IST199655
commited on
Commit
Β·
59e0922
1
Parent(s):
192eae5
app.py
CHANGED
|
@@ -4,71 +4,34 @@ from huggingface_hub import InferenceClient
|
|
| 4 |
"""
|
| 5 |
Copied from inference in colab notebook
|
| 6 |
"""
|
| 7 |
-
import torch
|
| 8 |
|
| 9 |
-
# Monkey-patch to avoid CUDA initialization issues
|
| 10 |
-
torch.cuda.get_device_capability = lambda *args, **kwargs: (0, 0)
|
| 11 |
|
| 12 |
-
from unsloth.chat_templates import get_chat_template
|
| 13 |
-
from unsloth import FastLanguageModel
|
| 14 |
|
| 15 |
-
# IMPORTING MODEL AND TOKENIZER ββββββββ
|
| 16 |
|
| 17 |
-
max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
|
| 18 |
-
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
| 19 |
-
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
|
| 20 |
|
| 21 |
-
model, tokenizer = FastLanguageModel.from_pretrained(
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
)
|
| 27 |
-
|
| 28 |
-
tokenizer = get_chat_template(
|
| 29 |
-
tokenizer,
|
| 30 |
-
chat_template = "llama-3.1",
|
| 31 |
-
)
|
| 32 |
-
FastLanguageModel.for_inference(model) # Enable native 2x faster inference
|
| 33 |
-
|
| 34 |
-
# RUNNING INFERENCE ββββββββββββββββββββββββ
|
| 35 |
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
temperature,
|
| 42 |
-
top_p,
|
| 43 |
-
):
|
| 44 |
-
messages = [{"role": "system", "content": system_message}]
|
| 45 |
-
|
| 46 |
-
for val in history:
|
| 47 |
-
if val[0]:
|
| 48 |
-
messages.append({"role": "user", "content": val[0]})
|
| 49 |
-
if val[1]:
|
| 50 |
-
messages.append({"role": "assistant", "content": val[1]})
|
| 51 |
-
|
| 52 |
-
messages.append({"role": "user", "content": message})
|
| 53 |
-
|
| 54 |
-
inputs = tokenizer.apply_chat_template(
|
| 55 |
-
messages,
|
| 56 |
-
tokenize = True,
|
| 57 |
-
add_generation_prompt = True, # Must add for generation
|
| 58 |
-
return_tensors = "pt",
|
| 59 |
-
)
|
| 60 |
-
|
| 61 |
-
outputs = model.generate(input_ids = inputs, max_new_tokens = max_tokens, use_cache = True,
|
| 62 |
-
temperature = 1.5, min_p = 0.1)
|
| 63 |
-
response = tokenizer.batch_decode(outputs)
|
| 64 |
-
|
| 65 |
-
yield response
|
| 66 |
-
|
| 67 |
-
"""
|
| 68 |
-
For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
|
| 69 |
-
"""
|
| 70 |
-
# client = InferenceClient("llama_lora_model_1")
|
| 71 |
|
|
|
|
| 72 |
|
| 73 |
# def respond(
|
| 74 |
# message,
|
|
@@ -88,19 +51,56 @@ For more information on `huggingface_hub` Inference API support, please check th
|
|
| 88 |
|
| 89 |
# messages.append({"role": "user", "content": message})
|
| 90 |
|
| 91 |
-
#
|
| 92 |
-
|
| 93 |
-
# for message in client.chat_completion(
|
| 94 |
# messages,
|
| 95 |
-
#
|
| 96 |
-
#
|
| 97 |
-
#
|
| 98 |
-
#
|
| 99 |
-
|
| 100 |
-
#
|
| 101 |
-
|
| 102 |
-
#
|
| 103 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 104 |
|
| 105 |
|
| 106 |
"""
|
|
|
|
| 4 |
"""
|
| 5 |
Copied from inference in colab notebook
|
| 6 |
"""
