SPB Git

Repository Public

ModelNaming.swift 7.0 KB 108 lines swift

1// Author: Simon-Pierre Boucher — contact@spboucher.ai2// File: Packages/LocalChatModelHub/Sources/ModelNaming.swift3// Description: Explicit rules that read quantization, parameter count and capabilities out of repository and file names.4import Foundation5import LocalChatCore67public struct QuantizationInfo: Sendable, Hashable {8    public var label: String            // "4bit", "4bit DWQ", "Q4_K_M", "bf16"9    public var bitsPerWeight: Double?   // nominal bits per weight of the scheme (not measured)10    public init(label: String, bitsPerWeight: Double?) { self.label = label; self.bitsPerWeight = bitsPerWeight }11}1213public enum ModelNaming {14    /// Nominal bits per weight of llama.cpp quantization types (llama.cpp documentation figures).15    static let ggufBits: [String: Double] = [16        "F32": 32, "F16": 16, "BF16": 16, "Q8_0": 8.5, "Q6_K": 6.56, "Q5_K_M": 5.69, "Q5_K_S": 5.54, "Q5_0": 5.5, "Q5_1": 6.0,17        "Q4_K_M": 4.85, "Q4_K_S": 4.58, "Q4_0": 4.55, "Q4_1": 5.0, "IQ4_NL": 4.5, "IQ4_XS": 4.25, "Q3_K_L": 4.27, "Q3_K_M": 3.91,18        "Q3_K_S": 3.5, "IQ3_M": 3.66, "IQ3_S": 3.44, "IQ3_XS": 3.3, "IQ3_XXS": 3.06, "Q2_K": 2.96, "Q2_K_S": 2.96, "IQ2_M": 2.7,19        "IQ2_S": 2.5, "IQ2_XS": 2.31, "IQ2_XXS": 2.06, "IQ1_M": 1.75, "IQ1_S": 1.56, "TQ1_0": 1.69, "TQ2_0": 2.06,20    ]2122    /// llama.cpp `general.file_type` values.23    static let ggufFileTypes: [Int: String] = [24        0: "F32", 1: "F16", 2: "Q4_0", 3: "Q4_1", 7: "Q8_0", 8: "Q5_0", 9: "Q5_1", 10: "Q2_K", 11: "Q3_K_S", 12: "Q3_K_M", 13: "Q3_K_L",25        14: "Q4_K_S", 15: "Q4_K_M", 16: "Q5_K_S", 17: "Q5_K_M", 18: "Q6_K", 19: "IQ2_XXS", 20: "IQ2_XS", 21: "Q2_K_S", 22: "IQ3_XS",26        23: "IQ3_XXS", 24: "IQ1_S", 25: "IQ4_NL", 26: "IQ3_S", 27: "IQ3_M", 28: "IQ2_S", 29: "IQ2_M", 30: "IQ4_XS", 31: "IQ1_M",27        32: "BF16", 36: "TQ1_0", 37: "TQ2_0",28    ]2930    private static func firstMatch(_ pattern: String, in text: String, group: Int = 1) -> String? {31        guard let regex = try? NSRegularExpression(pattern: pattern, options: [.caseInsensitive]) else { return nil }32        let range = NSRange(text.startIndex..., in: text)33        guard let m = regex.firstMatch(in: text, range: range), m.numberOfRanges > group, let r = Range(m.range(at: group), in: text) else { return nil }34        return String(text[r])35    }3637    /// GGUF quantization from a file name such as "Qwen3-4B-UD-Q4_K_XL.gguf" or "model.Q8_0.gguf".38    public static func ggufQuantization(fromName name: String) -> QuantizationInfo? {39        let pattern = "(?<![A-Za-z0-9])(IQ[1-4]_(?:XXS|XS|S|M|NL)|T?Q[1-8]_K_(?:XXL|XL|L|M|S)|Q[2-8]_K|T?Q[1-8]_[01]|BF16|F16|F32)(?![A-Za-z0-9])"40        guard let raw = firstMatch(pattern, in: name)?.uppercased() else { return nil }41        // unsloth "UD-…_XL" variants have no nominal figure: fall back to the base K type42        let bits = ggufBits[raw] ?? ggufBits[raw.replacingOccurrences(of: "_XXL", with: "_M").replacingOccurrences(of: "_XL", with: "_M")]43        return QuantizationInfo(label: raw, bitsPerWeight: bits)44    }4546    public static func ggufQuantization(fileType: Int) -> QuantizationInfo? {47        ggufFileTypes[fileType].map { QuantizationInfo(label: $0, bitsPerWeight: ggufBits[$0]) }48    }4950    /// MLX quantization from a repository name ("…-4bit", "…-8bit-DWQ", "…-bf16", "…-qat-4bit") or from config bits.51    public static func mlxQuantization(fromName name: String, configBits: Int? = nil) -> QuantizationInfo? {52        let lower = name.lowercased()53        var bits = configBits54        if bits == nil, let s = firstMatch("(?