Spaces:
Runtime error
Runtime error
Roger Condori
commited on
upload backend
Browse files- conversadocs/bones.py +78 -39
- conversadocs/llamacppmodels.py +309 -0
- conversadocs/llm_chess.py +101 -0
conversadocs/bones.py
CHANGED
@@ -7,7 +7,7 @@ from langchain.memory import ConversationBufferMemory
|
|
7 |
from langchain.chat_models import ChatOpenAI
|
8 |
from langchain.embeddings import HuggingFaceEmbeddings
|
9 |
from langchain import HuggingFaceHub
|
10 |
-
from langchain.llms import LlamaCpp
|
11 |
from huggingface_hub import hf_hub_download
|
12 |
import param
|
13 |
import os
|
@@ -23,37 +23,28 @@ from langchain.document_loaders import (
|
|
23 |
UnstructuredWordDocumentLoader,
|
24 |
PyPDFLoader,
|
25 |
)
|
|
|
|
|
|
|
26 |
|
27 |
#YOUR_HF_TOKEN = os.getenv("My_hf_token")
|
28 |
-
llm_api=HuggingFaceHub(
|
29 |
-
huggingfacehub_api_token=os.getenv("My_hf_token"),
|
30 |
-
repo_id="tiiuae/falcon-7b-instruct",
|
31 |
-
model_kwargs={
|
32 |
-
"temperature":0.2,
|
33 |
-
"max_new_tokens":500,
|
34 |
-
"top_k":50,
|
35 |
-
"top_p":0.95,
|
36 |
-
"repetition_penalty":1.2,
|
37 |
-
},), #ChatOpenAI(model_name=llm_name, temperature=0)
|
38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
40 |
#alter
|
41 |
def load_db(files):
|
42 |
-
EXTENSIONS = {
|
43 |
-
".txt": (TextLoader, {"encoding": "utf8"}),
|
44 |
-
".pdf": (PyPDFLoader, {}),
|
45 |
-
".doc": (UnstructuredWordDocumentLoader, {}),
|
46 |
-
".docx": (UnstructuredWordDocumentLoader, {}),
|
47 |
-
".enex": (EverNoteLoader, {}),
|
48 |
-
".epub": (UnstructuredEPubLoader, {}),
|
49 |
-
".html": (UnstructuredHTMLLoader, {}),
|
50 |
-
".md": (UnstructuredMarkdownLoader, {}),
|
51 |
-
".odt": (UnstructuredODTLoader, {}),
|
52 |
-
".ppt": (UnstructuredPowerPointLoader, {}),
|
53 |
-
".pptx": (UnstructuredPowerPointLoader, {}),
|
54 |
-
}
|
55 |
-
|
56 |
-
|
57 |
|
58 |
# select extensions loader
|
59 |
documents = []
|
@@ -102,14 +93,14 @@ class DocChat(param.Parameterized):
|
|
102 |
answer = param.String("")
|
103 |
db_query = param.String("")
|
104 |
db_response = param.List([])
|
105 |
-
llm = llm_api[0]
|
106 |
k_value = param.Integer(3)
|
107 |
-
|
108 |
|
109 |
def __init__(self, **params):
|
110 |
super(DocChat, self).__init__( **params)
|
111 |
-
self.loaded_file = "demo_docs/demo.txt"
|
112 |
self.db = load_db(self.loaded_file)
|
|
|
113 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
114 |
|
115 |
|
@@ -141,7 +132,8 @@ class DocChat(param.Parameterized):
|
|
141 |
try:
|
142 |
result = self.qa({"question": query, "chat_history": self.chat_history})
|
143 |
except:
|
144 |
-
|
|
|
145 |
self.qa = q_a(self.db, "stuff", k_max, self.llm)
|
146 |
result = self.qa({"question": query, "chat_history": self.chat_history})
|
147 |
|
@@ -151,17 +143,48 @@ class DocChat(param.Parameterized):
|
|
151 |
self.answer = result['answer']
|
152 |
return self.answer
|
153 |
|
154 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
|
156 |
if torch.cuda.is_available():
|
157 |
try:
|
158 |
model_path = hf_hub_download(repo_id=repo_, filename=file_)
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
|
160 |
self.llm = LlamaCpp(
|
161 |
model_path=model_path,
|
162 |
-
n_ctx=
