TheBitmonkey
commented
6 months ago Ollama needs to be sent the raw flag set to true.... but here is the catch, Langchain will not send it so this is what I do which seems to work to pass the template.
Modify the Ollama.py file in langchain-community.llms to this:
import json
from typing import Any, AsyncIterator, Dict, Iterator, List, Mapping, Optional, Union
import aiohttp
import requests
from langchain_core.callbacks import (
AsyncCallbackManagerForLLMRun,
CallbackManagerForLLMRun,
)
from langchain_core.language_models import BaseLanguageModel
from langchain_core.language_models.llms import BaseLLM
from langchain_core.outputs import GenerationChunk, LLMResult
from langchain_core.pydantic_v1 import Extra
def _stream_response_to_generation_chunk(
stream_response: str,
) -> GenerationChunk:
"""Convert a stream response to a generation chunk."""
parsed_response = json.loads(stream_response)
generation_info = parsed_response if parsed_response.get("done") is True else None
return GenerationChunk(
text=parsed_response.get("response", ""), generation_info=generation_info
)
class OllamaEndpointNotFoundError(Exception):
"""Raised when the Ollama endpoint is not found."""
class _OllamaCommon(BaseLanguageModel):
base_url: str = "http://localhost:11434"
"""Base url the model is hosted under."""
model: str = "llama2"
"""Model name to use."""
mirostat: Optional[int] = None
"""Enable Mirostat sampling for controlling perplexity.
(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)"""
mirostat_eta: Optional[float] = None
"""Influences how quickly the algorithm responds to feedback
from the generated text. A lower learning rate will result in
slower adjustments, while a higher learning rate will make
the algorithm more responsive. (Default: 0.1)"""
mirostat_tau: Optional[float] = None
"""Controls the balance between coherence and diversity
of the output. A lower value will result in more focused and
coherent text. (Default: 5.0)"""
num_ctx: Optional[int] = None
"""Sets the size of the context window used to generate the
next token. (Default: 2048) """
num_gpu: Optional[int] = None
"""The number of GPUs to use. On macOS it defaults to 1 to
enable metal support, 0 to disable."""
num_thread: Optional[int] = None
"""Sets the number of threads to use during computation.
By default, Ollama will detect this for optimal performance.
It is recommended to set this value to the number of physical
CPU cores your system has (as opposed to the logical number of cores)."""
num_predict: Optional[int] = None
"""Maximum number of tokens to predict when generating text.
(Default: 128, -1 = infinite generation, -2 = fill context)"""
repeat_last_n: Optional[int] = None
"""Sets how far back for the model to look back to prevent
repetition. (Default: 64, 0 = disabled, -1 = num_ctx)"""
repeat_penalty: Optional[float] = None
"""Sets how strongly to penalize repetitions. A higher value (e.g., 1.5)
will penalize repetitions more strongly, while a lower value (e.g., 0.9)
will be more lenient. (Default: 1.1)"""
temperature: Optional[float] = None
"""The temperature of the model. Increasing the temperature will
make the model answer more creatively. (Default: 0.8)"""
stop: Optional[List[str]] = None
"""Sets the stop tokens to use."""
tfs_z: Optional[float] = None
"""Tail free sampling is used to reduce the impact of less probable
tokens from the output. A higher value (e.g., 2.0) will reduce the
impact more, while a value of 1.0 disables this setting. (default: 1)"""
top_k: Optional[int] = None
"""Reduces the probability of generating nonsense. A higher value (e.g. 100)
will give more diverse answers, while a lower value (e.g. 10)
will be more conservative. (Default: 40)"""
top_p: Optional[float] = None
"""Works together with top-k. A higher value (e.g., 0.95) will lead
to more diverse text, while a lower value (e.g., 0.5) will
generate more focused and conservative text. (Default: 0.9)"""
system: Optional[str] = None
"""system prompt (overrides what is defined in the Modelfile)"""
template: Optional[str] = None
"""full prompt or prompt template (overrides what is defined in the Modelfile)"""
format: Optional[str] = None
"""Specify the format of the output (e.g., json)"""
timeout: Optional[int] = None
"""Timeout for the request stream"""
keep_alive: Optional[Union[int, str]] = None
"""How long the model will stay loaded into memory."""
raw: Optional[bool] = None
"""raw or not.""
The parameter (Default: 5 minutes) can be set to:
1. a duration string in Golang (such as "10m" or "24h");
2. a number in seconds (such as 3600);
3. any negative number which will keep the model loaded \
in memory (e.g. -1 or "-1m");
4. 0 which will unload the model immediately after generating a response;
See the [Ollama documents](https://github.com/ollama/ollama/blob/main/docs/faq.md#how-do-i-keep-a-model-loaded-in-memory-or-make-it-unload-immediately)"""
headers: Optional[dict] = None
"""Additional headers to pass to endpoint (e.g. Authorization, Referer).
