zerfoo/api.go at v1.60.0 · zerfoo/zerfoo · GitHub

Repository navigation

History

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

package zerfoo

import (

"context"

"encoding/json"

"fmt"

"math"

"os"

"path/filepath"

"strings"

"time"

"github.com/zerfoo/zerfoo/generate/grammar"

"github.com/zerfoo/zerfoo/inference"

"github.com/zerfoo/zerfoo/model/huggingface"

"github.com/zerfoo/zerfoo/serve"

)

// Model is a loaded language model ready for inference.

// A Model is created via [Load] and used for text generation, embedding,

// and tool-call detection. [Model.Close] must be called when the model is no

// longer needed to release GPU and CPU resources.

// Stable.

type Model struct {

inner *inference.Model

// generateFunc is an optional override for testing ChatStream.

// When non-nil, ChatStream uses this instead of inner.Generate.

generateFunc func(ctx context.Context, prompt string) (string, error)

}

// NewModel creates a Model with a custom generate function for testing

// and demonstrations without loading a real GGUF model file. The provided

// function is called by Generate, Chat, and ChatStream.

// This is useful for writing pkg.go.dev examples and unit tests:

// m := zerfoo.NewModel(func(ctx context.Context, prompt string) (string, error) {

// return "Hello from the model!", nil

// })

// result, _ := m.Generate(ctx, "Hi")

func NewModel(fn func(ctx context.Context, prompt string) (string, error)) *Model {

return &Model{generateFunc: fn}

}

// defaultQuant is the preferred quantization when none is specified.

const defaultQuant = "Q4_K_M"

// Load loads a model from a file path or HuggingFace model ID.

// Paths starting with "/", "./" or "../" are treated as local GGUF files.

// All other strings are treated as HuggingFace model IDs (e.g. "google/gemma-3-4b"

// or "google/gemma-3-4b/Q8_0"). If the model is not cached locally it will be

// downloaded from HuggingFace.

// Stable.

func Load(pathOrID string) (*Model, error) {

if isLocalPath(pathOrID) {

m, err := inference.LoadFile(pathOrID)

if err != nil {

return nil, fmt.Errorf("load %q: %w", pathOrID, err)

}

return &Model{inner: m}, nil

}

return loadFromHuggingFace(pathOrID)

}

// parseModelID splits a model identifier into repo ID and optional quant.

// Accepted formats:

// "owner/model" → ("owner/model", "Q4_K_M")

// "owner/model/Q8_0" → ("owner/model", "Q8_0")

func parseModelID(id string) (repoID, quant string) {

parts := strings.Split(id, "/")

if len(parts) == 3 {

return parts[0] + "/" + parts[1], parts[2]

}

return id, defaultQuant

}

// loadFromHuggingFace resolves a HuggingFace model ID through the cache,

// downloading the GGUF file if not already present.

func loadFromHuggingFace(id string) (*Model, error) {

repoID, quant := parseModelID(id)

manifest, err := huggingface.LoadManifest()

if err != nil {

return nil, fmt.Errorf("load %q: %w", id, err)

}

// Cache hit: model already downloaded.

if cached, ok := manifest.FindByRepo(repoID); ok {

if _, statErr := os.Stat(cached.Path); statErr == nil {

m, loadErr := inference.LoadFile(cached.Path)

if loadErr != nil {

return nil, fmt.Errorf("load cached %q: %w", cached.Path, loadErr)

}

return &Model{inner: m}, nil

}

// File missing on disk — fall through to re-download.

}

// Cache miss: resolve the best GGUF file and download it.

client := huggingface.NewClient()

fileInfo, err := client.ResolveGGUF(repoID, quant)

if err != nil {

return nil, fmt.Errorf("load %q: %w", id, err)

}

cacheDir, err := huggingface.CacheDir()

if err != nil {

return nil, fmt.Errorf("load %q: %w", id, err)

}

destPath := filepath.Join(cacheDir, repoID, fileInfo.Filename)

downloadURL := fmt.Sprintf("https://huggingface.co/%s/resolve/main/%s", repoID, fileInfo.Filename)

dl := huggingface.NewDownloader(nil)

if err := dl.Download(context.Background(), downloadURL, destPath); err != nil {

return nil, fmt.Errorf("load %q: download: %w", id, err)

}

// Update manifest.