|
163 |
n_batch=512,
|
164 |
-
n_gpu_layers=
|
165 |
max_tokens=max_tokens,
|
166 |
verbose=False,
|
167 |
temperature=temperature,
|
@@ -173,14 +196,20 @@ class DocChat(param.Parameterized):
|
|
173 |
self.k_value = k
|
174 |
return f"Loaded {file_} [GPU INFERENCE]"
|
175 |
except:
|
176 |
-
|
|
|
177 |
else:
|
178 |
try:
|
179 |
model_path = hf_hub_download(repo_id=repo_, filename=file_)
|
|
|
|
|
|
|
|
|
|
|
180 |
|
181 |
self.llm = LlamaCpp(
|
182 |
model_path=model_path,
|
183 |
-
n_ctx=
|
184 |
n_batch=8,
|
185 |
max_tokens=max_tokens,
|
186 |
verbose=False,
|
@@ -193,10 +222,20 @@ class DocChat(param.Parameterized):
|
|
193 |
self.k_value = k
|
194 |
return f"Loaded {file_} [CPU INFERENCE SLOW]"
|
195 |
except:
|
196 |
-
|
|
|
197 |
|
198 |
-
def default_falcon_model(self):
|
199 |
-
self.llm = llm_api
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
200 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
201 |
return "Loaded model Falcon 7B-instruct [API FAST INFERENCE]"
|
202 |
|
@@ -204,7 +243,7 @@ class DocChat(param.Parameterized):
|
|
204 |
self.llm = ChatOpenAI(temperature=0, openai_api_key=API_KEY, model_name='gpt-3.5-turbo')
|
205 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
206 |
API_KEY = ""
|
207 |
-
return "Loaded model OpenAI gpt-3.5-turbo [API FAST INFERENCE]
|
208 |
|
209 |
@param.depends('db_query ', )
|
210 |
def get_lquest(self):
|
|
|
7 |
from langchain.chat_models import ChatOpenAI
|
8 |
from langchain.embeddings import HuggingFaceEmbeddings
|
9 |
from langchain import HuggingFaceHub
|
10 |
+
from conversadocs.llamacppmodels import LlamaCpp #from langchain.llms import LlamaCpp
|
11 |
from huggingface_hub import hf_hub_download
|
12 |
import param
|
13 |
import os
|
|
|
23 |
UnstructuredWordDocumentLoader,
|
24 |
PyPDFLoader,
|
25 |
)
|
26 |
+
import gc
|
27 |
+
gc.collect()
|
28 |
+
torch.cuda.empty_cache()
|
29 |
|
30 |
#YOUR_HF_TOKEN = os.getenv("My_hf_token")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
|
32 |
+
EXTENSIONS = {
|
33 |
+
".txt": (TextLoader, {"encoding": "utf8"}),
|
34 |
+
".pdf": (PyPDFLoader, {}),
|
35 |
+
".doc": (UnstructuredWordDocumentLoader, {}),
|
36 |
+
".docx": (UnstructuredWordDocumentLoader, {}),
|
37 |
+
".enex": (EverNoteLoader, {}),
|
38 |
+
".epub": (UnstructuredEPubLoader, {}),
|
39 |
+
".html": (UnstructuredHTMLLoader, {}),
|
40 |
+
".md": (UnstructuredMarkdownLoader, {}),
|
41 |
+
".odt": (UnstructuredODTLoader, {}),
|
42 |
+
".ppt": (UnstructuredPowerPointLoader, {}),
|
43 |
+
".pptx": (UnstructuredPowerPointLoader, {}),
|
44 |
+
}
|
45 |
|
46 |
#alter
|
47 |
def load_db(files):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
# select extensions loader
|
50 |
documents = []
|
|
|
93 |
answer = param.String("")
|
94 |
db_query = param.String("")
|
95 |
db_response = param.List([])
|
|
|
96 |
k_value = param.Integer(3)
|
97 |
+
llm = None
|
98 |
|
99 |
def __init__(self, **params):
|
100 |
super(DocChat, self).__init__( **params)
|
101 |
+
self.loaded_file = ["demo_docs/demo.txt"]
|
102 |
self.db = load_db(self.loaded_file)
|
103 |
+
self.change_llm("TheBloke/Llama-2-7B-Chat-GGML", "llama-2-7b-chat.ggmlv3.q5_1.bin", max_tokens=256, temperature=0.2, top_p=0.95, top_k=50, repeat_penalty=1.2, k=3)
