UAIX.LmRuntime / Package guide

UAIX.LmRuntime.Sampling

Greedy and probability sampling, logit transforms, deterministic random state, top-k/top-p filtering, stop matching, and generation control.

Required For token selection and stop handling

UAIX.LmRuntime.Sampling

Greedy and probability sampling, logit transforms, deterministic random state, top-k/top-p filtering, stop matching, and generation control.

Open on NuGet Package family

Overview

Samplers and logits processors for pure C# local LLM runtime generation.

Who should use it Generation loops, model runtimes, and tests that need explicit token selection and termination behavior.

Install

.NET CLI

dotnet add package UAIX.LmRuntime.Sampling

Project file

<PackageReference Include="UAIX.LmRuntime.Sampling" />

Version policy: The documentation deliberately omits UAIX.LmRuntime package version numbers. Resolve and pin versions through your normal dependency-management and lock-file process.

Review the current package metadata, frameworks, dependencies, and downloads on NuGet ↗

Direct package dependencies

UAIX.LmRuntime.AbstractionsNuGet ↗

Package role and boundaries

Required For token selection and stop handling

You need deterministic greedy argmax selection.
You need seeded temperature, top-k, top-p, minimum-p, repetition, frequency, presence, or logit-bias processing.
You need byte-safe stop-sequence matching or a bounded generation controller.

Boundary

Computing logits or decoding model weights.
Treating a seed as sufficient reproducibility when model execution, tokenizer behavior, or hardware math differs.

Selection is separate from logits

The package receives a vocabulary-sized logit span. Model execution remains outside the sampling layer.

State is explicit

SamplingState owns deterministic random state and token-frequency history, so callers control whether state persists across generation steps.

Stops are byte-aware

StopSequenceMatcher retains partial UTF-8 byte prefixes across appended chunks and distinguishes visible output from matched stop bytes.

Key types

These are the main entry points for this package. The generated reference below includes every documented type and member represented by public package XML documentation.

GreedySampler ProbabilitySampler SamplingOptions SamplingState LogitProcessor TopKSelector StopSequenceMatcher GenerationController Xoshiro256StarStar

Coding examples

Examples use public package signatures documented on LMRuntime.com. Model paths, hashes, byte counts, prompts, and host-specific identifiers remain application inputs.

Select the greedy token

Choose the highest logit without allocating a probability distribution.

GreedySamplingExample.cs

using UAIX.LmRuntime.Sampling;

ReadOnlySpan<float> logits = [-1.2f, 0.5f, 2.75f, 1.1f];
int tokenId = GreedySampler.Select(logits);

Console.WriteLine($"Selected token: {tokenId}");

Run reproducible probability sampling

Create SamplingState once per generation sequence so random and repetition state advance together.

ProbabilitySamplingExample.cs

using UAIX.LmRuntime.Sampling;

public static class ReproducibleSamplingExample
{
    /// <summary>
    /// Selects one token from logits using an explicitly seeded sampling state.
    /// </summary>
    /// <param name="logits">The vocabulary logits for the current decoding step.</param>
    /// <returns>The selected token and candidate evidence.</returns>
    public static SamplingDecision Select(ReadOnlySpan<float> logits)
    {
        if (logits.IsEmpty)
        {
            throw new ArgumentException("At least one logit is required.", nameof(logits));
        }

        var options = new SamplingOptions
        {
            Temperature = 0.8f,
            TopK = Math.Min(40, logits.Length),
            TopP = 0.95f,
            MinimumP = 0.05f,
            RepetitionPenalty = 1.1f,
            FrequencyPenalty = 0,
            PresencePenalty = 0,
            MinimumGeneratedTokens = 0,
            MaximumGeneratedTokens = 128,
            MaximumContextTokens = 8_192,
            Seed = 42
        };

        options.Validate(vocabularySize: logits.Length);

        var state = new SamplingState(options);
        SamplingDecision decision =
            ProbabilitySampler.Select(logits, options, state);

        state.RecordToken(decision.TokenId);
        return decision;
    }
}

Boundary: Reproducibility also requires the same logits, tokenizer, settings, token history, and runtime math path.

Apply logit bias and history penalties

Inspect the transformed logits independently from token selection.

