Spaces:
Runtime error
Runtime error
File size: 6,166 Bytes
ed4d993 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 |
"""MLX Chat Wrapper."""
from typing import Any, Iterator, List, Optional
from langchain_core.callbacks.manager import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models.chat_models import BaseChatModel
from langchain_core.messages import (
AIMessage,
AIMessageChunk,
BaseMessage,
HumanMessage,
SystemMessage,
)
from langchain_core.outputs import (
ChatGeneration,
ChatGenerationChunk,
ChatResult,
LLMResult,
)
from langchain_community.llms.mlx_pipeline import MLXPipeline
DEFAULT_SYSTEM_PROMPT = """You are a helpful, respectful, and honest assistant."""
class ChatMLX(BaseChatModel):
"""MLX chat models.
Works with `MLXPipeline` LLM.
To use, you should have the ``mlx-lm`` python package installed.
Example:
.. code-block:: python
from langchain_community.chat_models import chatMLX
from langchain_community.llms import MLXPipeline
llm = MLXPipeline.from_model_id(
model_id="mlx-community/quantized-gemma-2b-it",
)
chat = chatMLX(llm=llm)
"""
llm: MLXPipeline
system_message: SystemMessage = SystemMessage(content=DEFAULT_SYSTEM_PROMPT)
tokenizer: Any = None
def __init__(self, **kwargs: Any):
super().__init__(**kwargs)
self.tokenizer = self.llm.tokenizer
def _generate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> ChatResult:
llm_input = self._to_chat_prompt(messages)
llm_result = self.llm._generate(
prompts=[llm_input], stop=stop, run_manager=run_manager, **kwargs
)
return self._to_chat_result(llm_result)
async def _agenerate(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> ChatResult:
llm_input = self._to_chat_prompt(messages)
llm_result = await self.llm._agenerate(
prompts=[llm_input], stop=stop, run_manager=run_manager, **kwargs
)
return self._to_chat_result(llm_result)
def _to_chat_prompt(
self,
messages: List[BaseMessage],
tokenize: bool = False,
return_tensors: Optional[str] = None,
) -> str:
"""Convert a list of messages into a prompt format expected by wrapped LLM."""
if not messages:
raise ValueError("At least one HumanMessage must be provided!")
if not isinstance(messages[-1], HumanMessage):
raise ValueError("Last message must be a HumanMessage!")
messages_dicts = [self._to_chatml_format(m) for m in messages]
return self.tokenizer.apply_chat_template(
messages_dicts,
tokenize=tokenize,
add_generation_prompt=True,
return_tensors=return_tensors,
)
def _to_chatml_format(self, message: BaseMessage) -> dict:
"""Convert LangChain message to ChatML format."""
if isinstance(message, SystemMessage):
role = "system"
elif isinstance(message, AIMessage):
role = "assistant"
elif isinstance(message, HumanMessage):
role = "user"
else:
raise ValueError(f"Unknown message type: {type(message)}")
return {"role": role, "content": message.content}
@staticmethod
def _to_chat_result(llm_result: LLMResult) -> ChatResult:
chat_generations = []
for g in llm_result.generations[0]:
chat_generation = ChatGeneration(
message=AIMessage(content=g.text), generation_info=g.generation_info
)
chat_generations.append(chat_generation)
return ChatResult(
generations=chat_generations, llm_output=llm_result.llm_output
)
@property
def _llm_type(self) -> str:
return "mlx-chat-wrapper"
def _stream(
self,
messages: List[BaseMessage],
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[ChatGenerationChunk]:
import mlx.core as mx
from mlx_lm.utils import generate_step
try:
import mlx.core as mx
from mlx_lm.utils import generate_step
except ImportError:
raise ImportError(
"Could not import mlx_lm python package. "
"Please install it with `pip install mlx_lm`."
)
model_kwargs = kwargs.get("model_kwargs", self.llm.pipeline_kwargs)
temp: float = model_kwargs.get("temp", 0.0)
max_new_tokens: int = model_kwargs.get("max_tokens", 100)
repetition_penalty: Optional[float] = model_kwargs.get(
"repetition_penalty", None
)
repetition_context_size: Optional[int] = model_kwargs.get(
"repetition_context_size", None
)
llm_input = self._to_chat_prompt(messages, tokenize=True, return_tensors="np")
prompt_tokens = mx.array(llm_input[0])
eos_token_id = self.tokenizer.eos_token_id
for (token, prob), n in zip(
generate_step(
prompt_tokens,
self.llm.model,
temp,
repetition_penalty,
repetition_context_size,
),
range(max_new_tokens),
):
# identify text to yield
text: Optional[str] = None
text = self.tokenizer.decode(token.item())
# yield text, if any
if text:
chunk = ChatGenerationChunk(message=AIMessageChunk(content=text))
yield chunk
if run_manager:
run_manager.on_llm_new_token(text, chunk=chunk)
# break if stop sequence found
if token == eos_token_id or (stop is not None and text in stop):
break
|