|
104 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
105 |
|
106 |
|
|
|
132 |
try:
|
133 |
result = self.qa({"question": query, "chat_history": self.chat_history})
|
134 |
except:
|
135 |
+
print("Error not get response from model, reloaded default llama-2 7B config")
|
136 |
+
self.change_llm("TheBloke/Llama-2-7B-Chat-GGML", "llama-2-7b-chat.ggmlv3.q5_1.bin", max_tokens=256, temperature=0.2, top_p=0.95, top_k=50, repeat_penalty=1.2, k=3)
|
137 |
self.qa = q_a(self.db, "stuff", k_max, self.llm)
|
138 |
result = self.qa({"question": query, "chat_history": self.chat_history})
|
139 |
|
|
|
143 |
self.answer = result['answer']
|
144 |
return self.answer
|
145 |
|
146 |
+
def summarize(self, chunk_size=2000, chunk_overlap=100):
|
147 |
+
# load docs
|
148 |
+
documents = []
|
149 |
+
for file in self.loaded_file:
|
150 |
+
ext = "." + file.rsplit(".", 1)[-1]
|
151 |
+
if ext in EXTENSIONS:
|
152 |
+
loader_class, loader_args = EXTENSIONS[ext]
|
153 |
+
loader = loader_class(file, **loader_args)
|
154 |
+
documents.extend(loader.load_and_split())
|
155 |
+
|
156 |
+
if documents == []:
|
157 |
+
return "Error in summarization"
|
158 |
+
|
159 |
+
# split documents
|
160 |
+
text_splitter = RecursiveCharacterTextSplitter(
|
161 |
+
chunk_size=chunk_size,
|
162 |
+
chunk_overlap=chunk_overlap,
|
163 |
+
separators=["\n\n", "\n", "(?<=\. )", " ", ""]
|
164 |
+
)
|
165 |
+
docs = text_splitter.split_documents(documents)
|
166 |
+
# summarize
|
167 |
+
from langchain.chains.summarize import load_summarize_chain
|
168 |
+
chain = load_summarize_chain(self.llm, chain_type='map_reduce', verbose=True)
|
169 |
+
return chain.run(docs)
|
170 |
+
|
171 |
+
def change_llm(self, repo_, file_, max_tokens=256, temperature=0.2, top_p=0.95, top_k=50, repeat_penalty=1.2, k=3):
|
172 |
|
173 |
if torch.cuda.is_available():
|
174 |
try:
|
175 |
model_path = hf_hub_download(repo_id=repo_, filename=file_)
|
176 |
+
|
177 |
+
self.qa = None
|
178 |
+
self.llm = None
|
179 |
+
gc.collect()
|
180 |
+
torch.cuda.empty_cache()
|
181 |
+
gpu_llm_layers = 35 if not '70B' in repo_.upper() else 25 # fix for 70B
|
182 |
|
183 |
self.llm = LlamaCpp(
|
184 |
model_path=model_path,
|
185 |
+
n_ctx=4096,
|
186 |
n_batch=512,
|
187 |
+
n_gpu_layers=gpu_llm_layers,
|
188 |
max_tokens=max_tokens,
|
189 |
verbose=False,
|
190 |
temperature=temperature,
|
|
|
196 |
self.k_value = k
|
197 |
return f"Loaded {file_} [GPU INFERENCE]"
|
198 |
except:
|
199 |
+
self.change_llm("TheBloke/Llama-2-7B-Chat-GGML", "llama-2-7b-chat.ggmlv3.q5_1.bin", max_tokens=256, temperature=0.2, top_p=0.95, top_k=50, repeat_penalty=1.2, k=3)
|
200 |
+
return "No valid model | Reloaded Reloaded default llama-2 7B config"
|
201 |
else:
|
202 |
try:
|
203 |
model_path = hf_hub_download(repo_id=repo_, filename=file_)
|
204 |
+
|
205 |
+
self.qa = None
|
206 |
+
self.llm = None
|
207 |
+
gc.collect()
|
208 |
+
torch.cuda.empty_cache()
|
209 |
|
210 |
self.llm = LlamaCpp(
|
211 |
model_path=model_path,
|
212 |
+
n_ctx=2048,
|
213 |
n_batch=8,
|
214 |
max_tokens=max_tokens,
|
215 |
verbose=False,
|
|
|
222 |
self.k_value = k
|
223 |
