnpm --filter @semianalysisai/inferencex-app test:unit — 1689/1689 passed
• pnpm lint — 0 new errors. The 41 errors reported are pre-existing on master (verified by git stash && pnpm lint); they're all no-underscore-dangle violations in files I didn't touch.

Not done — needs your verification

• No browser verification. The dev server in this sandbox can't fetch GitHub artifacts, so I can't load ?unofficialrun=25457884421 and visually confirm the overlay points render. AGENTS.md flags Playwright verification as mandatory for inference-path features — please load the Vercel preview at ?unofficialrun=25457884421 once it deploys and confirm DSv4 overlay points appear under the new 8K / 256 sequence option.
  • Branch: claude/huawei-950dt-mindie-unofficial

[Oseltamivir]

@claude

For unofficial runs, add a toggle where the dashboard treats them as ingested runs.

• They actual names in the legend, if a job has multiple they're treated as separate lines
• When this is toggled, the plots should be susceptible to the other toggles(such as "Optimal only", currently they don't.
• Ensure everything works for multiple unofficial runs
• Animation isn't necessary

[github-actions]

Claude finished @Oseltamivir's task in 24m 55s —— View job

---

Done — "Show as ingested" toggle for unofficial runs

Pushed 5e327a7 to claude/huawei-950dt-mindie-unofficial. When the toggle is on, unofficial-run rows are promoted into the same series pipeline as DB-ingested data, so they participate in Optimal-only, hardware toggles, precision filters, and per-roofline grouping.

How it works

• Each (run, original hwKey) pair is rewritten to a synth key ${origHwKey}__uorun${runId} (e.g. b200_vllm__uorun25457884421). The GPU base prefix is preserved so getModelSortIndex / isKnownGpu keep working — only the suffix is new.
• A synthesized HardwareEntry is generated for each synth key with label = "${origLabel} • ${branch}" so the legend shows e.g. B200 • infinitebench-dsv4-b200-20260506-125800 (vLLM).
• A per-key color override (from the existing overlayRunColor palette) is threaded through InferenceContext → ScatterGraph's resolveColor so each run still gets its distinct color even after the GPU vendor system would normally re-paint by base GPU.
• Multiple runs surface as separate legend entries; one job that benchmarks several GPUs surfaces those GPUs as separate legend entries too — (run, GPU) is the legend granularity.
• The X-shape overlay layer is suppressed when merged, and stale .unofficial-overlay-pt / .overlay-roofline-path DOM is cleaned up on the toggle-flip (the D3 layer system has no built-in teardown for dropped layers).

What changed

• packages/app/src/lib/unofficial-merge.ts (new) — mergeUnofficialIntoOfficial, makeSynthHwKey, parseSynthHwKey, isSynthHwKey, buildStubGraphsForMerge. Reuses processOverlayChartData so the merged points sit in the same x/y coordinate space the official pipeline uses.
• packages/app/src/lib/unofficial-merge.test.ts (new) — 10 tests covering synth-key round-trips, no-op behavior when no overlay matches the (model, sequence), single-run merge, multi-run separation (same GPU under two runs gets two synth keys + two colors), preserving official rows, and stub-graph synthesis when official graphs is empty.
• packages/app/src/components/unofficial-run-provider.tsx — adds mergeAsIngested + setMergeAsIngested to context. URL-synced via i_uoff_ingested=1 (with a hydration-safe useEffect re-sync).
• packages/app/src/components/inference/InferenceContext.tsx — calls the merger after useChartData, exposes merged graphs/hardwareConfig and hwColorOverrides through context. Result memoized, so the toggle adds zero overhead when off.
• packages/app/src/components/inference/ui/ScatterGraph.tsx — wraps resolveColor to consult hwColorOverrides first; adds the "Show as ingested" switch (only rendered when isUnofficialRun); cleanup effect removes stale overlay DOM when overlayData drops to null.
• packages/app/src/components/inference/ui/ChartDisplay.tsx — sets overlayDataByChartType to null when mergeAsIngested is on, so the X-shape overlay layer doesn't render alongside the merged points.
• packages/app/src/lib/url-state.ts — registers i_uoff_ingested as a known URL state key.
• packages/app/src/components/inference/types.ts, packages/app/cypress/support/mock-data.ts — type-level wiring for hwColorOverrides and the new toggle fields.

Analytics: latency_unofficial_as_ingested_toggled fires on every flip via the existing track() convention.