LogitProcessingExample.cs

using UAIX.LmRuntime.Sampling;

public static class LogitProcessingExample
{
    /// <summary>
    /// Applies caller-selected bias and history penalties to one logit vector.
    /// </summary>
    /// <param name="logits">The unmodified vocabulary logits.</param>
    /// <returns>A processed copy suitable for token selection.</returns>
    public static float[] Process(ReadOnlySpan<float> logits)
    {
        if (logits.Length < 2)
        {
            throw new ArgumentException(
                "At least two logits are required for this example.",
                nameof(logits));
        }

        var options = new SamplingOptions
        {
            Temperature = 1.0f,
            RepetitionPenalty = 1.15f,
            FrequencyPenalty = 0.2f,
            PresencePenalty = 0.1f,
            LogitBias = new Dictionary<int, float>
            {
                [0] = -2.0f,
                [1] = 1.5f
            },
            Seed = 7
        };

        options.Validate(vocabularySize: logits.Length);

        var state = new SamplingState(options);
        state.RecordToken(0);
        state.RecordToken(0);

        return LogitProcessor.Process(logits, options, state);
    }
}

Match a stop sequence across byte chunks

Detect a stop string even when its UTF-8 bytes span multiple generated tokens.

StopSequenceExample.cs

using System.Text;
using UAIX.LmRuntime.Sampling;

var matcher = new StopSequenceMatcher(
    stopSequences: ["</answer>"],
    includeMatchedBytes: false);

StopSequenceMatchResult first = matcher.Append(
    Encoding.UTF8.GetBytes("Result</ans"));

StopSequenceMatchResult second = matcher.Append(
    Encoding.UTF8.GetBytes("wer>ignored"));

byte[] remaining = matcher.Complete();

Console.Write(Encoding.UTF8.GetString(first.VisibleBytes));
Console.Write(Encoding.UTF8.GetString(second.VisibleBytes));
Console.Write(Encoding.UTF8.GetString(remaining));

Open this package on NuGet Compare the full package family →

Generated API reference

Expand a type to review its documented fields, properties, constructors, methods, parameter descriptions, and return descriptions. Browser Find also works across the closed detail elements.

GenerationControllerUAIX.LmRuntime.Sampling 9 members

Enforces stop, limit, cancellation, usage, and output-publication boundaries for one generation.

Property FinishReason

Gets the first decisive finish reason, or UAIX.LmRuntime.Sampling.GenerationFinishReason.None while active.

Property OutputTokenIds

Gets generated token identifiers retained under stop-token emission policy.

Property State

Gets the session-local sampling state used by the same generation.

Property Usage

Gets exact tokenizer-ID usage without deriving counts from text or bytes.

Method AcceptToken(int,System.ReadOnlySpan<byte>,System.Threading.CancellationToken)

Attempts to accept and publish one generated token at a bounded cancellation point.

tokenId: The token identifier to process; it must fall within the validated vocabulary and operation-specific range.
decodedBytes: The decoded bytes sequence used by this operation; its required length, ordering, and element bounds are validated before access.
cancellationToken: Cancellation observed before the token is recorded or published.

Returns: The token/output/finish transition produced by this step.

Method CompleteVisibleBytes

Completes an otherwise active stream and flushes any bytes retained as a possible stop prefix.

Returns: Remaining visible bytes; the finish reason remains GenerationFinishReason.None.

Method Fail(System.Exception)

Records an execution failure without exposing exception details through the stable finish reason.

exception: The non-null execution failure observed by the caller.

Returns: The GenerationFinishReason result produced by GenerationController.Fail for this contract: Records an execution failure without exposing exception details through the stable finish reason. It is published only after all documented validation and ownership transitions succeed.

Method GenerationController(int,int,UAIX.LmRuntime.Sampling.SamplingOptions)

Initializes a controller from validated vocabulary and prompt-token bounds.

vocabularySize: The positive tokenizer vocabulary size used to validate token identifiers and size bounded result buffers.
promptTokenCount: The prompt token count used to bound this operation; it must be nonnegative and within the supported range.
options: The optional SamplingOptions controlling GenerationController; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.

Method ObserveCancellation(System.Threading.CancellationToken)

Observes cancellation between decode steps without publishing another token.

cancellationToken: The caller-provided token used to cancel the operation before additional work or results are published.

