LLamaSharp/LLama/LLamaQuantizer.cs

154 lines
6.7 KiB
C#

using LLama.Native;
using System;
using System.Collections.Generic;
namespace LLama
{
/// <summary>
/// The quantizer to quantize the model.
/// </summary>
public static class LLamaQuantizer
{
/// <summary>
/// Quantize the model.
/// </summary>
/// <param name="srcFileName">The model file to be quantized.</param>
/// <param name="dstFilename">The path to save the quantized model.</param>
/// <param name="ftype">The type of quantization.</param>
/// <param name="nthread">Thread to be used during the quantization. By default it's the physical core number.</param>
/// <param name="allowRequantize"></param>
/// <param name="quantizeOutputTensor"></param>
/// <returns>Whether the quantization is successful.</returns>
/// <exception cref="ArgumentException"></exception>
public static bool Quantize(
string srcFileName, string dstFilename, LLamaFtype ftype, int nthread = -1, bool allowRequantize = true, bool quantizeOutputTensor = false)
{
if (!ValidateFtype(ftype))
{
throw new ArgumentException($"The type {Enum.GetName(typeof(LLamaFtype), ftype)} is not a valid type " +
$"to perform quantization.");
}
var quantizeParams = LLamaModelQuantizeParams.Default();
quantizeParams.ftype = ftype;
quantizeParams.nthread = nthread;
quantizeParams.allow_requantize = allowRequantize;
quantizeParams.quantize_output_tensor = quantizeOutputTensor;
// todo: fill in other quantize params fields.
// This method could probably do with a redesign - passing in a config object (maybe directly
// expose `LLamaModelQuantizeParams`) instead of an ever growing list of method parameters!
return NativeApi.llama_model_quantize(srcFileName, dstFilename, ref quantizeParams) == 0;
}
/// <summary>
/// Quantize the model.
/// </summary>
/// <param name="srcFileName">The model file to be quantized.</param>
/// <param name="dstFilename">The path to save the quantized model.</param>
/// <param name="ftype">The type of quantization.</param>
/// <param name="nthread">Thread to be used during the quantization. By default it's the physical core number.</param>
/// <param name="allowRequantize"></param>
/// <param name="quantizeOutputTensor"></param>
/// <returns>Whether the quantization is successful.</returns>
/// <exception cref="ArgumentException"></exception>
public static bool Quantize(string srcFileName, string dstFilename, string ftype, int nthread = -1, bool allowRequantize = true,
bool quantizeOutputTensor = false)
{
return Quantize(srcFileName, dstFilename, StringToFtype(ftype), nthread, allowRequantize, quantizeOutputTensor);
}
private static bool ValidateFtype(LLamaFtype ftype)
{
// Validation copies from here:
// https://github.com/ggerganov/llama.cpp/blob/f7001ccc5aa359fcf41bba19d1c99c3d25c9bcc7/llama.cpp#L13450
switch (ftype)
{
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q4_0:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q4_1:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q5_0:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q5_1:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q8_0:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_F16:
case LLamaFtype.LLAMA_FTYPE_ALL_F32:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q2_K_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q2_K:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ3_K_XS:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q3_K_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q3_K_M:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q3_K_L:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q4_K_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q4_K_M:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q5_K_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q5_K_M:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q6_K:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ2_XXS:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ2_XS:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ2_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ2_M:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ3_XXS:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ1_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ1_M:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ4_NL:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ4_XS:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ3_S:
case LLamaFtype.LLAMA_FTYPE_MOSTLY_IQ3_M:
return true;
case LLamaFtype.LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
case LLamaFtype.LLAMA_FTYPE_GUESSED:
default:
return false;
}
}
/// <summary>
/// Parse a string into a LLamaFtype. This is a "relaxed" parsing, which allows any string which is contained within
/// the enum name to be used.
///
/// For example "Q5_K_M" will convert to "LLAMA_FTYPE_MOSTLY_Q5_K_M"
/// </summary>
/// <param name="str"></param>
/// <returns></returns>
/// <exception cref="ArgumentException"></exception>
private static LLamaFtype StringToFtype(string str)
{
// Find all variants which contain the input string
var matches = new List<LLamaFtype>();
foreach (LLamaFtype ftype in Enum.GetValues(typeof(LLamaFtype)))
{
var name = Enum.GetName(typeof(LLamaFtype), ftype);
// Note: this is using "IndexOf" instead of "Contains" to be compatible with netstandard2.0
#pragma warning disable CA2249
if (name != null && name.IndexOf(str, StringComparison.OrdinalIgnoreCase) >= 0)
matches.Add(ftype);
#pragma warning restore CA2249
}
// If there was just one match, success!
if (matches.Count == 1)
return matches[0];
// If none matched throw a generic error
if (matches.Count == 0)
throw new ArgumentException($"Unknown ftype \"{str}\" for quantization.");
// There were several matches, throw an error asking the user to be more specific
throw new ArgumentException($"\"{str}\" matches multiple potential ftypes: {string.Join(",", matches)}");
}
}
}