Browser verification (Playwright)

Walked through the user-facing flow against localhost:3000:

• ?unofficialrun=25457884421 (InfiniteBench DSv4 8K/256, B200 vLLM) — toggling on flips the 2 X-shape overlay points to ingested-style dot-groups with synth hwKey = "b200_vllm__uorun25457884421". Optimal-only is on; both points are Pareto-optimal so both render. Toggling off returns to overlay X-shapes (URL param drops too).
• ?unofficialruns=25014782858 (Mock-ascend, DSv4 8K/1K) — synth hwKey = "950dt_huawei-mindie_mtp__uorun25014782858" co-renders with 374 official DSv4 points; no overlay X-shapes remain.
• ?unofficialruns=25457884421,25014782858 — both run chips show in the banner; each contributes its own synth-keyed legend entry on the slices where it has data.
• Reload with i_uoff_ingested=1 in the URL → toggle starts checked (state is hydration-safe, with a useEffect re-sync to handle the SSR default).

Checks

• pnpm typecheck — clean
• pnpm --filter @semianalysisai/inferencex-app test:unit — 1699/1699 passed (10 new tests in unofficial-merge.test.ts)
• pnpm --filter @semianalysisai/inferencex-constants test:unit — 22/22 passed
• pnpm lint — 0 new errors (the 41 reported are pre-existing on master, all no-underscore-dangle in files I didn't touch — verified by git stash && pnpm lint)
• Browser-verified with Playwright per the cases above
  · branch: claude/huawei-950dt-mindie-unofficial

[Oseltamivir]

@claude

for the previous edit, the results look good, but continue to build on it:

• I don't want the name "infinitebench-dsv4-b200-20260506-125800" to be in the legend
• I want the colours to follow ingested runs, the B300 plots are showing up as the same red. They should be different shades of green for Nvidia

[github-actions]

Claude finished @Oseltamivir's task in 12m 4s —— View job

---

Done — bare GPU labels + vendor-zone colors for ingested unofficial rows

Pushed 323d1b4 to claude/huawei-950dt-mindie-unofficial. Both fixes live in packages/app/src/lib/unofficial-merge.ts.

What changed

• Legend label drops the branch. makeSynthHardwareEntry previously produced "${baseLabel} • ${branch}"; now it returns just baseLabel (e.g. Ascend 950DT (Huawei MindIE, MTP) instead of ... • mock-ascend-950dt-deepseekv4-20260427-185322). The branch + run URL still surface in gpu, so the row tooltip retains run provenance.
• No more per-run color override. The merge no longer populates colorOverrides[synthHwKey] = overlayRunColor(runIdx). Synth keys preserve the original GPU base prefix (b200_vllm__uorun123), so getVendor() recognizes them and generateVendorColors distributes a hue from the vendor's OKLCh zone — exactly the path official rows already use. That's why two NVIDIA GPUs from one unofficial run now get different shades of green rather than two copies of the same --overlay-run-0 red.

The wiring (hwColorOverrides → ScatterGraph.resolveColor) stays in place; the merge just returns an empty {} map. Future callers can still pin a specific color per synth key without further refactors.

Browser verification (Playwright)

Loaded ?unofficialruns=25457884421,25014782858&i_uoff_ingested=1 against localhost:3000, switched to DSv4 / 8K / 1K, hit "Reset filter" to expand the legend, and dumped the swatch styles:

• Ascend 950DT (Huawei MindIE, MTP) → oklch(0.64 0.18 75) — Huawei amber zone (60–90)
• GB300 NVL72 (Dynamo SGLang) → oklch(0.78 0.15 121.7) — NVIDIA green
• B300 (vLLM) → oklch(0.62 0.15 148.3) — NVIDIA green, not red
• B200 (vLLM) → oklch(0.52 0.15 165) — NVIDIA green, distinct shade
• MI355X (SGLang) → oklch(0.5 0.22 34.5) — AMD red
• MI355X (ATOM¹) → oklch(0.78 0.22 19.5) — AMD red

Each NVIDIA GPU sits in the 121.7–168.3 hue band — a different shade per GPU, exactly the spec. No legend label contains "infinitebench-..." or "mock-ascend-..." anywhere. Switched to 8K / 256 to confirm the InfiniteBench overlay (B200 (vLLM)) also receives a green hue (oklch(0.64 0.15 145)) instead of the previous overlay-run red.

Tests

Added 3 integration cases to packages/app/src/lib/unofficial-merge.test.ts:

• two NVIDIA synth keys (B200 + B300) from a single run get distinct colors;
• vendor-zone hues stay inside their bands (NVIDIA 120–170, AMD 12–42 per VENDOR_OKLCH_ZONES);
• two synth keys for the same GPU but different runs still get distinct hues.

