|
| 1 | +import Foundation |
| 2 | +import llama |
| 3 | + |
| 4 | +let arguments = CommandLine.arguments |
| 5 | + |
| 6 | +// Check that we have at least one argument (the model path) |
| 7 | +guard arguments.count > 1 else { |
| 8 | + print("Usage: swift MODEL_PATH [PROMPT] [PARALLEL]") |
| 9 | + exit(1) |
| 10 | +} |
| 11 | + |
| 12 | +let modelPath: String = arguments[1] |
| 13 | +let prompt: String = arguments.count > 2 ? arguments[2] : "Hello my name is" |
| 14 | +let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(arguments[3])! : 1 |
| 15 | + |
| 16 | +// total length of the sequences including the prompt |
| 17 | +let n_len: Int = 32 |
| 18 | + |
| 19 | +// init LLM |
| 20 | +llama_backend_init(false) |
| 21 | +defer { |
| 22 | + llama_backend_free() |
| 23 | +} |
| 24 | + |
| 25 | +let model_params = llama_model_default_params() |
| 26 | +guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), model_params) else { |
| 27 | + print("Failed to load model") |
| 28 | + exit(1) |
| 29 | +} |
| 30 | + |
| 31 | +defer { |
| 32 | + llama_free_model(model) |
| 33 | +} |
| 34 | + |
| 35 | +var tokens = tokenize(text: prompt, add_bos: true) |
| 36 | + |
| 37 | +let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel) |
| 38 | + |
| 39 | +var context_params = llama_context_default_params() |
| 40 | +context_params.seed = 1234 |
| 41 | +context_params.n_ctx = n_kv_req |
| 42 | +context_params.n_batch = UInt32(max(n_len, n_parallel)) |
| 43 | +context_params.n_threads = 8 |
| 44 | +context_params.n_threads_batch = 8 |
| 45 | + |
| 46 | +let context = llama_new_context_with_model(model, context_params) |
| 47 | +guard context != nil else { |
| 48 | + print("Failed to initialize context") |
| 49 | + exit(1) |
| 50 | +} |
| 51 | + |
| 52 | +defer { |
| 53 | + llama_free(context) |
| 54 | +} |
| 55 | + |
| 56 | +let n_ctx = llama_n_ctx(context) |
| 57 | + |
| 58 | +print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n") |
| 59 | + |
| 60 | +if n_kv_req > n_ctx { |
| 61 | + print("error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", n_kv_req) |
| 62 | + exit(1) |
| 63 | +} |
| 64 | + |
| 65 | +var buffer: [CChar] = [] |
| 66 | +for id: llama_token in tokens { |
| 67 | + print(token_to_piece(token: id, buffer: &buffer) ?? "", terminator: "") |
| 68 | +} |
| 69 | + |
| 70 | +print("\n") |
| 71 | + |
| 72 | +var batch = llama_batch_init(max(Int32(tokens.count), Int32(n_parallel)), 0) |
| 73 | +defer { |
| 74 | + llama_batch_free(batch) |
| 75 | +} |
| 76 | + |
| 77 | +// evaluate the initial prompt |
| 78 | +batch.n_tokens = Int32(tokens.count) |
| 79 | + |
| 80 | +for (i, token) in tokens.enumerated() { |
| 81 | + batch.token[i] = token |
| 82 | + batch.pos[i] = Int32(i) |
| 83 | + batch.seq_id[i] = 0 |
| 84 | + batch.logits[i] = 0 |
| 85 | +} |
| 86 | + |
| 87 | +// llama_decode will output logits only for the last token of the prompt |
| 88 | +batch.logits[Int(batch.n_tokens) - 1] = 1 |
| 89 | + |
| 90 | +if llama_decode(context, batch) != 0 { |
| 91 | + print("llama_decode() failed") |
| 92 | + exit(1) |
| 93 | +} |
| 94 | + |
| 95 | +for i in 1 ..< n_parallel { |
| 96 | + llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens) |
| 97 | +} |
| 98 | + |
| 99 | +if n_parallel > 1 { |
| 100 | + print("generating \(n_parallel) sequences ...\n") |
| 101 | +} |
| 102 | + |
| 103 | +var streams: [String] = .init(repeating: "", count: n_parallel) |
| 104 | +var streamBuffers: [[CChar]] = .init(repeating: [], count: n_parallel) |
| 105 | +var i_batch = [Int32](repeating: batch.n_tokens - 1, count: n_parallel) |
| 106 | + |
| 107 | +var n_cur = batch.n_tokens |
| 108 | +var n_decode = 0 |
| 109 | + |
| 110 | +let t_main_start = ggml_time_us() |
| 111 | + |
| 112 | +while n_cur <= n_len { |
| 113 | + // prepare the next batch |
| 114 | + batch.n_tokens = 0 |
| 115 | + |
| 116 | + // sample the next token for each parallel sequence / stream |
| 117 | + for i in 0 ..< n_parallel { |
| 118 | + if i_batch[i] < 0 { |
| 119 | + // the stream has already finished |
| 120 | + continue |
| 121 | + } |
| 122 | + |
| 123 | + var n_vocab = llama_n_vocab(model) |
| 124 | + var logits = llama_get_logits_ith(context, i_batch[i]) |
| 125 | + |
| 126 | + var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab)) |
| 127 | + |
| 128 | + for token_id in 0 ..< n_vocab { |
| 129 | + candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0)) |
| 130 | + } |
| 131 | + |
