tested app_modules/llm_loader.py
Browse files- .env.example +3 -3
- app_modules/llm_loader.py +15 -18
- test.py +19 -10
.env.example
CHANGED
@@ -54,13 +54,13 @@ MOSAICML_MODEL_NAME_OR_PATH="mosaicml/mpt-7b-instruct"
|
|
54 |
|
55 |
FALCON_MODEL_NAME_OR_PATH="tiiuae/falcon-7b-instruct"
|
56 |
|
57 |
-
GPT4ALL_J_MODEL_PATH="../models/llama-2-7b-chat.ggmlv3.
|
58 |
GPT4ALL_J_DOWNLOAD_LINK=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q4_0.bin
|
59 |
|
60 |
-
GPT4ALL_MODEL_PATH="
|
61 |
GPT4ALL_DOWNLOAD_LINK=https://gpt4all.io/models/ggml-nous-gpt4-vicuna-13b.bin
|
62 |
|
63 |
-
LLAMACPP_MODEL_PATH="
|
64 |
LLAMACPP_DOWNLOAD_LINK=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q4_K_M.bin
|
65 |
|
66 |
# Index for AI Books PDF files - chunk_size=1024 chunk_overlap=512
|
|
|
54 |
|
55 |
FALCON_MODEL_NAME_OR_PATH="tiiuae/falcon-7b-instruct"
|
56 |
|
57 |
+
GPT4ALL_J_MODEL_PATH="../models/llama-2-7b-chat.ggmlv3.q4_K_M.bin"
|
58 |
GPT4ALL_J_DOWNLOAD_LINK=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q4_0.bin
|
59 |
|
60 |
+
GPT4ALL_MODEL_PATH="../models/ggml-nous-gpt4-vicuna-13b.bin"
|
61 |
GPT4ALL_DOWNLOAD_LINK=https://gpt4all.io/models/ggml-nous-gpt4-vicuna-13b.bin
|
62 |
|
63 |
+
LLAMACPP_MODEL_PATH="../models/llama-2-7b-chat.ggmlv3.q4_K_M.bin"
|
64 |
LLAMACPP_DOWNLOAD_LINK=https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGML/resolve/main/llama-2-7b-chat.ggmlv3.q4_K_M.bin
|
65 |
|
66 |
# Index for AI Books PDF files - chunk_size=1024 chunk_overlap=512
|
app_modules/llm_loader.py
CHANGED
@@ -30,7 +30,7 @@ from transformers import (
|
|
30 |
)
|
31 |
|
32 |
from app_modules.instruct_pipeline import InstructionTextGenerationPipeline
|
33 |
-
from app_modules.utils import ensure_model_is_downloaded
|
34 |
|
35 |
|
36 |
class TextIteratorStreamer(TextStreamer, StreamingStdOutCallbackHandler):
|
@@ -336,7 +336,6 @@ class LLMLoader:
|
|
336 |
)
|
337 |
else:
|
338 |
if os.environ.get("DISABLE_MODEL_PRELOADING") != "true":
|
339 |
-
use_auth_token = None
|
340 |
model = (
|
341 |
AutoModelForSeq2SeqLM.from_pretrained(
|
342 |
MODEL_NAME_OR_PATH,
|
@@ -354,25 +353,23 @@ class LLMLoader:
|
|
354 |
)
|
355 |
print(f"Model memory footprint: {model.get_memory_footprint()}")
|
356 |
else:
|
357 |
-
use_auth_token = token
|
358 |
model = MODEL_NAME_OR_PATH
|
359 |
|
360 |
pipe = pipeline(
|
361 |
-
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
|
374 |
-
|
375 |
-
)
|
376 |
|
377 |
self.llm = HuggingFacePipeline(pipeline=pipe, callbacks=callbacks)
|
378 |
elif self.llm_model_type == "mosaicml":
|
|
|
30 |
)
|
31 |
|
32 |
from app_modules.instruct_pipeline import InstructionTextGenerationPipeline
|
33 |
+
from app_modules.utils import ensure_model_is_downloaded
|
34 |
|
35 |
|
36 |
class TextIteratorStreamer(TextStreamer, StreamingStdOutCallbackHandler):
|
|
|
336 |
)
|
337 |
else:
|
338 |
if os.environ.get("DISABLE_MODEL_PRELOADING") != "true":
|
|
|
339 |
model = (
|
340 |
AutoModelForSeq2SeqLM.from_pretrained(
|
341 |
MODEL_NAME_OR_PATH,
|
|
|
353 |
)
|
354 |
print(f"Model memory footprint: {model.get_memory_footprint()}")
|
355 |
else:
|
|
|
356 |
model = MODEL_NAME_OR_PATH
|
357 |
|
358 |
pipe = pipeline(
|
359 |
+
task,
|
360 |
+
model=model,
|
361 |
+
tokenizer=tokenizer,
|
362 |
+
streamer=self.streamer,
|
363 |
+