info, _ := os.Stat(destPath)

manifest.Add(huggingface.CachedModel{

RepoID: repoID,

Filename: fileInfo.Filename,

Path: destPath,

Size: info.Size(),

})

if err := huggingface.SaveManifest(manifest); err != nil {

return nil, fmt.Errorf("load %q: save manifest: %w", id, err)

}

m, err := inference.LoadFile(destPath)

if err != nil {

return nil, fmt.Errorf("load %q: %w", id, err)

}

return &Model{inner: m}, nil

}

// isLocalPath returns true if the string looks like a local file path.

func isLocalPath(s string) bool {

return strings.HasPrefix(s, "/") || strings.HasPrefix(s, "./") || strings.HasPrefix(s, "../")

}

// Chat runs a simple one-shot generation and returns the generated text.

// Stable.

func (m *Model) Chat(prompt string) (string, error) {

result, err := m.Generate(context.Background(), prompt)

if err != nil {

return "", err

}

return result.Text, nil

}

// Generate runs text generation with the given prompt and options.

// Stable.

func (m *Model) Generate(ctx context.Context, prompt string, opts ...GenerateOption) (*GenerateResult, error) {

var gopts generateOptions

for _, o := range opts {

o(&gopts)

}

var infOpts []inference.GenerateOption

if gopts.maxTokens > 0 {

infOpts = append(infOpts, inference.WithMaxTokens(gopts.maxTokens))

}

if gopts.temperature > 0 {

infOpts = append(infOpts, inference.WithTemperature(float64(gopts.temperature)))

}

if gopts.topP > 0 && gopts.topP < 1.0 {

infOpts = append(infOpts, inference.WithTopP(float64(gopts.topP)))

}

if gopts.schema != nil {

g, err := grammar.Convert(gopts.schema)

if err != nil {

return nil, fmt.Errorf("convert schema to grammar: %w", err)

}

infOpts = append(infOpts, inference.WithGrammar(g))

}

start := time.Now()

var text string

var err error

if m.generateFunc != nil {

text, err = m.generateFunc(ctx, prompt)

} else {

text, err = m.inner.Generate(ctx, prompt, infOpts...)

}

if err != nil {

return nil, err

}

duration := time.Since(start)

// Estimate token count from the tokenizer.

tokenCount := 0

if m.inner != nil {

tok := m.inner.Tokenizer()

if tok != nil {

if ids, encErr := tok.Encode(text); encErr == nil {

tokenCount = len(ids)

}

result := &GenerateResult{

Text: text,

TokenCount: tokenCount,

Duration: duration,

}

// Detect tool calls if tools are configured.

if len(gopts.tools) > 0 {

tc := serve.ToolChoice{Mode: "auto"}

if gopts.toolChoice != nil {

tc = *gopts.toolChoice

}

if tcResult, ok := serve.DetectToolCall(text, gopts.tools, tc); ok {

result.ToolCalls = []ToolCall{{

ID: tcResult.ID,

FunctionName: tcResult.FunctionName,

Arguments: tcResult.Arguments,

}}

}

return result, nil

}

// Embed returns embeddings for the given texts.

// Each input string is tokenized, its token embeddings are looked up from the

// model's embedding table, mean-pooled, and L2-normalized.

// Stable.

func (m *Model) Embed(texts []string) ([]Embedding, error) {

if len(texts) == 0 {

return nil, nil

}

results := make([]Embedding, len(texts))

for i, text := range texts {

vec, err := m.inner.Embed(text)

if err != nil {

return nil, fmt.Errorf("embed text %d: %w", i, err)

}

results[i] = Embedding{Vector: vec}

}

return results, nil

}

// Close releases model resources.

// Stable.

func (m *Model) Close() error {

return m.inner.Close()

}

// GenerateResult holds the result of a text generation call.

// Stable.

type GenerateResult struct {

Text string

TokenCount int

Duration time.Duration

ToolCalls []ToolCall

}

// Embedding holds a text embedding vector.

// Stable.

type Embedding struct {

Vector []float32

}

// CosineSimilarity computes the cosine similarity between two embeddings.

