mirror of
https://github.com/morluto/rea.git
synced 2026-11-19 21:23:45 +00:00
131 lines
4.5 KiB
TypeScript
131 lines
4.5 KiB
TypeScript
import { Budget } from "../domain/budget.js";
|
|
import {
|
|
malformedInput,
|
|
ParseFailureError,
|
|
unsupportedInput,
|
|
} from "../domain/parseFailure.js";
|
|
|
|
/** Producing constraint for a model capacity refusal. Recovery uses this, not message text. */
|
|
export const MODEL_CAPACITY_LIMITS = [
|
|
"header-bytes",
|
|
"report-bytes",
|
|
"representation-bytes",
|
|
"exact-offset",
|
|
"metadata-nesting",
|
|
"json-nesting",
|
|
] as const;
|
|
|
|
/** Which fixed model allowance refused the request. */
|
|
export type ModelCapacityLimit = (typeof MODEL_CAPACITY_LIMITS)[number];
|
|
|
|
/** The ParseFailure format of GGUF and Safetensors model headers. */
|
|
export const MODEL_FORMAT = "model";
|
|
|
|
/** A fixed model allowance refused the input; the limit name is its {@link ModelCapacityLimit}. */
|
|
export const modelCapacityFailure = (
|
|
limit: ModelCapacityLimit,
|
|
detail: string,
|
|
): ParseFailureError<"limit"> =>
|
|
new ParseFailureError({
|
|
kind: "limit",
|
|
format: MODEL_FORMAT,
|
|
detail,
|
|
limit: {
|
|
name: limit,
|
|
maximum: Object.values(modelCapacityReportedLimits(limit))[0] ?? 0,
|
|
},
|
|
});
|
|
|
|
/** The model allowance a failure names, or undefined for any other failure. */
|
|
export const modelCapacityLimitOf = (
|
|
cause: ParseFailureError,
|
|
): ModelCapacityLimit | undefined => {
|
|
const { failure } = cause;
|
|
if (failure.kind !== "limit") return undefined;
|
|
return MODEL_CAPACITY_LIMITS.find((limit) => limit === failure.limit.name);
|
|
};
|
|
|
|
export function rejectModel(message: string): never {
|
|
throw malformedInput(MODEL_FORMAT, message);
|
|
}
|
|
|
|
export function rejectPickle(): never {
|
|
throw unsupportedInput(
|
|
MODEL_FORMAT,
|
|
"PyTorch pickle/ZIP model deserialization is unsupported; supply GGUF or safetensors instead",
|
|
);
|
|
}
|
|
|
|
// Shared header ceiling for GGUF and Safetensors. Safetensors counts declared
|
|
// JSON bytes after the eight-byte length prefix; GGUF counts the complete
|
|
// header. Retained rows and the serialized report have separate allowances.
|
|
export const MODEL_HEADER_BYTES = 100_000_000;
|
|
export const MODEL_REPORT_BYTES = 32 * 1024 * 1024;
|
|
export const MODEL_ARRAY_PREVIEW = 8;
|
|
export const MODEL_METADATA_NESTING = 32;
|
|
export const MODEL_JSON_NESTING = 32;
|
|
|
|
const fixedModelLimit = (allowance: string): string =>
|
|
`${allowance}. This release has no runtime override. Additional host memory does not raise it. Repeating the unchanged request fails the same way.`;
|
|
|
|
/** Recovery for the allowance that actually refused the model. Host memory does not raise it. */
|
|
export const modelCapacityRemediation = (limit: ModelCapacityLimit): string => {
|
|
switch (limit) {
|
|
case "header-bytes":
|
|
return fixedModelLimit("The model header allowance is 100,000,000 bytes");
|
|
case "report-bytes":
|
|
return fixedModelLimit(
|
|
`The retained model report allowance is ${String(MODEL_REPORT_BYTES)} bytes`,
|
|
);
|
|
case "representation-bytes":
|
|
return fixedModelLimit(
|
|
`The retained model representation allowance is ${String(MODEL_REPORT_BYTES)} bytes`,
|
|
);
|
|
case "exact-offset":
|
|
return fixedModelLimit(
|
|
`Exact host offset precision stops at ${String(Number.MAX_SAFE_INTEGER)}`,
|
|
);
|
|
case "metadata-nesting":
|
|
return fixedModelLimit(
|
|
`GGUF metadata nesting is limited to ${String(MODEL_METADATA_NESTING)} array frames`,
|
|
);
|
|
case "json-nesting":
|
|
return fixedModelLimit(
|
|
`Safetensors JSON nesting is limited to ${String(MODEL_JSON_NESTING)} levels`,
|
|
);
|
|
}
|
|
};
|
|
|
|
/** Limits that belong to the producing constraint, without unrelated ceilings. */
|
|
export const modelCapacityReportedLimits = (
|
|
limit: ModelCapacityLimit,
|
|
): Readonly<Record<string, number>> => {
|
|
switch (limit) {
|
|
case "header-bytes":
|
|
return { maximum_header_bytes: MODEL_HEADER_BYTES };
|
|
case "report-bytes":
|
|
return { maximum_report_bytes: MODEL_REPORT_BYTES };
|
|
case "representation-bytes":
|
|
return { maximum_representation_bytes: MODEL_REPORT_BYTES };
|
|
case "exact-offset":
|
|
return { maximum_exact_offset: Number.MAX_SAFE_INTEGER };
|
|
case "metadata-nesting":
|
|
return { maximum_array_frames: MODEL_METADATA_NESTING };
|
|
case "json-nesting":
|
|
return { maximum_json_depth: MODEL_JSON_NESTING };
|
|
}
|
|
};
|
|
|
|
/** Charge representation growth before retaining rows or decoding long strings. */
|
|
export class ModelReportBudget {
|
|
readonly #bytes = new Budget(MODEL_REPORT_BYTES);
|
|
charge(bytes: number): void {
|
|
this.#bytes.charge(bytes, () =>
|
|
modelCapacityFailure(
|
|
"representation-bytes",
|
|
"Model facts exceed the 32 MiB retained representation budget. This release has no runtime override.",
|
|
),
|
|
);
|
|
}
|
|
}
|