return f"Loaded {file_} [CPU INFERENCE SLOW]"
|
224 |
except:
|
225 |
+
self.change_llm("TheBloke/Llama-2-7B-Chat-GGML", "llama-2-7b-chat.ggmlv3.q5_1.bin", max_tokens=256, temperature=0.2, top_p=0.95, top_k=50, repeat_penalty=1.2, k=3)
|
226 |
+
return "No valid model | Reloaded default llama-2 7B config"
|
227 |
|
228 |
+
def default_falcon_model(self, HF_TOKEN):
|
229 |
+
self.llm = llm_api=HuggingFaceHub(
|
230 |
+
huggingfacehub_api_token=HF_TOKEN,
|
231 |
+
repo_id="tiiuae/falcon-7b-instruct",
|
232 |
+
model_kwargs={
|
233 |
+
"temperature":0.2,
|
234 |
+
"max_new_tokens":500,
|
235 |
+
"top_k":50,
|
236 |
+
"top_p":0.95,
|
237 |
+
"repetition_penalty":1.2,
|
238 |
+
},)
|
239 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
240 |
return "Loaded model Falcon 7B-instruct [API FAST INFERENCE]"
|
241 |
|
|
|
243 |
self.llm = ChatOpenAI(temperature=0, openai_api_key=API_KEY, model_name='gpt-3.5-turbo')
|
244 |
self.qa = q_a(self.db, "stuff", self.k_value, self.llm)
|
245 |
API_KEY = ""
|
246 |
+
return "Loaded model OpenAI gpt-3.5-turbo [API FAST INFERENCE]"
|
247 |
|
248 |
@param.depends('db_query ', )
|
249 |
def get_lquest(self):
|
conversadocs/llamacppmodels.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
from typing import Any, Dict, Iterator, List, Optional
|
3 |
+
|
4 |
+
from pydantic import Field, root_validator
|
5 |
+
|
6 |
+
from langchain.callbacks.manager import CallbackManagerForLLMRun
|
7 |
+
from langchain.llms.base import LLM
|
8 |
+
from langchain.schema.output import GenerationChunk
|
9 |
+
|
10 |
+
logger = logging.getLogger(__name__)
|
11 |
+
|
12 |
+
|
13 |
+
class LlamaCpp(LLM):
|
14 |
+
"""llama.cpp model.
|
15 |
+
|
16 |
+
To use, you should have the llama-cpp-python library installed, and provide the
|
17 |
+
path to the Llama model as a named parameter to the constructor.
|
18 |
+
Check out: https://github.com/abetlen/llama-cpp-python
|
19 |
+
|
20 |
+
Example:
|
21 |
+
.. code-block:: python
|
22 |
+
|
23 |
+
from langchain.llms import LlamaCpp
|
24 |
+
llm = LlamaCpp(model_path="/path/to/llama/model")
|
25 |
+
"""
|
26 |
+
|
27 |
+
client: Any #: :meta private:
|
28 |
+
model_path: str
|
29 |
+
"""The path to the Llama model file."""
|
30 |
+
|
31 |
+
lora_base: Optional[str] = None
|
32 |
+
"""The path to the Llama LoRA base model."""
|
33 |
+
|
34 |
+
lora_path: Optional[str] = None
|
35 |
+
"""The path to the Llama LoRA. If None, no LoRa is loaded."""
|
36 |
+
|
37 |
+
n_ctx: int = Field(512, alias="n_ctx")
|
38 |
+
"""Token context window."""
|
39 |
+
|
40 |
+
n_parts: int = Field(-1, alias="n_parts")
|
41 |
+
"""Number of parts to split the model into.
|
42 |
+
If -1, the number of parts is automatically determined."""
|
43 |
+
|
44 |
+
seed: int = Field(-1, alias="seed")
|
45 |
+
"""Seed. If -1, a random seed is used."""
|
46 |
+
|
47 |
+
f16_kv: bool = Field(True, alias="f16_kv")
|
48 |
+
"""Use half-precision for key/value cache."""
|
49 |
+
|
50 |
+
logits_all: bool = Field(False, alias="logits_all")
|
51 |
+
"""Return logits for all tokens, not just the last token."""
|
52 |
+
|
53 |
+
vocab_only: bool = Field(False, alias="vocab_only")
|
54 |
+
"""Only load the vocabulary, no weights."""