This is useful when Ollama is hosted on cloud services that require
tokens for authentication.
"""
@property
def _default_params(self) -> Dict[str, Any]:
"""Get the default parameters for calling Ollama."""
return {
"model": self.model,
"format": self.format,
"options": {
"mirostat": self.mirostat,
"mirostat_eta": self.mirostat_eta,
"mirostat_tau": self.mirostat_tau,
"num_ctx": self.num_ctx,
"num_gpu": self.num_gpu,
"num_thread": self.num_thread,
"num_predict": self.num_predict,
"repeat_last_n": self.repeat_last_n,
"repeat_penalty": self.repeat_penalty,
"temperature": self.temperature,
"stop": self.stop,
"tfs_z": self.tfs_z,
"top_k": self.top_k,
"top_p": self.top_p,
},
"system": self.system,
"template": self.template,
"keep_alive": self.keep_alive,
"raw": self.raw,
}
@property
def _identifying_params(self) -> Mapping[str, Any]:
"""Get the identifying parameters."""
return {**{"model": self.model, "format": self.format}, **self._default_params}
def _create_generate_stream(
self,
prompt: str,
stop: Optional[List[str]] = None,
images: Optional[List[str]] = None,
**kwargs: Any,
) -> Iterator[str]:
payload = {"prompt": prompt, "images": images}
yield from self._create_stream(
payload=payload,
stop=stop,
api_url=f"{self.base_url}/api/generate",
**kwargs,
)
async def _acreate_generate_stream(
self,
prompt: str,
stop: Optional[List[str]] = None,
images: Optional[List[str]] = None,
**kwargs: Any,
) -> AsyncIterator[str]:
payload = {"prompt": prompt, "images": images}
async for item in self._acreate_stream(
payload=payload,
stop=stop,
api_url=f"{self.base_url}/api/generate",
**kwargs,
):
yield item
def _create_stream(
self,
api_url: str,
payload: Any,
stop: Optional[List[str]] = None,
**kwargs: Any,
) -> Iterator[str]:
if self.stop is not None:
stop = self.stop
params = self._default_params
for key in self._default_params:
if key in kwargs:
params[key] = kwargs[key]
if "options" in kwargs:
params["options"] = kwargs["options"]
else:
params["options"] = {
**params["options"],
"stop": stop,
**{k: v for k, v in kwargs.items() if k not in self._default_params},
}
if payload.get("messages"):
request_payload = {"messages": payload.get("messages", []), **params}
else:
request_payload = {
"prompt": payload.get("prompt"),
"images": payload.get("images", []),
**params,
}
response = requests.post(
url=api_url,
headers={
"Content-Type": "application/json",
**(self.headers if isinstance(self.headers, dict) else {}),
},
json=request_payload,
stream=True,
timeout=self.timeout,
)
response.encoding = "utf-8"
if response.status_code != 200:
if response.status_code == 404:
raise OllamaEndpointNotFoundError(
"Ollama call failed with status code 404. "
"Maybe your model is not found "
f"and you should pull the model with `ollama pull {self.model}`."
)
else:
optional_detail = response.text
raise ValueError(
f"Ollama call failed with status code {response.status_code}."
f" Details: {optional_detail}"
)
return response.iter_lines(decode_unicode=True)
async def _acreate_stream(
self,
api_url: str,
payload: Any,
stop: Optional[List[str]] = None,
**kwargs: Any,
) -> AsyncIterator[str]:
if self.stop is not None and stop is not None:
raise ValueError("`stop` found in both the input and default params.")
elif self.stop is not None:
stop = self.stop
params = self._default_params
for key in self._default_params:
if key in kwargs:
params[key] = kwargs[key]
if "options" in kwargs:
params["options"] = kwargs["options"]
else:
params["options"] = {
**params["options"],
"stop": stop,
**{k: v for k, v in kwargs.items() if k not in self._default_params},
}
if payload.get("messages"):
request_payload = {"messages": payload.get("messages", []), **params}
else:
request_payload = {
"prompt": payload.get("prompt"),
"images": payload.get("images", []),
**params,
}
async with aiohttp.ClientSession() as session:
async with session.post(
url=api_url,
headers={
"Content-Type": "application/json",
**(self.headers if isinstance(self.headers, dict) else {}),
},
json=request_payload,
timeout=self.timeout,
) as response:
if response.status != 200:
if response.status == 404:
raise OllamaEndpointNotFoundError(
"Ollama call failed with status code 404."
