Merge pull request #348 from martindevans/new_object_based_sampling_pipeline

martindevans · web-flow · commit d87d654a3421 · 2023-12-11T21:40:53.000Z
Custom Sampling Pipelines
diff --git a/LLama.Unittest/GrammarParserTest.cs b/LLama.Unittest/GrammarParserTest.cs
@@ -1,5 +1,4 @@
-﻿using System.Text;
-using LLama.Exceptions;
+﻿using LLama.Exceptions;
 using LLama.Native;
 using LLama.Grammars;
 
diff --git a/LLama.Unittest/StatelessExecutorTest.cs b/LLama.Unittest/StatelessExecutorTest.cs
@@ -1,5 +1,6 @@
 using System.Diagnostics;
 using LLama.Common;
+using LLama.Sampling;
 using Xunit.Abstractions;
 
 namespace LLama.Unittest
@@ -30,10 +31,13 @@ public void Dispose()
         [Fact]
         public async Task Stateless()
         {
+            // Create a custom pipeline that mimics the default pipeline
+            var pipeline = new DefaultSamplingPipeline();
+
             var executor = new StatelessExecutor(_weights, _params);
 
             const string question = "Question. what is a cat?\nAnswer: ";
-            var @params = new InferenceParams { MaxTokens = 32, AntiPrompts = new[] { "." } };
+            var @params = new InferenceParams { MaxTokens = 32, AntiPrompts = new[] { "." }, SamplingPipeline = pipeline };
 
             var timer = new Stopwatch();
             timer.Start();
diff --git a/LLama.Web/Common/InferenceOptions.cs b/LLama.Web/Common/InferenceOptions.cs
@@ -1,6 +1,9 @@
-﻿using LLama.Common;
+﻿#nullable enable
+
+using LLama.Common;
 using LLama.Abstractions;
 using LLama.Native;
+using LLama.Sampling;
 
 namespace LLama.Web.Common
 {
@@ -64,6 +67,9 @@ public class InferenceOptions
         /// <summary>
         /// A grammar to constrain possible tokens
         /// </summary>
-        public SafeLLamaGrammarHandle Grammar { get; set; } = null;
+        public SafeLLamaGrammarHandle? Grammar { get; set; }
+
+        /// <inheritdoc />
+        public ISamplingPipeline? SamplingPipeline { get; set; }
     }
 }
diff --git a/LLama/Abstractions/IInferenceParams.cs b/LLama/Abstractions/IInferenceParams.cs
@@ -1,6 +1,7 @@
 ﻿using System.Collections.Generic;
 using LLama.Common;
 using LLama.Native;
+using LLama.Sampling;
 
 namespace LLama.Abstractions
 {  
@@ -108,5 +109,10 @@ public interface IInferenceParams
 		/// Grammar to constrain possible tokens
 		/// </summary>
 		SafeLLamaGrammarHandle? Grammar { get; set; }
+
+		/// <summary>
+		/// Set a custom sampling pipeline to use. <b>If this is set All other sampling parameters are ignored!</b>
+		/// </summary>
+		ISamplingPipeline? SamplingPipeline { get; set; }
 	}
 }
diff --git a/LLama/Common/InferenceParams.cs b/LLama/Common/InferenceParams.cs
@@ -2,6 +2,7 @@
 using System;
 using System.Collections.Generic;
 using LLama.Native;
+using LLama.Sampling;
 
 namespace LLama.Common
 {
@@ -76,6 +77,9 @@ public record InferenceParams
 
         /// <inheritdoc />
         public SafeLLamaGrammarHandle? Grammar { get; set; }
+
+        /// <inheritdoc />
+        public ISamplingPipeline? SamplingPipeline { get; set; }
     }
 
     /// <summary>
diff --git a/LLama/LLamaContext.cs b/LLama/LLamaContext.cs
@@ -10,6 +10,7 @@
 using System.Runtime.InteropServices;
 using LLama.Extensions;
 using LLama.Abstractions;
+using LLama.Sampling;
 using Microsoft.Extensions.Logging;
 
 namespace LLama
@@ -212,6 +213,17 @@ public void LoadState(State state)
             }
         }
 
+        /// <summary>
+        /// Sample a single token from this context, using the given sampling pipeline
+        /// </summary>
+        /// <param name="pipeline">The pipeline to use to process the logits and to select a token</param>
+        /// <param name="lastTokens">The tokens recently returned from the model</param>
+        /// <returns>The selected token</returns>
+        public llama_token Sample(ISamplingPipeline pipeline, ReadOnlySpan<llama_token> lastTokens)
+        {
+            return pipeline.Sample(NativeHandle, NativeHandle.GetLogits(), lastTokens);
+        }
+
         /// <summary>
         /// Perform the sampling. Please don't use it unless you fully know what it does.
         /// </summary>
diff --git a/LLama/LLamaInstructExecutor.cs b/LLama/LLamaInstructExecutor.cs
@@ -210,16 +210,24 @@ protected override Task InferInternal(IInferenceParams inferenceParams, InferSta
                     SaveSessionFile(_pathSession);
                 }
 