|
55 |
+
|
56 |
+
use_mlock: bool = Field(False, alias="use_mlock")
|
57 |
+
"""Force system to keep model in RAM."""
|
58 |
+
|
59 |
+
n_threads: Optional[int] = Field(None, alias="n_threads")
|
60 |
+
"""Number of threads to use.
|
61 |
+
If None, the number of threads is automatically determined."""
|
62 |
+
|
63 |
+
n_batch: Optional[int] = Field(8, alias="n_batch")
|
64 |
+
"""Number of tokens to process in parallel.
|
65 |
+
Should be a number between 1 and n_ctx."""
|
66 |
+
|
67 |
+
n_gpu_layers: Optional[int] = Field(None, alias="n_gpu_layers")
|
68 |
+
"""Number of layers to be loaded into gpu memory. Default None."""
|
69 |
+
|
70 |
+
suffix: Optional[str] = Field(None)
|
71 |
+
"""A suffix to append to the generated text. If None, no suffix is appended."""
|
72 |
+
|
73 |
+
max_tokens: Optional[int] = 256
|
74 |
+
"""The maximum number of tokens to generate."""
|
75 |
+
|
76 |
+
temperature: Optional[float] = 0.8
|
77 |
+
"""The temperature to use for sampling."""
|
78 |
+
|
79 |
+
top_p: Optional[float] = 0.95
|
80 |
+
"""The top-p value to use for sampling."""
|
81 |
+
|
82 |
+
logprobs: Optional[int] = Field(None)
|
83 |
+
"""The number of logprobs to return. If None, no logprobs are returned."""
|
84 |
+
|
85 |
+
echo: Optional[bool] = False
|
86 |
+
"""Whether to echo the prompt."""
|
87 |
+
|
88 |
+
stop: Optional[List[str]] = []
|
89 |
+
"""A list of strings to stop generation when encountered."""
|
90 |
+
|
91 |
+
repeat_penalty: Optional[float] = 1.1
|
92 |
+
"""The penalty to apply to repeated tokens."""
|
93 |
+
|
94 |
+
top_k: Optional[int] = 40
|
95 |
+
"""The top-k value to use for sampling."""
|
96 |
+
|
97 |
+
last_n_tokens_size: Optional[int] = 64
|
98 |
+
"""The number of tokens to look back when applying the repeat_penalty."""
|
99 |
+
|
100 |
+
use_mmap: Optional[bool] = True
|
101 |
+
"""Whether to keep the model loaded in RAM"""
|
102 |
+
|
103 |
+
rope_freq_scale: float = 1.0
|
104 |
+
"""Scale factor for rope sampling."""
|
105 |
+
|
106 |
+
rope_freq_base: float = 10000.0
|
107 |
+
"""Base frequency for rope sampling."""
|
108 |
+
|
109 |
+
streaming: bool = True
|
110 |
+
"""Whether to stream the results, token by token."""
|
111 |
+
|
112 |
+
verbose: bool = True
|
113 |
+
"""Print verbose output to stderr."""
|
114 |
+
|
115 |
+
n_gqa: Optional[int] = None
|
116 |
+
|
117 |
+
@root_validator()
|
118 |
+
def validate_environment(cls, values: Dict) -> Dict:
|
119 |
+
"""Validate that llama-cpp-python library is installed."""
|
120 |
+
|
121 |
+
|
122 |
+
model_path = values["model_path"]
|
123 |
+
model_param_names = [
|
124 |
+
"n_gqa",
|
125 |
+
"rope_freq_scale",
|
126 |
+
"rope_freq_base",
|
127 |
+
"lora_path",
|
128 |
+
"lora_base",
|
129 |
+
"n_ctx",
|
130 |
+
"n_parts",
|
131 |
+
"seed",
|
132 |
+
"f16_kv",
|
133 |
+
"logits_all",
|
134 |
+
"vocab_only",
|
135 |
+
"use_mlock",
|
136 |
+
"n_threads",
|
137 |
+
"n_batch",
|
138 |
+
"use_mmap",
|
139 |
+
"last_n_tokens_size",
|
140 |
+
"verbose",
|
141 |
+
]
|
142 |
+
model_params = {k: values[k] for k in model_param_names}
|
143 |
+
|
144 |
+
model_params['n_gqa'] = 8 if '70B' in model_path.upper() else None # (TEMPORARY) must be 8 for llama2 70b