Returns: The GenerationFinishReason result produced by GenerationController.ObserveCancellation for this contract: Observes cancellation between decode steps without publishing another token. It is published only after all documented validation and ownership transitions succeed.

GenerationFinishReasonUAIX.LmRuntime.Sampling 7 members

Identifies the first decisive condition that ended a generation.

Field Cancelled

Cancellation was observed before publishing another token.

Field ContextLimit

The configured prompt-plus-generation context bound was reached.

Field ExecutionError

An execution error ended generation.

Field None

Generation remains active.

Field StopText

A configured decoded UTF-8 stop sequence ended generation.

Field StopToken

An exact configured token identifier ended generation.

Field TokenLimit

The configured maximum generated-token count was reached.

GenerationStepResultUAIX.LmRuntime.Sampling 3 members

Represents the observable result of attempting to publish one generated token.

Property FinishReason

Gets the stable finish reason after this step.

Property TokenAccepted

Gets whether the token identifier was retained in generated-token output.

Property VisibleBytes

Gets bytes newly safe to publish after stop-prefix matching.

GenerationUsageUAIX.LmRuntime.Sampling 3 members

Records tokenizer-ID-based usage accounting.

Property CompletionTokens

Gets the exact number of generated token identifiers accepted by the controller.

Property PromptTokens

Gets the exact number of prompt token identifiers consumed.

Property TotalTokens

Gets the checked sum of prompt and completion token counts.

GreedySamplerUAIX.LmRuntime.Sampling 1 member

Provides deterministic greedy token selection with explicit non-finite input policy.

Method Select(System.ReadOnlySpan<float>)

Selects the highest logit index with deterministic lower-index tie-breaking.

logits: The non-empty source logits. NaN is rejected; infinities compare normally.

Returns: The int value computed by GreedySampler.Select for this contract: Selects the highest logit index with deterministic lower-index tie-breaking. Range, finite-value, and overflow checks are completed before the value is returned.

LogitProcessorUAIX.LmRuntime.Sampling 4 members

Applies validated, deterministic token-history and bias policies to logits.

Method ApplyBias(System.Span<float>,System.Collections.Generic.IReadOnlyDictionary<int,float>)

Adds all validated per-token biases exactly once.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
biases: The ordered biases collection of type IReadOnlyDictionary<int, float>; LogitProcessor.ApplyBias validates nullability, count, and element constraints before consuming or snapshotting it and does not retain a mutable caller alias.

Method

ApplyHistoryPenalties(System.Span<float>,UAIX.LmRuntime.Sampling.SamplingOptions,System.Collections.Generic.IReadOnlyDictionary<int,int>)

Applies sign-aware repetition, frequency, and presence penalties in one deterministic pass.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
options: The optional SamplingOptions controlling ApplyHistoryPenalties; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.
tokenCounts: The ordered token counts collection of type IReadOnlyDictionary<int, int>; LogitProcessor.ApplyHistoryPenalties validates nullability, count, and element constraints before consuming or snapshotting it and does not retain a mutable caller alias.

Method Process(System.ReadOnlySpan<float>,UAIX.LmRuntime.Sampling.SamplingOptions,UAIX.LmRuntime.Sampling.SamplingState)

Produces a processed copy of the source logits without exposing partially mutated caller data on validation failure.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
options: The optional SamplingOptions controlling Process; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.
state: The validated state value consumed by the operation; mutations, when applicable, are limited to the explicitly documented state owner.

Returns: A newly allocated float[] containing the ordered result of LogitProcessor.Process: Produces a processed copy of the source logits without exposing partially mutated caller data on validation failure. The caller owns the returned array and later mutation cannot alter the source object.

Method SuppressEarlyStopTokens(System.Span<float>,UAIX.LmRuntime.Sampling.SamplingOptions,int)

Marks configured stop tokens ineligible before the exact minimum-token boundary.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
options: The optional SamplingOptions controlling SuppressEarlyStopTokens; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.
generatedTokenCount: The generated token count used to bound this operation; it must be nonnegative and within the supported range.

LogitScoreUAIX.LmRuntime.Sampling 2 members

Represents a scored token candidate.

Property Score

Gets the logit score.

Property TokenIndex

Gets the token index.

ProbabilitySamplerUAIX.LmRuntime.Sampling 2 members

Builds stable normalized distributions and samples them with session-local deterministic state.

