|
| 1 | +import dataclasses |
| 2 | +from typing import TYPE_CHECKING, Generator, Iterator, List, Optional, Tuple, Union |
| 3 | + |
| 4 | +from .transformers import TransformerTokenizer |
| 5 | + |
| 6 | +if TYPE_CHECKING: |
| 7 | + import mlx.core as mx |
| 8 | + import mlx.nn as nn |
| 9 | + from transformers import PreTrainedTokenizer |
| 10 | + |
| 11 | + from outlines.generate.api import GenerationParameters, SamplingParameters |
| 12 | + from outlines.processors import BaseLogitsProcessor |
| 13 | + |
| 14 | + |
| 15 | +class MLXLM: |
| 16 | + """ |
| 17 | + Represents an `mlx_lm` model |
| 18 | + """ |
| 19 | + |
| 20 | + def __init__( |
| 21 | + self, |
| 22 | + model: "nn.Module", |
| 23 | + tokenizer: "PreTrainedTokenizer", |
| 24 | + ): |
| 25 | + self.model = model |
| 26 | + self.mlx_tokenizer = tokenizer # returns mlx tensors, used for encode() |
| 27 | + self.tokenizer = TransformerTokenizer( |
| 28 | + tokenizer._tokenizer |
| 29 | + ) # _tokenizer is HF Tokenizer |
| 30 | + |
| 31 | + def generate( |
| 32 | + self, |
| 33 | + prompts: Union[str, List[str]], |
| 34 | + generation_parameters: "GenerationParameters", |
| 35 | + logits_processor, |
| 36 | + sampling_parameters: "SamplingParameters", |
| 37 | + ) -> str: |
| 38 | + streamer = self.stream( |
| 39 | + prompts, generation_parameters, logits_processor, sampling_parameters |
| 40 | + ) |
| 41 | + return "".join(list(streamer)) |
| 42 | + |
| 43 | + def stream( |
| 44 | + self, |
| 45 | + prompts: Union[str, List[str]], |
| 46 | + generation_parameters: "GenerationParameters", |
| 47 | + logits_processor, |
| 48 | + sampling_parameters: "SamplingParameters", |
| 49 | + ) -> Iterator[str]: |
| 50 | + """Generate text using `mlx_lm`. |
| 51 | +
|
| 52 | + Arguments |
| 53 | + --------- |
| 54 | + prompts |
| 55 | + A prompt or list of prompts. |
| 56 | + generation_parameters |
| 57 | + An instance of `GenerationParameters` that contains the prompt, |
| 58 | + the maximum number of tokens, stop sequences and seed. All the |
| 59 | + arguments to `SequenceGeneratorAdapter`'s `__cal__` method. |
| 60 | + logits_processor |
| 61 | + The logits processor to use when generating text. |
| 62 | + sampling_parameters |
| 63 | + An instance of `SamplingParameters`, a dataclass that contains |
| 64 | + the name of the sampler to use and related parameters as available |
| 65 | + in Outlines. |
| 66 | + Returns |
| 67 | + ------- |
| 68 | + The generated text. |
| 69 | + """ |
| 70 | + import mlx.core as mx |
| 71 | + |
| 72 | + max_tokens, stop_at, seed = dataclasses.astuple(generation_parameters) |
| 73 | + sampler, num_samples, top_p, top_k, temperature = dataclasses.astuple( |
| 74 | + sampling_parameters |
| 75 | + ) |
| 76 | + if max_tokens is None: |
| 77 | + max_tokens = int(1e9) |
| 78 | + |
| 79 | + if not isinstance(prompts, str): |
| 80 | + raise NotImplementedError( |
| 81 | + "The `mlx-lm` library does not support batch inference." |
| 82 | + ) |
| 83 | + if sampler == "beam_search": |
| 84 | + raise NotImplementedError( |
| 85 | + "The `mlx-lm` library does not support Beam Search." |
| 86 | + ) |
| 87 | + if num_samples != 1: |