|
145 |
+
# For backwards compatibility, only include if non-null.
|
146 |
+
if values["n_gpu_layers"] is not None:
|
147 |
+
model_params["n_gpu_layers"] = values["n_gpu_layers"]
|
148 |
+
|
149 |
+
try:
|
150 |
+
from llama_cpp import Llama
|
151 |
+
|
152 |
+
values["client"] = Llama(model_path, **model_params)
|
153 |
+
except ImportError:
|
154 |
+
raise ImportError(
|
155 |
+
"Could not import llama-cpp-python library. "
|
156 |
+
"Please install the llama-cpp-python library to "
|
157 |
+
"use this embedding model: pip install llama-cpp-python"
|
158 |
+
)
|
159 |
+
except Exception as e:
|
160 |
+
raise ValueError(
|
161 |
+
f"Could not load Llama model from path: {model_path}. "
|
162 |
+
f"Received error {e}"
|
163 |
+
)
|
164 |
+
|
165 |
+
return values
|
166 |
+
|
167 |
+
@property
|
168 |
+
def _default_params(self) -> Dict[str, Any]:
|
169 |
+
"""Get the default parameters for calling llama_cpp."""
|
170 |
+
return {
|
171 |
+
"suffix": self.suffix,
|
172 |
+
"max_tokens": self.max_tokens,
|
173 |
+
"temperature": self.temperature,
|
174 |
+
"top_p": self.top_p,
|
175 |
+
"logprobs": self.logprobs,
|
176 |
+
"echo": self.echo,
|
177 |
+
"stop_sequences": self.stop, # key here is convention among LLM classes
|
178 |
+
"repeat_penalty": self.repeat_penalty,
|
179 |
+
"top_k": self.top_k,
|
180 |
+
}
|
181 |
+
|
182 |
+
@property
|
183 |
+
def _identifying_params(self) -> Dict[str, Any]:
|
184 |
+
"""Get the identifying parameters."""
|
185 |
+
return {**{"model_path": self.model_path}, **self._default_params}
|
186 |
+
|
187 |
+
@property
|
188 |
+
def _llm_type(self) -> str:
|
189 |
+
"""Return type of llm."""
|
190 |
+
return "llamacpp"
|
191 |
+
|
192 |
+
def _get_parameters(self, stop: Optional[List[str]] = None) -> Dict[str, Any]:
|
193 |
+
"""
|
194 |
+
Performs sanity check, preparing parameters in format needed by llama_cpp.
|
195 |
+
|
196 |
+
Args:
|
197 |
+
stop (Optional[List[str]]): List of stop sequences for llama_cpp.
|
198 |
+
|
199 |
+
Returns:
|
200 |
+
Dictionary containing the combined parameters.
|
201 |
+
"""
|
202 |
+
|
203 |
+
# Raise error if stop sequences are in both input and default params
|
204 |
+
if self.stop and stop is not None:
|
205 |
+
raise ValueError("`stop` found in both the input and default params.")
|
206 |
+
|
207 |
+
params = self._default_params
|
208 |
+
|
209 |
+
# llama_cpp expects the "stop" key not this, so we remove it:
|
210 |
+
params.pop("stop_sequences")
|
211 |
+
|
212 |
+
# then sets it as configured, or default to an empty list:
|
213 |
+
params["stop"] = self.stop or stop or []
|
214 |
+
|
215 |
+
return params
|
216 |
+
|
217 |
+
def _call(
|
218 |
+
self,
|
219 |
+
prompt: str,
|
220 |
+
stop: Optional[List[str]] = None,
|
221 |
+
run_manager: Optional[CallbackManagerForLLMRun] = None,
|
222 |
+
**kwargs: Any,
|
223 |
+
) -> str:
|
224 |
+
"""Call the Llama model and return the output.
|
225 |
+
|
226 |
+
Args:
|
227 |
+
prompt: The prompt to use for generation.
|
228 |
+
stop: A list of strings to stop generation when encountered.
|
229 |
+
|
230 |
+
Returns:
|
231 |
+
The generated text.
|
232 |
+
|
233 |
+
Example:
|
234 |
+
.. code-block:: python
|
235 |
+
|
236 |
+
from langchain.llms import LlamaCpp
|
237 |
+
llm = LlamaCpp(model_path="/path/to/local/llama/model.bin")
|
238 |
+
llm("This is a prompt.")