| 132 | + var candidates_p: llama_token_data_array = .init( |
| 133 | + data: &candidates, |
| 134 | + size: candidates.count, |
| 135 | + sorted: false |
| 136 | + ) |
| 137 | + |
| 138 | + let top_k: Int32 = 40 |
| 139 | + let top_p: Float = 0.9 |
| 140 | + let temp: Float = 0.4 |
| 141 | + |
| 142 | + llama_sample_top_k(context, &candidates_p, top_k, 1) |
| 143 | + llama_sample_top_p(context, &candidates_p, top_p, 1) |
| 144 | + llama_sample_temp(context, &candidates_p, temp) |
| 145 | + |
| 146 | + let new_token_id = llama_sample_token(context, &candidates_p) |
| 147 | + |
| 148 | + // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p); |
| 149 | + |
| 150 | + // is it an end of stream? -> mark the stream as finished |
| 151 | + if new_token_id == llama_token_eos(context) || n_cur == n_len { |
| 152 | + i_batch[i] = -1 |
| 153 | + // print("") |
| 154 | + if n_parallel > 1 { |
| 155 | + print("stream \(i) finished at n_cur = \(n_cur)") |
| 156 | + } |
| 157 | + |
| 158 | + continue |
| 159 | + } |
| 160 | + |
| 161 | + let nextStringPiece = token_to_piece(token: new_token_id, buffer: &streamBuffers[i]) ?? "" |
| 162 | + |
| 163 | + // if there is only one stream, we print immediately to stdout |
| 164 | + if n_parallel == 1 { |
| 165 | + print(nextStringPiece, terminator: "") |
| 166 | + } |
| 167 | + streams[i] += nextStringPiece |
| 168 | + |
| 169 | + // push this new token for next evaluation |
| 170 | + batch.token[Int(batch.n_tokens)] = new_token_id |
| 171 | + batch.pos[Int(batch.n_tokens)] = n_cur |
| 172 | + batch.seq_id[Int(batch.n_tokens)] = Int32(i) |
| 173 | + batch.logits[Int(batch.n_tokens)] = 1 |
| 174 | + |
| 175 | + i_batch[i] = batch.n_tokens |
| 176 | + |
| 177 | + batch.n_tokens += 1 |
| 178 | + |
| 179 | + n_decode += 1 |
| 180 | + } |
| 181 | + |
| 182 | + // all streams are finished |
| 183 | + if batch.n_tokens == 0 { |
| 184 | + break |
| 185 | + } |
| 186 | + |
| 187 | + n_cur += 1 |
| 188 | + |
| 189 | + // evaluate the current batch with the transformer model |
| 190 | + if llama_decode(context, batch) != 0 { |
| 191 | + print("llama_decode() failed") |
| 192 | + exit(1) |
| 193 | + } |
| 194 | +} |
| 195 | + |
| 196 | +if n_parallel > 1 { |
| 197 | + print("\n") |
| 198 | + for (i, stream) in streams.enumerated() { |
| 199 | + print("sequence \(i):\n\n\(prompt)\(stream)\n") |
| 200 | + } |
| 201 | +} |
| 202 | + |
| 203 | +let t_main_end = ggml_time_us() |
| 204 | + |
| 205 | +print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n") |
| 206 | + |
| 207 | +llama_print_timings(context) |
| 208 | + |
| 209 | +private func tokenize(text: String, add_bos: Bool) -> [llama_token] { |
| 210 | + let n_tokens = text.count + (add_bos ? 1 : 0) |
| 211 | + let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens) |
| 212 | + let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos) |
| 213 | + var swiftTokens: [llama_token] = [] |
| 214 | + for i in 0 ..< tokenCount { |
| 215 | + swiftTokens.append(tokens[Int(i)]) |
| 216 | + } |
| 217 | + tokens.deallocate() |
| 218 | + return swiftTokens |
| 219 | +} |
| 220 | + |
| 221 | +private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? { |
| 222 | + var result = [CChar](repeating: 0, count: 8) |
| 223 | + let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count)) |
| 224 | + if nTokens < 0 { |
| 225 | + if result.count >= -Int(nTokens) { |
| 226 | + result.removeLast(-Int(nTokens)) |
| 227 | + } else { |
| 228 | + result.removeAll() |
| 229 | + } |
| 230 | + let check = llama_token_to_piece( |
| 231 | + model, |
| 232 | + token, |
| 233 | + &result, |
| 234 | + Int32(result.count) |
| 235 | + ) |
| 236 | + assert(check == nTokens) |
| 237 | + } else { |
| 238 | + result.removeLast(result.count - Int(nTokens)) |
| 239 | + } |
| 240 | + if buffer.isEmpty, let utfString = String(cString: result + [0], encoding: .utf8) { |
| 241 | + return utfString |
| 242 | + } else { |
| 243 | + buffer.append(contentsOf: result) |
| 244 | + let data = Data(buffer.map { UInt8(bitPattern: $0) }) |
| 245 | + if buffer.count >= 4 { // 4 bytes is the max length of a utf8 character so if we're here we need to reset the buffer |
| 246 | + buffer = [] |
| 247 | + } |
| 248 | + guard let bufferString = String(data: data, encoding: .utf8) else { |
| 249 | + return nil |
| 250 | + } |
| 251 | + buffer = [] |
| 252 | + return bufferString |
| 253 | + } |
| 254 | + return nil |
| 255 | +} |
0 commit comments