|
| 7 |
+
# import torch
|
| 8 |
|
| 9 |
+
# # Monkey-patch to avoid CUDA initialization issues
|
| 10 |
+
# torch.cuda.get_device_capability = lambda *args, **kwargs: (0, 0)
|
| 11 |
|
| 12 |
+
# from unsloth.chat_templates import get_chat_template
|
| 13 |
+
# from unsloth import FastLanguageModel
|
| 14 |
|
| 15 |
+
# # IMPORTING MODEL AND TOKENIZER ββββββββ
|
| 16 |
|
| 17 |
+
# max_seq_length = 2048 # Choose any! We auto support RoPE Scaling internally!
|
| 18 |
+
# dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
|
| 19 |
+
# load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
|
| 20 |
|
| 21 |
+
# model, tokenizer = FastLanguageModel.from_pretrained(
|
| 22 |
+
# model_name = "llama_lora_model_1",
|
| 23 |
+
# max_seq_length = max_seq_length,
|
| 24 |
+
# dtype = dtype,
|
| 25 |
+
# load_in_4bit = load_in_4bit,
|
| 26 |
+
# )
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
+
# tokenizer = get_chat_template(
|
| 29 |
+
# tokenizer,
|
| 30 |
+
# chat_template = "llama-3.1",
|
| 31 |
+
# )
|
| 32 |
+
# FastLanguageModel.for_inference(model) # Enable native 2x faster inference
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
+
# # RUNNING INFERENCE ββββββββββββββββββββββββ
|
| 35 |
|
| 36 |
# def respond(
|
| 37 |
# message,
|
|
|
|
| 51 |
|
| 52 |
# messages.append({"role": "user", "content": message})
|
| 53 |
|
| 54 |
+
# inputs = tokenizer.apply_chat_template(
|
|
|
|
|
|
|
| 55 |
# messages,
|
| 56 |
+
# tokenize = True,
|
| 57 |
+
# add_generation_prompt = True, # Must add for generation
|
| 58 |
+
# return_tensors = "pt",
|
| 59 |
+
# )
|
| 60 |
+
|
| 61 |
+
# outputs = model.generate(input_ids = inputs, max_new_tokens = max_tokens, use_cache = True,
|
| 62 |
+
# temperature = 1.5, min_p = 0.1)
|
| 63 |
+
# response = tokenizer.batch_decode(outputs)
|
| 64 |
+
|
| 65 |
+
# yield response
|
| 66 |
+
|
| 67 |
+
"""
|
| 68 |
+
For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
|
| 69 |
+
"""
|
| 70 |
+
client = InferenceClient("llama_lora_model_1")
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def respond(
|
| 74 |
+
message,
|
| 75 |
+
history: list[tuple[str, str]],
|
| 76 |
+
system_message,
|
| 77 |
+
max_tokens,
|
| 78 |
+
temperature,
|
| 79 |
+
top_p,
|
| 80 |
+
):
|
| 81 |
+
messages = [{"role": "system", "content": system_message}]
|
| 82 |
+
|
| 83 |
+
for val in history:
|
| 84 |
+
if val[0]:
|
| 85 |
+
messages.append({"role": "user", "content": val[0]})
|
| 86 |
+
if val[1]:
|
| 87 |
+
messages.append({"role": "assistant", "content": val[1]})
|
| 88 |
+
|
| 89 |
+
messages.append({"role": "user", "content": message})
|
| 90 |
+
|
| 91 |
+
response = ""
|
| 92 |
+
|
| 93 |
+
for message in client.chat_completion(
|
| 94 |
+
messages,
|
| 95 |
+
max_tokens=max_tokens,
|
| 96 |
+
stream=True,
|
| 97 |
+
temperature=temperature,
|
| 98 |
+
top_p=top_p,
|
| 99 |
+
):
|
| 100 |
+
token = message.choices[0].delta.content
|
| 101 |
+
|
| 102 |
+
response += token
|
| 103 |
+
yield response
|
| 104 |
|
| 105 |
|
| 106 |
"""
|