|
239 |
+
"""
|
240 |
+
if self.streaming:
|
241 |
+
# If streaming is enabled, we use the stream
|
242 |
+
# method that yields as they are generated
|
243 |
+
# and return the combined strings from the first choices's text:
|
244 |
+
combined_text_output = ""
|
245 |
+
for chunk in self._stream(
|
246 |
+
prompt=prompt, stop=stop, run_manager=run_manager, **kwargs
|
247 |
+
):
|
248 |
+
combined_text_output += chunk.text
|
249 |
+
return combined_text_output
|
250 |
+
else:
|
251 |
+
params = self._get_parameters(stop)
|
252 |
+
params = {**params, **kwargs}
|
253 |
+
result = self.client(prompt=prompt, **params)
|
254 |
+
return result["choices"][0]["text"]
|
255 |
+
|
256 |
+
def _stream(
|
257 |
+
self,
|
258 |
+
prompt: str,
|
259 |
+
stop: Optional[List[str]] = None,
|
260 |
+
run_manager: Optional[CallbackManagerForLLMRun] = None,
|
261 |
+
**kwargs: Any,
|
262 |
+
) -> Iterator[GenerationChunk]:
|
263 |
+
"""Yields results objects as they are generated in real time.
|
264 |
+
|
265 |
+
It also calls the callback manager's on_llm_new_token event with
|
266 |
+
similar parameters to the OpenAI LLM class method of the same name.
|
267 |
+
|
268 |
+
Args:
|
269 |
+
prompt: The prompts to pass into the model.
|
270 |
+
stop: Optional list of stop words to use when generating.
|
271 |
+
|
272 |
+
Returns:
|
273 |
+
A generator representing the stream of tokens being generated.
|
274 |
+
|
275 |
+
Yields:
|
276 |
+
A dictionary like objects containing a string token and metadata.
|
277 |
+
See llama-cpp-python docs and below for more.
|
278 |
+
|
279 |
+
Example:
|
280 |
+
.. code-block:: python
|
281 |
+
|
282 |
+
from langchain.llms import LlamaCpp
|
283 |
+
llm = LlamaCpp(
|
284 |
+
model_path="/path/to/local/model.bin",
|
285 |
+
temperature = 0.5
|
286 |
+
)
|
287 |
+
for chunk in llm.stream("Ask 'Hi, how are you?' like a pirate:'",
|
288 |
+
stop=["'","\n"]):
|
289 |
+
result = chunk["choices"][0]
|
290 |
+
print(result["text"], end='', flush=True)
|
291 |
+
|
292 |
+
"""
|
293 |
+
params = {**self._get_parameters(stop), **kwargs}
|
294 |
+
result = self.client(prompt=prompt, stream=True, **params)
|
295 |
+
for part in result:
|
296 |
+
logprobs = part["choices"][0].get("logprobs", None)
|
297 |
+
chunk = GenerationChunk(
|
298 |
+
text=part["choices"][0]["text"],
|
299 |
+
generation_info={"logprobs": logprobs},
|
300 |
+
)
|
301 |
+
yield chunk
|
302 |
+
if run_manager:
|
303 |
+
run_manager.on_llm_new_token(
|
304 |
+
token=chunk.text, verbose=self.verbose, log_probs=logprobs
|
305 |
+
)
|
306 |
+
|
307 |
+
def get_num_tokens(self, text: str) -> int:
|
308 |
+
tokenized_text = self.client.tokenize(text.encode("utf-8"))
|
309 |
+
return len(tokenized_text)
|
conversadocs/llm_chess.py
ADDED
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import chess
|
3 |
+
import chess.pgn
|
4 |
+
|
5 |
+
# credit code https://github.com/notnil/chess-gpt
|
6 |
+
def get_legal_moves(board):
|
7 |
+
"""Returns a list of legal moves in UCI notation."""
|
8 |
+
return list(map(board.san, board.legal_moves))
|
9 |
+
|
10 |
+
def init_game() -> tuple[chess.pgn.Game, chess.Board]:
|
11 |
+
"""Initializes a new game."""