-                var tokenDataArray = Context.ApplyPenalty(_last_n_tokens, inferenceParams.LogitBias, repeat_last_n,
-                    inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
+                llama_token id;
+                if (inferenceParams.SamplingPipeline is not null)
+                {
+                    id = inferenceParams.SamplingPipeline.Sample(Context.NativeHandle, Context.NativeHandle.GetLogits(), _last_n_tokens.ToArray());
+                }
+                else
+                {
+                    var tokenDataArray = Context.ApplyPenalty(_last_n_tokens, inferenceParams.LogitBias, repeat_last_n,
+                        inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
 
-                var mu = MirostatMu;
-                var id = Context.Sample(
-                    tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
-                    inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
-                    inferenceParams.MinP
-                );
-                MirostatMu = mu;
+                    var mu = MirostatMu;
+                    id = Context.Sample(
+                        tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
+                        inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
+                        inferenceParams.MinP
+                    );
+                    MirostatMu = mu;
+                }
 
                 _last_n_tokens.Enqueue(id);
 
diff --git a/LLama/LLamaInteractExecutor.cs b/LLama/LLamaInteractExecutor.cs
@@ -189,16 +189,24 @@ protected override async Task InferInternal(IInferenceParams inferenceParams, In
                     SaveSessionFile(_pathSession);
                 }
 
-                var tokenDataArray = Context.ApplyPenalty(_last_n_tokens, inferenceParams.LogitBias, repeat_last_n,
-                    inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
-
-                var mu = MirostatMu;
-                var id = Context.Sample(
-                    tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
-                    inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
-                    inferenceParams.MinP
-                );
-                MirostatMu = mu;
+                llama_token id;
+                if (inferenceParams.SamplingPipeline is not null)
+                {
+                    id = inferenceParams.SamplingPipeline.Sample(Context.NativeHandle, Context.NativeHandle.GetLogits(), _last_n_tokens.ToArray());
+                }
+                else
+                {
+                    var tokenDataArray = Context.ApplyPenalty(_last_n_tokens, inferenceParams.LogitBias, repeat_last_n,
+                        inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
+
+                    var mu = MirostatMu;
+                    id = Context.Sample(
+                        tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
+                        inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
+                        inferenceParams.MinP
+                    );
+                    MirostatMu = mu;
+                }
 
                 _last_n_tokens.Enqueue(id);
 
diff --git a/LLama/LLamaStatelessExecutor.cs b/LLama/LLamaStatelessExecutor.cs
@@ -7,6 +7,7 @@
 using System.Threading;
 using System.Threading.Tasks;
 using LLama.Native;
+using LLama.Sampling;
 using Microsoft.Extensions.Logging;
 
 namespace LLama
@@ -85,16 +86,24 @@ public async IAsyncEnumerable<string> InferAsync(string prompt, IInferenceParams
             var max_tokens = inferenceParams.MaxTokens < 0 ? int.MaxValue : inferenceParams.MaxTokens;
             for(var i = 0; i < max_tokens && !cancellationToken.IsCancellationRequested; i++)
             {
-                // Penalize the generated tokens by various penalties
-                var tokenDataArray = Context.ApplyPenalty(lastTokens, inferenceParams.LogitBias, repeat_last_n,
-                    inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
-
-                // Sample a single token
-                var id = Context.Sample(
-                    tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
-                    inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
-                    inferenceParams.MinP
-                );
+                llama_token id;
+                if (inferenceParams.SamplingPipeline is not null)
+                {
+                    id = inferenceParams.SamplingPipeline.Sample(Context.NativeHandle, Context.NativeHandle.GetLogits(), lastTokens);
+                }
+                else
+                {
+                    // Penalize the generated tokens by various penalties
+                    var tokenDataArray = Context.ApplyPenalty(lastTokens, inferenceParams.LogitBias, repeat_last_n,
+                        inferenceParams.RepeatPenalty, inferenceParams.FrequencyPenalty, inferenceParams.PresencePenalty, inferenceParams.PenalizeNL);
+
+                    // Sample a single token
+                    id = Context.Sample(
+                        tokenDataArray, ref mu, inferenceParams.Temperature, inferenceParams.Mirostat, inferenceParams.MirostatTau,
+                        inferenceParams.MirostatEta, inferenceParams.TopK, inferenceParams.TopP, inferenceParams.TfsZ, inferenceParams.TypicalP, inferenceParams.Grammar,
+                        inferenceParams.MinP
+                    );
+                }
 
                 // Decode this token into text
                 decoder.Add(id);
diff --git a/LLama/Native/LLamaTokenDataArray.cs b/LLama/Native/LLamaTokenDataArray.cs
@@ -46,14 +46,41 @@ public static LLamaTokenDataArray Create(ReadOnlySpan<float> logits)
             return new LLamaTokenDataArray(candidates);
         }
 