// Stable.

func (e Embedding) CosineSimilarity(other Embedding) float32 {

if len(e.Vector) == 0 || len(e.Vector) != len(other.Vector) {

return 0

}

var dot, normA, normB float64

for i := range e.Vector {

a := float64(e.Vector[i])

b := float64(other.Vector[i])

dot += a * b

normA += a * a

normB += b * b

}

denom := math.Sqrt(normA) * math.Sqrt(normB)

if denom == 0 {

return 0

}

return float32(dot / denom)

}

// generateOptions holds parsed generation options.

type generateOptions struct {

maxTokens int

temperature float32

topP float32

tools []serve.Tool

toolChoice *serve.ToolChoice

schema *grammar.JSONSchema

}

// GenerateOption configures the behavior of [Model.Generate].

// Stable.

type GenerateOption func(*generateOptions)

// WithGenMaxTokens sets the maximum number of tokens to generate.

// Stable.

func WithGenMaxTokens(n int) GenerateOption {

return func(o *generateOptions) {

o.maxTokens = n

}

// WithGenTemperature sets the sampling temperature.

// Stable.

func WithGenTemperature(t float32) GenerateOption {

return func(o *generateOptions) {

o.temperature = t

}

// WithGenTopP sets the top-p (nucleus) sampling parameter.

// Stable.

func WithGenTopP(p float32) GenerateOption {

return func(o *generateOptions) {

o.topP = p

}

// StreamToken represents a token received during streaming generation.

// Stable.

type StreamToken struct {

Text string

Done bool

}

// ChatStream starts streaming generation and returns a receive-only channel

// that yields [StreamToken] values as they are generated. The channel is closed

// when generation completes or ctx is canceled. The error return is non-nil

// only if startup fails (e.g. the model is not loaded).

// Stable.

func (m *Model) ChatStream(ctx context.Context, prompt string, opts ...GenerateOption) (<-chan StreamToken, error) {

if m.inner == nil && m.generateFunc == nil {

return nil, fmt.Errorf("model not loaded")

}

// Build the generate function: use the override if set, otherwise wrap inner.Generate.

genFn := m.generateFunc

if genFn == nil {

var gopts generateOptions

for _, o := range opts {

o(&gopts)

}

var infOpts []inference.GenerateOption

if gopts.maxTokens > 0 {

infOpts = append(infOpts, inference.WithMaxTokens(gopts.maxTokens))

}

if gopts.temperature > 0 {

infOpts = append(infOpts, inference.WithTemperature(float64(gopts.temperature)))

}

if gopts.topP > 0 && gopts.topP < 1.0 {

infOpts = append(infOpts, inference.WithTopP(float64(gopts.topP)))

}

genFn = func(ctx context.Context, prompt string) (string, error) {

return m.inner.Generate(ctx, prompt, infOpts...)

}

ch := make(chan StreamToken, 64)

go func() {

defer close(ch)

text, err := genFn(ctx, prompt)

if err != nil {

return

}

// Split the result into words and stream them as tokens.

words := strings.Fields(text)

for i, word := range words {

if ctx.Err() != nil {

return

}

if i < len(words)-1 {

word += " "

}

select {

case ch <- StreamToken{Text: word}:

case <-ctx.Done():

return

}

// Send done signal.

select {

case ch <- StreamToken{Done: true}:

case <-ctx.Done():

}

}()

return ch, nil

}

// ToolCall represents a tool invocation detected in model output.

// Experimental.

type ToolCall struct {

ID string

FunctionName string

Arguments json.RawMessage

}

// WithTools configures the tools available for tool call detection.

// When tools are provided, [Model.Generate] will attempt to detect tool calls

// in the model output and populate [GenerateResult.ToolCalls].

// Experimental.

func WithTools(tools ...serve.Tool) GenerateOption {

return func(o *generateOptions) {

o.tools = tools

}

// WithToolChoice sets the tool choice mode for tool call detection.

// Experimental.

func WithToolChoice(choice serve.ToolChoice) GenerateOption {

return func(o *generateOptions) {

o.toolChoice = &choice

}

// WithSchema enables grammar-guided decoding.

// The model's output will be constrained to valid JSON matching the given schema.

// Experimental.

func WithSchema(schema grammar.JSONSchema) GenerateOption {

return func(o *generateOptions) {

o.schema = &schema

}

zerfoo/api.go at v1.60.0 · zerfoo/zerfoo · GitHub

Repository navigation

Expand file tree

Breadcrumbs

api.go

Latest commit

History

Breadcrumbs

api.go

File metadata and controls