Added lots of comments to all the LLamaFtype variants

LLama/Native/LLamaFtype.cs

@@ -1,29 +1,109 @@
-using System;
-using System.Collections.Generic;
-using System.Text;
-
-namespace LLama.Native
+namespace LLama.Native
 {
+    /// <summary>
+    /// Supported model file types
+    /// </summary>
     public enum LLamaFtype
     {
+        /// <summary>
+        /// All f32
+        /// </summary>
+        /// <remarks>Benchmark: 26GB @ 7B parameters</remarks>
         LLAMA_FTYPE_ALL_F32 = 0,
-        LLAMA_FTYPE_MOSTLY_F16 = 1,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q2_K = 10,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_S = 11,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_M = 12,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q3_K_L = 13,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_S = 14,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_K_M = 15,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,// except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q6_K = 18,// except 1d tensors
+
+        /// <summary>
+        /// Mostly f16
+        /// </summary>
+        /// <remarks>Benchmark: 13GB @ 7B parameters</remarks>
+        LLAMA_FTYPE_MOSTLY_F16 = 1,
+
+        /// <summary>
+        /// Mostly 8 bit
+        /// </summary>
+        /// <remarks>Benchmark: 6.7GB @ 7B parameters, +0.0004ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,
+
+        /// <summary>
+        /// Mostly 4 bit
+        /// </summary>
+        /// <remarks>Benchmark: 3.50GB @ 7B parameters, +0.2499 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,
+
+        /// <summary>
+        /// Mostly 4 bit
+        /// </summary>
+        /// <remarks>Benchmark: 3.90GB @ 7B parameters, +0.1846 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,
+
+        /// <summary>
+        /// Mostly 4 bit, tok_embeddings.weight and output.weight are f16
+        /// </summary>
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4,
+
+        /// <summary>
+        /// Mostly 5 bit
+        /// </summary>
+        /// <remarks>Benchmark: 4.30GB @ 7B tokens, +0.0796 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,
+
+        /// <summary>
+        /// Mostly 5 bit
+        /// </summary>
+        /// <remarks>Benchmark: 4.70GB @ 7B parameters, +0.0415 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,
+
+        /// <summary>
+        /// K-Quant 2 bit
+        /// </summary>
+        /// <remarks>Benchmark: 2.67GB @ 7N parameters, +0.8698 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q2_K = 10,
+
+        /// <summary>
+        /// K-Quant 3 bit (Small)
+        /// </summary>
+        /// <remarks>Benchmark: 2.75GB @ 7B parameters, +0.5505 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q3_K_S = 11,
+
+        /// <summary>
+        /// K-Quant 3 bit (Medium)
+        /// </summary>
+        /// <remarks>Benchmark: 3.06GB @ 7B parameters, +0.2437 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q3_K_M = 12,
+
+        /// <summary>
+        /// K-Quant 3 bit (Large)
+        /// </summary>
+        /// <remarks>Benchmark: 3.35GB @ 7B parameters, +0.1803 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q3_K_L = 13,
+
+        /// <summary>
+        /// K-Quant 4 bit (Small)
+        /// </summary>
+        /// <remarks>Benchmark: 3.56GB @ 7B parameters, +0.1149 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q4_K_S = 14,
+
+        /// <summary>
+        /// K-Quant 4 bit (Medium)
+        /// </summary>
+        /// <remarks>Benchmark: 3.80GB @ 7B parameters, +0.0535 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q4_K_M = 15,
+
+        /// <summary>
+        /// K-Quant 5 bit (Small)
+        /// </summary>
+        /// <remarks>Benchmark: 4.33GB @ 7B parameters, +0.0353 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q5_K_S = 16,
+
+        /// <summary>
+        /// K-Quant 5 bit (Medium)
+        /// </summary>
+        /// <remarks>Benchmark: 4.45GB @ 7B parameters, +0.0142 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q5_K_M = 17,
+
+        /// <summary>
+        /// K-Quant 6 bit
+        /// </summary>
+        /// <remarks>Benchmark: 5.15GB @ 7B parameters, +0.0044 ppl</remarks>
+        LLAMA_FTYPE_MOSTLY_Q6_K = 18,
     }
 }

LLama/Native/LLamaModelQuantizeParams.cs

@@ -1,25 +1,28 @@
-using System;
-using System.Collections.Generic;
-using System.Runtime.InteropServices;
-using System.Text;
+using System.Runtime.InteropServices;
 
 namespace LLama.Native
 {
+    /// <summary>
+    /// Quantizer parameters used in the native API
+    /// </summary>
     public struct LLamaModelQuantizeParams
     {
         /// <summary>
         /// number of threads to use for quantizing, if &lt;=0 will use std::thread::hardware_concurrency()
         /// </summary>
         public int nthread;
+
         /// <summary>
         /// quantize to this llama_ftype
         /// </summary>
         public LLamaFtype ftype;
+
         /// <summary>
         /// allow quantizing non-f32/f16 tensors
         /// </summary>
         [MarshalAs(UnmanagedType.I1)]
         public bool allow_requantize;
+
         /// <summary>
         /// quantize output.weight
         /// </summary>