| 88 | + raise NotImplementedError( |
| 89 | + "The `mlx-lm` library does not allow to take several samples." |
| 90 | + ) |
| 91 | + if top_k is not None: |
| 92 | + raise NotImplementedError("The `mlx-lm` library does not support top_k.") |
| 93 | + if seed is not None: |
| 94 | + raise NotImplementedError("The `mlx-lm` library does not support seed.") |
| 95 | + if stop_at is not None: |
| 96 | + raise NotImplementedError("The `mlx-lm` library does not support stop_at.") |
| 97 | + |
| 98 | + generate_kwargs = { |
| 99 | + "temp": temperature, |
| 100 | + "top_p": top_p, |
| 101 | + "sampler": sampler, |
| 102 | + "logits_processor": logits_processor, |
| 103 | + } |
| 104 | + |
| 105 | + # Adapted from |
| 106 | + # https://github.com/ml-explore/mlx-examples/blob/4872727/llms/mlx_lm/utils.py#L267 |
| 107 | + prompt_tokens = mx.array(self.mlx_tokenizer.encode(prompts)) |
| 108 | + |
| 109 | + for (token, prob), n in zip( |
| 110 | + self.generate_step(prompt_tokens, **generate_kwargs), |
| 111 | + range(max_tokens), |
| 112 | + ): |
| 113 | + if token == self.tokenizer.eos_token_id: |
| 114 | + break |
| 115 | + yield self.tokenizer.decode([token])[0] |
| 116 | + |
| 117 | + def generate_step( |
| 118 | + self, |
| 119 | + prompt: "mx.array", |
| 120 | + temp: Optional[float], |
| 121 | + top_p: Optional[float], |
| 122 | + sampler: str, |
| 123 | + logits_processor: "BaseLogitsProcessor", |
| 124 | + ) -> Generator[Tuple[int, float], None, None]: |
| 125 | + """ |
| 126 | + Adapted from |
| 127 | + https://github.com/ml-explore/mlx-examples/blob/4872727/llms/mlx_lm/utils.py#L129 |
| 128 | +
|
| 129 | + A generator producing token ids based on the given prompt from the model. |
| 130 | +
|
| 131 | + Args: |
| 132 | + prompt (mx.array): The input prompt. |
| 133 | + temp (float): The temperature for sampling, if 0 the argmax is used. |
| 134 | + Default: ``0``. |
| 135 | + top_p (float, optional): Nulceus sampling, higher means model considers |
| 136 | + more less likely words. |
| 137 | + sampler (str): The sampler string defined by SequenceGeneratorAdapter |
| 138 | + logits_processor (BaseLogitsProcessor): Augment logits before sampling. |
| 139 | + """ |
| 140 | + import mlx.core as mx |
| 141 | + import mlx_lm |
| 142 | + |
| 143 | + temperature: float = temp or 1.0 |
| 144 | + |
| 145 | + def sample(logits: "mx.array") -> Tuple["mx.array", float]: |
| 146 | + softmax_logits = mx.softmax(logits) |
| 147 | + |
| 148 | + if temperature == 0.0 or sampler == "greedy": |
| 149 | + token = mx.argmax(logits, axis=-1) |
| 150 | + elif sampler == "multinomial": |
| 151 | + if top_p is not None and top_p > 0 and top_p < 1.0: |
| 152 | + token = mlx_lm.sample_utils.top_p_sampling( |
| 153 | + logits, top_p, temperature |
| 154 | + ) |
| 155 | + else: |
| 156 | + token = mx.random.categorical(logits * (1 / temperature)) |
| 157 | + else: |
| 158 | + raise ValueError(f"Invalid mlx-lm sampler: `{sampler}`") |
| 159 | + |
| 160 | + prob = softmax_logits[0, token] |
| 161 | + return token, prob |
| 162 | + |
| 163 | + kv_heads = ( |
| 164 | + [self.model.n_kv_heads] * len(self.model.layers) |