Method BuildDistribution(System.ReadOnlySpan<float>,UAIX.LmRuntime.Sampling.SamplingOptions)

Builds a stable, filtered, and normalized probability distribution without consuming random state.

processedLogits: The processed logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
options: The optional SamplingOptions controlling BuildDistribution; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.

Returns: Candidates ordered by descending probability and ascending token identifier for ties.

Method Select(System.ReadOnlySpan<float>,UAIX.LmRuntime.Sampling.SamplingOptions,UAIX.LmRuntime.Sampling.SamplingState)

Processes logits and selects one token under the supplied deterministic policy.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
options: The optional SamplingOptions controlling Select; null selects the documented defaults, supplied limits are validated before allocation, and the instance is not mutated.
state: The validated state value consumed by the operation; mutations, when applicable, are limited to the explicitly documented state owner.

Returns: The SamplingDecision result produced by ProbabilitySampler.Select for this contract: Processes logits and selects one token under the supplied deterministic policy. It is published only after all documented validation and ownership transitions succeed.

SamplingCandidateUAIX.LmRuntime.Sampling 2 members

Represents one normalized token candidate retained after all filters.

Property Probability

Gets the normalized candidate probability.

Property TokenId

Gets the token identifier.

SamplingDecisionUAIX.LmRuntime.Sampling 3 members

Describes one deterministic or stochastic sampling decision.

Property Candidates

Gets the candidate distribution used for selection, ordered by probability and token identifier.

Property IsGreedy

Gets whether the zero-temperature greedy path made the decision.

Property TokenId

Gets the selected token identifier.

SamplingOptionsUAIX.LmRuntime.Sampling 17 members

Defines deterministic logit-processing, candidate-selection, and generation-stop policies for one sampling session.

Instances are treated as immutable configuration after a UAIX.LmRuntime.Sampling.SamplingState is created. Validation is intentionally performed before a logit buffer is modified so an invalid option cannot leave partially adjusted data.

Property FrequencyPenalty

Gets the amount subtracted for each prior occurrence of a token.

Property IncludeStopSequence

Gets whether matched stop-sequence bytes are included in visible output.

Property IncludeStopToken

Gets whether a matched stop token is retained in emitted token identifiers.

Property LogitBias

Gets finite per-token additive logit biases.

Property MaximumContextTokens

Gets the maximum prompt-plus-generation token count.

Property MaximumGeneratedTokens

Gets the maximum number of generated tokens. Zero permits no generated tokens.

Property MinimumGeneratedTokens

Gets the minimum generated-token count before stop-token or stop-text policies become eligible.

Property MinimumP

Gets the minimum probability relative to the highest candidate probability.

Property PresencePenalty

Gets the amount subtracted once from any token that has previously appeared.

Property RepetitionPenalty

Gets the positive sign-aware repetition penalty.

Property Seed

Gets the deterministic per-session pseudo-random generator seed.

Property StopSequences

Gets UTF-8 stop strings matched across decoded token boundaries.

Property StopTokenIds

Gets exact token identifiers that terminate generation after the minimum-token boundary.

Property Temperature

Gets the non-negative temperature. Zero selects the deterministic greedy path.

Property TopK

Gets the maximum candidate count. Zero disables top-k filtering.

Property TopP

Gets the normalized nucleus probability threshold in the inclusive range zero through one.

Method Validate(int)

Validates every option and token-indexed policy against a vocabulary size without mutating caller data.

vocabularySize: The positive number of logits accepted by the model.

SamplingStateUAIX.LmRuntime.Sampling 6 members

Stores token history and deterministic pseudo-random state for exactly one generation session.

A state instance must not be shared by independent requests. Keeping history and random state together makes session isolation explicit and prevents interleaved requests from consuming one another's random sequence.

Property GeneratedTokenCount

Gets the number of generated token identifiers recorded by this session.

Property Random

Gets the session-local deterministic random generator.

Property TokenCounts

Gets the session-owned prior-token counts through a read-only interface.

Method GetTokenCount(int)

Gets the prior count for one token without adding it to history.

tokenId: The token identifier to process; it must fall within the validated vocabulary and operation-specific range.

Returns: The recorded count, or zero when the token has not appeared.