Updated the existing label/color assertions to expect bare GPU labels and an empty colorOverrides map. The gpu field is still asserted to contain UNOFFICIAL: <branch> so we don't regress the tooltip provenance.

Checks

• pnpm typecheck — clean
• pnpm exec vitest run src/lib/unofficial-merge.test.ts — 13/13 passed (10 original + 3 new)
• pnpm exec vitest run (whole app) — 1702/1702 passed
• pnpm exec oxlint packages/app/src/lib/unofficial-merge.ts unofficial-merge.test.ts — 0 errors. The 41 pnpm lint repo-wide errors are pre-existing on master (no-underscore-dangle in files I didn't touch).
• Browser-verified with Playwright per the swatch dump above
  · branch: claude/huawei-950dt-mindie-unofficial

[Oseltamivir]

@claude

Just for this branch, temporarily set default view to "8k / 256" and "Output Token Throughput per GPU". This default should apply after unofficial runs automatically changes model (introduced in #243 )

[github-actions]

Claude finished @Oseltamivir's task in 8m 40s —— View job

---

Done — temp default to 8K / 256 + Output Token Throughput per GPU on unofficial-run load

Pushed 462dc0f to claude/huawei-950dt-mindie-unofficial. When ?unofficialrun= loads and the URL doesn't pin i_seq / i_metric, the dashboard now defaults the sequence to 8K / 256 and the y-axis metric to Output Token Throughput per GPU, alongside the existing model auto-switch from #243.

What changed

• packages/app/src/lib/unofficial-run-auto-switch.ts — added a sibling pure helper computeUnofficialOverrideDecision(unofficialAvailable, urlValue, lastKey) that mirrors computeAutoSwitchDecision's dedupe pattern: empty overlay → reset key, URL-pinned param → no-op (preserve the same key so a later URL clear can re-fire), same model-set key → no-op, otherwise → fire override and advance key. Same sorted-models key shape as the auto-switch so a sequence-only delta doesn't invalidate a manual user pick.
• packages/app/src/components/GlobalFilterContext.tsx — added a useEffect that calls the helper with getUrlParam('i_seq') and a lastUnofficialSeqOverrideRef, then setSelectedSequence(Sequence.EightK_256) when shouldOverride is true. Lives next to the existing lastAutoSwitchKeyRef effect so the two are reviewed together.
• packages/app/src/components/inference/InferenceContext.tsx — analogous effect using getUrlParam('i_metric') + lastUnofficialMetricOverrideRef, calling setSelectedYAxisMetric('y_outputTputPerGpu'). Pulled availableModelsAndSequences out of useUnofficialRun() (already in the same destructure block) so the effect can key off the same overlay set the auto-switch sees.
• packages/app/src/lib/unofficial-run-auto-switch.test.ts — added 6 unit tests for the new helper: empty overlay resets, fresh fire with no URL pin, URL pin respected, no re-fire after applied, re-arm after overlay clears, sequence-only delta ignored, deterministic key across insertion orders.

Both effects are clearly tagged TEMPORARY (this branch only) so the override is easy to revert.

Browser verification (Playwright on localhost:3000)

• ?unofficialrun=25457884421 (no pins) — model auto-switched to DeepSeek V4 Pro, sequence dropdown landed on 8K / 256, y-axis on Output Token Throughput per GPU, NON-OFFICIAL banner present.
• ?unofficialrun=25457884421&i_seq=1k%2F1k&i_metric=y_tpPerGpu — sequence stayed at 1K / 1K, y-axis stayed at Token Throughput per GPU (URL pins respected). Model still auto-switched because g_model wasn't pinned.

Checks

• pnpm typecheck — clean
• pnpm --filter @semianalysisai/inferencex-app exec vitest run src/lib/unofficial-run-auto-switch.test.ts — 15/15 passed
• pnpm --filter @semianalysisai/inferencex-app test:unit — 1773/1773 passed
• pnpm lint — 0 warnings, 0 errors
• Browser-verified per the cases above

Notes

• The 2 console errors observed during the URL-pinned navigation case are pre-existing (a script-tag warning and a SSR/CSR hydration mismatch on SequenceSelector from the existing useState initializer reading the URL on the client — happens for any i_seq= URL pin, regardless of this change). Not introduced by this commit.
• The override is gated by both the dedupe ref and the URL-pin check, so manual sequence/metric changes (which write back to the in-memory URL store and then to the address bar) are safe — the next render reads the pinned URL value via getUrlParam and bails. Switching to a different unofficial run with a different model set re-arms the override (same behavior as the model auto-switch).
  · branch: claude/huawei-950dt-mindie-unofficial