<![0-9a-z])([2-8])[-_]?bits?(?![a-z0-9])", in: lower) { bits = Int(s) }55        if let bits {56            var label = "\(bits)bit"57            if lower.contains("dwq") { label += " DWQ" } else if lower.contains("qat") { label += " QAT" }58            // MLX affine quantization stores a scale and a bias per group of 64 weights: bits + 32/6459            return QuantizationInfo(label: label, bitsPerWeight: Double(bits) + 0.5)60        }61        if let f = firstMatch("(?<![a-z0-9])(bf16|fp16|f16|fp32|f32)(?![a-z0-9])", in: lower) {62            return QuantizationInfo(label: f, bitsPerWeight: f.hasSuffix("32") ? 32 : 16)63        }64        return nil65    }6667    /// Parameter count written in a name: "7B", "1.5b", "350M", "30B-A3B" (→ 30B). Returns nil when nothing unambiguous is found.68    public static func parameterCount(fromName name: String) -> Int64? {69        guard let regex = try? NSRegularExpression(pattern: "(?<![A-Za-z0-9.])([0-9]+(?:\\.[0-9]+)?)([BbMm])(?![A-Za-z0-9])") else { return nil }70        let base = (name as NSString).lastPathComponent71        for m in regex.matches(in: base, range: NSRange(base.startIndex..., in: base)) {72            guard let nr = Range(m.range(at: 1), in: base), let ur = Range(m.range(at: 2), in: base), let value = Double(base[nr]) else { continue }73            let unit = base[ur].lowercased()74            let count = value * (unit == "b" ? 1e9 : 1e6)75            // plausible language-model sizes only (rejects things like "2M" context hints or version numbers)76            if count >= 5e6, count <= 2e12 { return Int64(count) }77        }78        return nil79    }8081    static let codeHints = ["coder", "codestral", "devstral", "starcoder", "codegemma", "codellama", "deepseek-coder", "-code-", "codeqwen", "mellum"]82    static let reasoningHints = ["deepseek-r1", "qwq", "thinking", "reasoning", "-r1-", "gpt-oss", "magistral"]83    static let embeddingHints = ["embed", "bge-", "e5-", "gte-", "minilm", "arctic-embed", "colbert"]8485    /// Capabilities with one explicit rule each; nothing is assumed beyond what the Hub metadata or the name states.86    public static func capabilities(name: String, pipelineTag: String?, tags: [String], chatTemplate: String?, hasVisionConfig: Bool,87                                    hasProjector: Bool = false) -> Set<ModelCapability> {88        let lower = name.lowercased()89        let embeddingPipeline = pipelineTag == "sentence-similarity" || pipelineTag == "feature-extraction"90        let rerank = pipelineTag == "text-ranking" || lower.contains("rerank")91        if embeddingPipeline || (!rerank && chatTemplate == nil && embeddingHints.contains { lower.contains($0) }) || (lower.contains("embedding") && !rerank) {92            return [.embeddings]93        }94        var caps: Set<ModelCapability> = [.chat]95        if pipelineTag == "image-text-to-text" || pipelineTag == "video-text-to-text" || pipelineTag == "any-to-any" || hasVisionConfig || hasProjector {96            caps.insert(.vision)97        }98        if pipelineTag == "any-to-any" || tags.contains("audio-text-to-text") || pipelineTag == "audio-text-to-text" { caps.insert(.audio) }99        if let t = chatTemplate {100            if t.contains("tools") || t.contains("tool_call") { caps.insert(.tools) }101            if t.contains("enable_thinking") || t.contains("<think>") || t.contains("reasoning_content") { caps.insert(.reasoning) }102        }103        if reasoningHints.contains(where: { lower.contains($0) }) { caps.insert(.reasoning) }104        if codeHints.contains(where: { lower.contains($0) }) { caps.insert(.code) }105        return caps106    }107}108
Files main
Menu

Theme