[markyu0401]

The prefill bench's custom-size field rejected what its own hint suggested, and its ceiling ignored the model.

• The input took digits only, and parseContextSize ran parseInt("200k") → 200, so a typed 200k/128k never became a size.
• PREFILL_MAX_CONTEXT_SIZE was a fixed 300k, so a model with a larger window could not be asked for more than 300k.

Changes

• parseContextSize(raw, maxTokens) accepts a k/m suffix in the binary sense the decode dialog already uses (128k = 131072, 1m = 1048576) and validates against the caller's ceiling.
• prefillCeilingFor(contextLength) takes the ceiling from the model's own context. There is deliberately no fixed upper bound — models already ship windows past 1M, and any constant here is just the next thing to raise. 300k stays only as the fallback when the context is unknown.
• The dialog uses that ceiling for the custom field and accepts the suffix in the input; the hint now shows 200k as an example.
• The server keeps its structural guard: with a known context it accepts any structurally valid size and lets the existing fit check emit the skipped row; with an unknown context the 300k bound is the only limit available.
• formatContextSize reads 204800 back as 200k and 1048576 as 1m.

Not in scope

The decode bench's start-context field: it has no model context at start and already parses 128k. Also worth knowing that timeoutMsForSize still caps at 45 min, which is the practical ceiling for very large sizes on a slow engine.

Tests

Parser/ceiling/normalise cases in server/collectors/__tests__/PrefillBench.test.js, plus three dialog cases (600k accepted on a 1M context, 1.5m accepted on a 2M context, 600k refused with the real ceiling on 128k). npm run typecheck, build, i18n:check, full vitest (808) and server/__tests__ + PrefillBench (69) pass.

Overlap

#323 (i18n of the benchmark workflows) touches PrefillBenchDialog.tsx and is currently unmergeable; a rebase over it is mechanical (the hint/aria strings move into the catalog).

Reported by @de-flandres in #343.

[MiaAI-Lab]

Ported to main as 3bcea09. No constant ceiling of its own is the right call — a model past 1M would just make it the next number to raise; the k/m parse, the explaining skip row and the dialog ceiling came over intact. Full gate green after resolving against the current dialog. CHANGELOG credits you. Closing as completed (closes #343).

(Review notes prepared with Claude Code on behalf of the maintainer.)

[MiaAI-Lab]

now on main as 3bcea09.