Skip to content
ignisignis

Context that follows the card

How long a context a card can hold depends on how much VRAM is left for the KV pool, and the 5090 profile runs the 27B at 917,504 tokens (896K) with vision and DFlash2 on. Both models were trained over 262,144 positions, so a context past that needs a rescaled rotary table. ignis computes it for you: auto leaves positions alone up to 262,144 and applies YaRN at max_context / 262,144 beyond, so changing the context changes the table with it and a mismatch cannot happen silently.

How it works

Section titled “How it works”

Before ADR 0057 each profile spelled a matching yarn:F per model, and a factor that did not match the context was silent: a lower one leaves positions extrapolated, a higher one degrades short contexts for nothing. Now auto resolves per model at load from its effective model.max_context. Within 262,144 it is no scaling at all, the original table and not “YaRN at 1”. Beyond it, it is YaRN at the exact ratio with the default temperature and ramp: 2 at 524,288, 3 at 786,432, 3.5 at the 5090’s 917,504, 4 at 1,048,576. Past 64 times the envelope (16,777,216 tokens) the start is refused, naming the context.

Flash-Next applies YaRN’s attention factor on queries and keys alike, from one table that serves its attention, its sparse-attention indexer and its MTP head (ADR 0059). none and yarn:F[,t=..][,bf=..][,bs=..] override auto from any source, a model’s own section included. Because the factor follows the context, a check that compares two runs of the 27B must pin max_context: when free VRAM changes the context a load can hold, the factor changes with it.

Which contexts fit on which card, and what they cost in KV, is in Profiles and Context and speculation.

The numbers

Section titled “The numbers”
Figure Measured on Source
Trained envelope 262,144 positions; factor = max_context / 262,144; widest factor 64 the rule, not a measurement ADR 0057
The 5090 profile’s 27B context 917,504 tokens with vision and DFlash2 on and a KV pool of 944,832 tokens; 950,272 is refused by 59 MB RTX 5090, the plan raised until it refuses Prefill attention without prompt planes
A needle at 320,076 tokens is recalled both with yarn:4 and with none 27B, one needle YaRN needle at 320K does not discriminate
Flash-Next at about 512K on the 5090: auto 5/5 needles and 1/2 chains, none 4/5 and 1/2. At 1M on the RTX PRO 6000 with MTP: auto 4/5 and 0/2, none 5/5 and 1/2 one seed, one filler, one question per cell Flash-Next YaRN in the reference engines

The last two rows are the whole evidence on quality, and neither separates the tables: the none control did not degrade past 262,144 on these probes, so they cannot say what auto is worth. The ADR itself calls the 27B’s factor on the 5090 profile unmeasured against the yarn:4 it replaces, and Flash-Next’s factors unmeasured at long context.

See it, control it

Section titled “See it, control it”
  • The setting: model.rope_scaling (auto, none or yarn:F[,...]; auto by default, patchable with a reload), and model.max_context.
  • At load: the ignis.model.loaded line names the table that ran (rope_scaling) beside the setting (rope_scaling_setting).
  • Live: GET /v1/config shows the setting as configured, and the Playground’s Server view shows the running configuration.

Limits

Section titled “Limits”
  • A literal needle at 320K does not discriminate. A probe of long-context quality has to ask about relative position, and no such probe has yet separated the tables.
  • Context costs KV. The 5090’s 1,048,576 tokens for the 27B is 1.07 GB short of fitting, which is why its profile stops at 917,504.
  • Flash-Next prefill at 1M tokens takes longer than the default request timeout; raise it for such a prompt (Deadlines).

Read more

Section titled “Read more”