)
else:
optional_detail = response.text
raise ValueError(
f"Ollama call failed with status code {response.status}."
f" Details: {optional_detail}"
)
async for line in response.content:
yield line.decode("utf-8")
def _stream_with_aggregation(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
verbose: bool = False,
**kwargs: Any,
) -> GenerationChunk:
final_chunk: Optional[GenerationChunk] = None
for stream_resp in self._create_generate_stream(prompt, stop, **kwargs):
if stream_resp:
chunk = _stream_response_to_generation_chunk(stream_resp)
if final_chunk is None:
final_chunk = chunk
else:
final_chunk += chunk
if run_manager:
run_manager.on_llm_new_token(
chunk.text,
verbose=verbose,
)
if final_chunk is None:
raise ValueError("No data received from Ollama stream.")
return final_chunk
async def _astream_with_aggregation(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
verbose: bool = False,
**kwargs: Any,
) -> GenerationChunk:
final_chunk: Optional[GenerationChunk] = None
async for stream_resp in self._acreate_generate_stream(prompt, stop, **kwargs):
if stream_resp:
chunk = _stream_response_to_generation_chunk(stream_resp)
if final_chunk is None:
final_chunk = chunk
else:
final_chunk += chunk
if run_manager:
await run_manager.on_llm_new_token(
chunk.text,
verbose=verbose,
)
if final_chunk is None:
raise ValueError("No data received from Ollama stream.")
return final_chunk
class Ollama(BaseLLM, _OllamaCommon):
"""Ollama locally runs large language models.
To use, follow the instructions at https://ollama.ai/.
Example:
.. code-block:: python
from langchain_community.llms import Ollama
ollama = Ollama(model="llama2")
"""
class Config:
"""Configuration for this pydantic object."""
extra = Extra.forbid
@property
def _llm_type(self) -> str:
"""Return type of llm."""
return "ollama-llm"
def _generate( # type: ignore[override]
self,
prompts: List[str],
stop: Optional[List[str]] = None,
images: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> LLMResult:
"""Call out to Ollama's generate endpoint.
Args:
prompt: The prompt to pass into the model.
stop: Optional list of stop words to use when generating.
Returns:
The string generated by the model.
Example:
.. code-block:: python
response = ollama("Tell me a joke.")
"""
# TODO: add caching here.
generations = []
for prompt in prompts:
final_chunk = super()._stream_with_aggregation(
prompt,
stop=stop,
images=images,
run_manager=run_manager,
verbose=self.verbose,
**kwargs,
)
generations.append([final_chunk])
return LLMResult(generations=generations)
async def _agenerate( # type: ignore[override]
self,
prompts: List[str],
stop: Optional[List[str]] = None,
images: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> LLMResult:
"""Call out to Ollama's generate endpoint.
Args:
prompt: The prompt to pass into the model.
stop: Optional list of stop words to use when generating.
Returns:
The string generated by the model.
Example:
.. code-block:: python
response = ollama("Tell me a joke.")
"""
# TODO: add caching here.
generations = []
for prompt in prompts:
final_chunk = await super()._astream_with_aggregation(
prompt,
stop=stop,
images=images,
run_manager=run_manager, # type: ignore[arg-type]
verbose=self.verbose,
**kwargs,
)
generations.append([final_chunk])
return LLMResult(generations=generations)
def _stream(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> Iterator[GenerationChunk]:
for stream_resp in self._create_generate_stream(prompt, stop, **kwargs):
if stream_resp:
chunk = _stream_response_to_generation_chunk(stream_resp)
if run_manager:
run_manager.on_llm_new_token(
chunk.text,
verbose=self.verbose,
)
yield chunk
async def _astream(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[AsyncCallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> AsyncIterator[GenerationChunk]:
async for stream_resp in self._acreate_generate_stream(prompt, stop, **kwargs):
if stream_resp:
chunk = _stream_response_to_generation_chunk(stream_resp)
if run_manager:
await run_manager.on_llm_new_token(
chunk.text,
verbose=self.verbose,
)
yield chunkThen call it like so:
llm = Ollama(model=models["h2p-ll3"]["model"], stop=models["h2p-ll3"]["stop"], base_url=models["h2p-ll3"]["base_url"], num_ctx=models["h2p-ll3"]["num_ctx"], raw=True)
I am importing the model details from a json file so for reference here is that:
"h2p-ll3": { "model":"adrienbrault/nous-hermes2pro-llama3-8b:q8_0", "system_template": "<|im_start|>system\n{{ .System }}<|im_end|>", "prompt_template": "<|im_start|>user\n{{ .Prompt }}<|im_end|><|im_start|>assistant\n", "response_template": "", "stop": ["<|im_start|>", "<|start_header_id|>", "<|end_header_id|>", "<|eot_id|>", "<|reserved_special_token", "assistant", "<|im_end|>"], "num_ctx":8192, "base_url":"http://192.168.1.134:11434" }
github-actions[bot]
I'm trying to make Phi3 (3.8B) work. The model seems to have no problems using tools, but uses (different?) system messages, which leads to the chain not ending when it is supposed to:
this is the modelfile
Using
crewai 0.30.8I copied system/prompt/responese template is from #554, but it seems that it either not used, or just doesn't work with phi3.