| 165 | + if isinstance(self.model.n_kv_heads, int) |
| 166 | + else self.model.n_kv_heads |
| 167 | + ) |
| 168 | + cache = [mlx_lm.models.base.KVCache(self.model.head_dim, n) for n in kv_heads] |
| 169 | + |
| 170 | + # kv cache contains processed input IDs, we pass the unprocessed inputs and cache to model() |
| 171 | + unprocessed_input_ids = prompt |
| 172 | + generated_ids: List[int] = [] |
| 173 | + |
| 174 | + while True: |
| 175 | + logits = self.model(unprocessed_input_ids[None], cache=cache) |
| 176 | + logits = logits[:, -1, :] |
| 177 | + |
| 178 | + if logits_processor is not None: |
| 179 | + # convert to logits_processor 1d expectation, apply, then convert back |
| 180 | + logits_1d = logits.reshape(-1) |
| 181 | + logits_1d = logits_processor(generated_ids, logits_1d) |
| 182 | + logits = logits_1d.reshape(1, -1) |
| 183 | + |
| 184 | + new_token_single, prob = sample(logits) |
| 185 | + new_token = new_token_single.item() |
| 186 | + yield new_token, prob |
| 187 | + |
| 188 | + generated_ids.append(new_token) |
| 189 | + unprocessed_input_ids = new_token_single |
| 190 | + |
| 191 | + |
| 192 | +def mlxlm( |
| 193 | + model_name: str, |
| 194 | + tokenizer_config: dict = {}, |
| 195 | + model_config: dict = {}, |
| 196 | + adapter_path: Optional[str] = None, |
| 197 | + lazy: bool = False, |
| 198 | +): |
| 199 | + """Instantiate a model from the `mlx_lm` library and its tokenizer. |
| 200 | +
|
| 201 | + Signature adapted from |
| 202 | + https://github.com/ml-explore/mlx-examples/blob/4872727/llms/mlx_lm/utils.py#L422 |
| 203 | +
|
| 204 | + Parameters |
| 205 | + ---------- |
| 206 | + Args: |
| 207 | + path_or_hf_repo (Path): The path or the huggingface repository to load the model from. |
| 208 | + tokenizer_config (dict, optional): Configuration parameters specifically for the tokenizer. |
| 209 | + Defaults to an empty dictionary. |
| 210 | + model_config(dict, optional): Configuration parameters specifically for the model. |
| 211 | + Defaults to an empty dictionary. |
| 212 | + adapter_path (str, optional): Path to the LoRA adapters. If provided, applies LoRA layers |
| 213 | + to the model. Default: ``None``. |
| 214 | + lazy (bool): If False eval the model parameters to make sure they are |
| 215 | + loaded in memory before returning, otherwise they will be loaded |
| 216 | + when needed. Default: ``False`` |
| 217 | +
|
| 218 | + Returns |
| 219 | + ------- |
| 220 | + A `MLXLM` model instance. |
| 221 | +
|
| 222 | + """ |
| 223 | + try: |
| 224 | + import mlx.core as mx |
| 225 | + import mlx_lm |
| 226 | + except ImportError: |
| 227 | + raise ImportError( |
| 228 | + "The `mlx_lm` library needs to be installed in order to use `mlx_lm` models." |
| 229 | + ) |
| 230 | + if not mx.metal.is_available(): |
| 231 | + raise RuntimeError("You cannot use `mlx_lm` without Apple Silicon (Metal)") |
| 232 | + |
| 233 | + model, tokenizer = mlx_lm.load( |
| 234 | + model_name, |
| 235 | + tokenizer_config=tokenizer_config, |
| 236 | + model_config=model_config, |
| 237 | + adapter_path=adapter_path, |
| 238 | + lazy=lazy, |
| 239 | + ) |
| 240 | + return MLXLM(model, tokenizer) |
0 commit comments