Method RecordToken(int)

Records one generated token for repetition, frequency, presence, and usage policies.

tokenId: The non-negative tokenizer identifier generated by the model.

Method SamplingState(UAIX.LmRuntime.Sampling.SamplingOptions)

Initializes isolated state from the immutable session options.

options: The options whose seed initializes the session generator.

StopSequenceMatcherUAIX.LmRuntime.Sampling 5 members

Matches UTF-8 stop sequences across arbitrary decoded-byte boundaries while retaining only a bounded possible prefix.

Property MaximumRetainedBytes

Gets the maximum retained prefix bytes, bounded by the longest configured stop sequence.

Property RetainedByteCount

Gets the current possible stop-prefix byte count.

Method Append(System.ReadOnlySpan<byte>)

Appends one decoded byte chunk and releases bytes that can no longer participate in a stop match.

bytes: The next decoded UTF-8 bytes; chunks may split a code point or stop sequence.

Returns: The newly visible bytes and optional terminal match.

Method Complete

Completes matching and releases any retained non-matching prefix bytes.

Returns: A newly allocated byte[] containing the ordered result of StopSequenceMatcher.Complete: Completes matching and releases any retained non-matching prefix bytes. The caller owns the returned array and later mutation cannot alter the source object.

Method StopSequenceMatcher(System.Collections.Generic.IEnumerable<string>,bool)

Initializes a matcher from non-empty stop strings.

stopSequences: The exact Unicode strings encoded as UTF-8 for byte matching.
includeMatchedBytes: Whether a terminal match is included in visible output.

StopSequenceMatchResultUAIX.LmRuntime.Sampling 3 members

Represents visible bytes released by one bounded stop-sequence matching step.

Property Matched

Gets whether a terminal stop sequence has matched.

Property MatchedStopSequence

Gets the exact stop string that completed during this step, if any.

Property VisibleBytes

Gets newly visible bytes that cannot participate in a future stop match.

TopKSelectorUAIX.LmRuntime.Sampling 1 member

Provides partial top-k selection for logit arrays.

Method SelectTopK(System.ReadOnlySpan<float>,int)

Selects the highest scoring token candidates without sorting the full input.

logits: The logits sequence used by this operation; its required length, ordering, and element bounds are validated before access.
k: The numeric k consumed by SelectTopK; it must satisfy the member's documented range, geometry, and finite-value requirements.

Returns: The selected candidates in descending score order with deterministic index tie-breaking.

Xoshiro256StarStarUAIX.LmRuntime.Sampling 3 members

Implements the xoshiro256** generator with SplitMix64 seed expansion and deterministic unsigned arithmetic.

The generator is intentionally session-local and not thread-safe. Its transition follows the published xoshiro256** reference algorithm; seed expansion prevents the prohibited all-zero state for a zero seed.

Method NextUInt64

Returns the next 64-bit output and advances the generator exactly once.

Returns: The ulong value computed by Xoshiro256StarStar.NextUInt64 for this contract: Returns the next 64-bit output and advances the generator exactly once. Range, finite-value, and overflow checks are completed before the value is returned.

Method NextUnitDouble

Returns a uniformly distributed value in the half-open interval [0, 1).

Returns: The double value computed by Xoshiro256StarStar.NextUnitDouble for this contract: Returns a uniformly distributed value in the half-open interval [0, 1). Range, finite-value, and overflow checks are completed before the value is returned.

Method Xoshiro256StarStar(ulong)

Initializes the generator from one deterministic 64-bit seed.

seed: The numeric seed consumed by Xoshiro256StarStar; it must satisfy the member's documented range, geometry, and finite-value requirements.

Frequently asked questions

When should I use GreedySampler?

Use it for deterministic argmax generation, parity tests, and local facade paths that intentionally do not expose probabilistic settings.

Where should SamplingState live?

Create one state per generation sequence and retain it across steps. Starting a new state resets random progression and token-frequency history.

Does a fixed seed guarantee identical text everywhere?

No. It controls the package random generator. Identical output still depends on identical logits, processing settings, token history, tokenizer, floating-point behavior, and model execution.

Why match stop sequences as bytes?

Generated tokens may split one Unicode scalar or stop string across token boundaries. Byte-level retention avoids decoding incomplete fragments prematurely.