+        /// <summary>
+        /// Overwrite the logit values for all given tokens
+        /// </summary>
+        /// <param name="values">tuples of token and logit value to overwrite</param>
+        public void OverwriteLogits(ReadOnlySpan<(llama_token token, float logit)> values)
+        {
+            if (values.Length == 0)
+                return;
+
+            var dataSpan = data.Span;
+            foreach (var (token, value) in values)
+            {
+                for (var i = 0; i < data.Length; i++)
+                {
+                    if (dataSpan[i].id == token)
+                    {
+                        dataSpan[i].logit = value;
+                        break;
+                    }
+                }   
+            }
+            sorted = false;
+        }
+
         #region sampling
         /// <summary>
         /// Apply grammar rules to candidate tokens
         /// </summary>
         /// <param name="ctx"></param>
         /// <param name="grammar"></param>
-        public void ApplyGrammar(SafeLLamaContextHandle ctx, SafeLLamaGrammarHandle grammar)
+        public void ApplyGrammar(SafeLLamaContextHandle ctx, SafeLLamaGrammarHandle? grammar)
         {
+            if (grammar == null)
+                return;
+
             using (LLamaTokenDataArrayNative.Create(this, out var st))
             {
                 NativeApi.llama_sample_grammar(ctx, ref st, grammar);
@@ -145,15 +172,17 @@ public void LocallyTypical(SafeLLamaContextHandle context, float p, ulong min_ke
         /// <param name="penalty_repeat"></param>
         /// <param name="penalty_freq"></param>
         /// <param name="penalty_present"></param>
-        public void RepetitionPenalty(SafeLLamaContextHandle context, Memory<llama_token> last_tokens, float penalty_repeat, float penalty_freq, float penalty_present)
+        public void RepetitionPenalty(SafeLLamaContextHandle context, ReadOnlySpan<llama_token> last_tokens, float penalty_repeat, float penalty_freq, float penalty_present)
         {
             unsafe
             {
                 using (LLamaTokenDataArrayNative.Create(this, out var st))
-                using (var last_tokens_handle = last_tokens.Pin())
                 {
-                    NativeApi.llama_sample_repetition_penalties(context, ref st, (int*)last_tokens_handle.Pointer, (ulong)last_tokens.Length, penalty_repeat, penalty_freq, penalty_present);
-                    sorted = st.sorted;
+                    fixed (int* last_tokens_handle = last_tokens)
+                    {
+                        NativeApi.llama_sample_repetition_penalties(context, ref st, last_tokens_handle, (ulong)last_tokens.Length, penalty_repeat, penalty_freq, penalty_present);
+                        sorted = st.sorted;
+                    }
                 }
             }
         }
diff --git a/LLama/Sampling/BaseSamplingPipeline.cs b/LLama/Sampling/BaseSamplingPipeline.cs
diff --git a/LLama/Sampling/DefaultSamplingPipeline.cs b/LLama/Sampling/DefaultSamplingPipeline.cs
diff --git a/LLama/Sampling/ISamplingPipeline.cs b/LLama/Sampling/ISamplingPipeline.cs

Original file line number	Diff line number	Diff line change
`@@ -1,6 +1,7 @@`
`1`	`1`	`using System.Collections.Generic;`
`2`	`2`	`using LLama.Common;`
`3`	`3`	`using LLama.Native;`
	`4`	`+using LLama.Sampling;`
`4`	`5`
`5`	`6`	`namespace LLama.Abstractions`
`6`	`7`	`{`
`@@ -108,5 +109,10 @@ public interface IInferenceParams`
`108`	`109`	`/// Grammar to constrain possible tokens`
`109`	`110`	`/// </summary>`
`110`	`111`	`SafeLLamaGrammarHandle? Grammar { get; set; }`
	`112`	`+`
	`113`	`+ /// <summary>`
	`114`	`+ /// Set a custom sampling pipeline to use. <b>If this is set All other sampling parameters are ignored!</b>`
	`115`	`+ /// </summary>`
	`116`	`+ ISamplingPipeline? SamplingPipeline { get; set; }`
`111`	`117`	`}`
`112`	`118`	`}`
Original file line number	Diff line number	Diff line change
`@@ -2,6 +2,7 @@`
`2`	`2`	`using System;`
`3`	`3`	`using System.Collections.Generic;`
`4`	`4`	`using LLama.Native;`
	`5`	`+using LLama.Sampling;`
`5`	`6`
`6`	`7`	`namespace LLama.Common`
`7`	`8`	`{`
`@@ -76,6 +77,9 @@ public record InferenceParams`
`76`	`77`
`77`	`78`	`/// <inheritdoc />`
`78`	`79`	`public SafeLLamaGrammarHandle? Grammar { get; set; }`
	`80`	`+`
	`81`	`+ /// <inheritdoc />`
	`82`	`+ public ISamplingPipeline? SamplingPipeline { get; set; }`
`79`	`83`	`}`
`80`	`84`
`81`	`85`	`/// <summary>`