|
12 |
+
board = chess.Board()
|
13 |
+
game = chess.pgn.Game()
|
14 |
+
game.headers["White"] = "User"
|
15 |
+
game.headers["Black"] = "Chess-engine"
|
16 |
+
del game.headers["Event"]
|
17 |
+
del game.headers["Date"]
|
18 |
+
del game.headers["Site"]
|
19 |
+
del game.headers["Round"]
|
20 |
+
del game.headers["Result"]
|
21 |
+
game.setup(board)
|
22 |
+
return game, board
|
23 |
+
|
24 |
+
def generate_prompt(game: chess.pgn.Game, board: chess.Board) -> str:
|
25 |
+
|
26 |
+
moves = get_legal_moves(board)
|
27 |
+
moves_str = ",".join(moves)
|
28 |
+
return f"""
|
29 |
+
The task is play a Chess game:
|
30 |
+
You are the Chess-engine playing a chess match against the user as black and trying to win.
|
31 |
+
|
32 |
+
The current FEN notation is:
|
33 |
+
{board.fen()}
|
34 |
+
|
35 |
+
The next valid moves are:
|
36 |
+
{moves_str}
|
37 |
+
|
38 |
+
Continue the game.
|
39 |
+
{str(game)[:-2]}"""
|
40 |
+
|
41 |
+
def get_move(content, moves):
|
42 |
+
lines = content.splitlines()
|
43 |
+
for line in lines:
|
44 |
+
for lm in moves:
|
45 |
+
if lm in line:
|
46 |
+
return lm
|
47 |
+
|
48 |
+
class ChessGame:
|
49 |
+
def __init__(self, docschatllm):
|
50 |
+
self.docschatllm = docschatllm
|
51 |
+
|
52 |
+
def start_game(self):
|
53 |
+
self.game, self.board = init_game()
|
54 |
+
self.game_cp, _ = init_game()
|
55 |
+
self.node = self.game
|
56 |
+
self.node_copy = self.game_cp
|
57 |
+
|
58 |
+
svg_board = chess.svg.board(self.board, size=350)
|
59 |
+
return svg_board, "Valid moves: "+",".join(get_legal_moves(self.board)) # display(self.board)
|
60 |
+
|
61 |
+
def user_move(self, move_input):
|
62 |
+
try:
|
63 |
+
self.board.push_san(move_input)
|
64 |
+
except ValueError:
|
65 |
+
print("Invalid move")
|
66 |
+
svg_board = chess.svg.board(self.board, size=350)
|
67 |
+
return svg_board, "Valid moves: "+",".join(get_legal_moves(self.board)), 'Invalid move'
|
68 |
+
self.node = self.node.add_variation(self.board.move_stack[-1])
|
69 |
+
self.node_copy = self.node_copy.add_variation(self.board.move_stack[-1])
|
70 |
+
|
71 |
+
if self.board.is_game_over():
|
72 |
+
svg_board = chess.svg.board(self.board, size=350)
|
73 |
+
return svg_board, ",".join(get_legal_moves(self.board)), 'GAME OVER'
|
74 |
+
|
75 |
+
prompt = generate_prompt(self.game, self.board)
|
76 |
+
print("Prompt: \n"+prompt)
|
77 |
+
print("#############")
|
78 |
+
for i in range(10): #tries
|
79 |
+
if i == 9:
|
80 |
+
svg_board = chess.svg.board(self.board, size=350)
|
81 |
+
return svg_board, ",".join(get_legal_moves(self.board)), "The model can't do a valid move"
|
82 |
+
try:
|
83 |
+
"""Returns the move from the prompt."""
|
84 |
+
content = self.docschatllm.llm.predict(prompt) ### from selected model ###
|
85 |
+
#print(moves)
|
86 |
+
print("Response: \n"+content)
|
87 |
+
print("#############")
|
88 |
+
|
89 |
+
moves = get_legal_moves(self.board)
|
90 |
+
move = get_move(content, moves)
|
91 |
+
print(move)
|
92 |
+
print("#############")
|
93 |
+
self.board.push_san(move)
|
94 |
+
break
|
95 |
+
except:
|
96 |
+
prompt = prompt[1:]
|
97 |
+
print("attempt a move.")
|
98 |
+
self.node = self.node.add_variation(self.board.move_stack[-1])
|
99 |
+
self.node_copy = self.node_copy.add_variation(self.board.move_stack[-1])
|
100 |
+
svg_board = chess.svg.board(self.board, size=350)
|
101 |
+
return svg_board, "Valid moves: "+",".join(get_legal_moves(self.board)), ''
|