return_full_text=return_full_text, # langchain expects the full text
|
364 |
+
device=hf_pipeline_device_type,
|
365 |
+
torch_dtype=torch_dtype,
|
366 |
+
max_new_tokens=2048,
|
367 |
+
trust_remote_code=True,
|
368 |
+
temperature=temperature,
|
369 |
+
top_p=0.95,
|
370 |
+
top_k=0, # select from top 0 tokens (because zero, relies on top_p)
|
371 |
+
repetition_penalty=1.115,
|
372 |
+
)
|
|
|
373 |
|
374 |
self.llm = HuggingFacePipeline(pipeline=pipe, callbacks=callbacks)
|
375 |
elif self.llm_model_type == "mosaicml":
|
test.py
CHANGED
@@ -1,14 +1,21 @@
|
|
1 |
# project/test.py
|
2 |
|
|
|
3 |
import unittest
|
|
|
4 |
|
5 |
from langchain.callbacks.base import BaseCallbackHandler
|
6 |
from langchain.schema import HumanMessage
|
7 |
|
8 |
from app_modules.llm_loader import LLMLoader
|
9 |
-
from
|
10 |
|
11 |
-
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
|
14 |
class MyCustomHandler(BaseCallbackHandler):
|
@@ -32,7 +39,9 @@ class TestLLMLoader(unittest.TestCase):
|
|
32 |
def run_test_case(self, llm_model_type, query):
|
33 |
llm_loader = LLMLoader(llm_model_type)
|
34 |
start = timer()
|
35 |
-
llm_loader.init(
|
|
|
|
|
36 |
end = timer()
|
37 |
print(f"Model loaded in {end - start:.3f}s")
|
38 |
|
@@ -43,17 +52,17 @@ class TestLLMLoader(unittest.TestCase):
|
|
43 |
print(f"Inference completed in {end2 - end:.3f}s")
|
44 |
print(result)
|
45 |
|
46 |
-
def
|
47 |
-
self.run_test_case("openai",
|
48 |
|
49 |
-
def
|
50 |
-
self.run_test_case("llamacpp",
|
51 |
|
52 |
-
def
|
53 |
-
self.run_test_case("gpt4all-j",
|
54 |
|
55 |
def test_huggingface(self):
|
56 |
-
self.run_test_case("huggingface",
|
57 |
|
58 |
|
59 |
if __name__ == "__main__":
|
|
|
1 |
# project/test.py
|
2 |
|
3 |
+
import os
|
4 |
import unittest
|
5 |
+
from timeit import default_timer as timer
|
6 |
|
7 |
from langchain.callbacks.base import BaseCallbackHandler
|
8 |
from langchain.schema import HumanMessage
|
9 |
|
10 |
from app_modules.llm_loader import LLMLoader
|
11 |
+
from app_modules.utils import *
|
12 |
|
13 |
+
user_question = "What's the capital city of Malaysia?"
|
14 |
+
n_threds = int(os.environ.get("NUMBER_OF_CPU_CORES") or "4")
|
15 |
+
|
16 |
+
hf_embeddings_device_type, hf_pipeline_device_type = get_device_types()
|
17 |
+
print(f"hf_embeddings_device_type: {hf_embeddings_device_type}")
|
18 |
+
print(f"hf_pipeline_device_type: {hf_pipeline_device_type}")
|
19 |
|
20 |
|
21 |
class MyCustomHandler(BaseCallbackHandler):
|
|
|
39 |
def run_test_case(self, llm_model_type, query):
|
40 |
llm_loader = LLMLoader(llm_model_type)
|
41 |
start = timer()
|
42 |
+
llm_loader.init(
|
43 |
+
n_threds=n_threds, hf_pipeline_device_type=hf_pipeline_device_type
|
44 |
+
)
|
45 |
end = timer()
|
46 |
print(f"Model loaded in {end - start:.3f}s")
|
47 |
|
|
|
52 |
print(f"Inference completed in {end2 - end:.3f}s")
|
53 |
print(result)
|
54 |
|
55 |
+
def test_openai(self):
|
56 |
+
self.run_test_case("openai", user_question)
|
57 |
|
58 |
+
def test_llamacpp(self):
|
59 |
+
self.run_test_case("llamacpp", user_question)
|
60 |
|
61 |
+
def test_gpt4all_j(self):
|
62 |
+
self.run_test_case("gpt4all-j", user_question)
|
63 |
|
64 |
def test_huggingface(self):
|
65 |
+
self.run_test_case("huggingface", user_question)
|
66 |
|
67 |
|
68 |
if __name__ == "__main__":
|