Spaces:
Paused
Paused
Commit
·
2995eda
1
Parent(s):
3124ebe
Create app_v3.py
Browse files
app_v3.py
ADDED
|
@@ -0,0 +1,69 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import AutoTokenizer, TextStreamer, pipeline, logging
|
| 2 |
+
from auto_gptq import AutoGPTQForCausalLM, BaseQuantizeConfig
|
| 3 |
+
import time
|
| 4 |
+
|
| 5 |
+
model_name_or_path = "TheBloke/llama2_7b_chat_uncensored-GPTQ"
|
| 6 |
+
model_basename = "gptq_model-4bit-128g"
|
| 7 |
+
|
| 8 |
+
use_triton = False
|
| 9 |
+
|
| 10 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True, legacy=False)
|
| 11 |
+
|
| 12 |
+
model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,
|
| 13 |
+
model_basename=model_basename,
|
| 14 |
+
use_safetensors=True,
|
| 15 |
+
trust_remote_code=True,
|
| 16 |
+
device="cuda:0",
|
| 17 |
+
use_triton=use_triton,
|
| 18 |
+
quantize_config=None)
|
| 19 |
+
|
| 20 |
+
"""
|
| 21 |
+
To download from a specific branch, use the revision parameter, as in this example:
|
| 22 |
+
|
| 23 |
+
model = AutoGPTQForCausalLM.from_quantized(model_name_or_path,
|
| 24 |
+
revision="gptq-4bit-32g-actorder_True",
|
| 25 |
+
model_basename=model_basename,
|
| 26 |
+
use_safetensors=True,
|
| 27 |
+
trust_remote_code=True,
|
| 28 |
+
device="cuda:0",
|
| 29 |
+
quantize_config=None)
|
| 30 |
+
"""
|
| 31 |
+
|
| 32 |
+
prompt = "Tell me about AI"
|
| 33 |
+
prompt_template=f'''### HUMAN:
|
| 34 |
+
{prompt}
|
| 35 |
+
|
| 36 |
+
### RESPONSE:
|
| 37 |
+
'''
|
| 38 |
+
print("\n\n*** Generate:")
|
| 39 |
+
start_time = time.time()
|
| 40 |
+
input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
|
| 41 |
+
streamer = TextStreamer(tokenizer)
|
| 42 |
+
# output = model.generate(inputs=input_ids, temperature=0.7, max_new_tokens=512)
|
| 43 |
+
# print(tokenizer.decode(output[0]))
|
| 44 |
+
|
| 45 |
+
_ = model.generate(inputs=input_ids, streamer=streamer, temperature=0.7, max_new_tokens=512)
|
| 46 |
+
print(f"Inference time: {time.time() - start_time:.4f} seconds")
|
| 47 |
+
|
| 48 |
+
# Inference can also be done using transformers' pipeline
|
| 49 |
+
|
| 50 |
+
# Prevent printing spurious transformers error when using pipeline with AutoGPTQ
|
| 51 |
+
logging.set_verbosity(logging.CRITICAL)
|
| 52 |
+
|
| 53 |
+
print("*** Pipeline:")
|
| 54 |
+
start_time = time.time()
|
| 55 |
+
|
| 56 |
+
pipe = pipeline(
|
| 57 |
+
"text-generation",
|
| 58 |
+
model=model,
|
| 59 |
+
tokenizer=tokenizer,
|
| 60 |
+
streamer=streamer,
|
| 61 |
+
max_new_tokens=512,
|
| 62 |
+
temperature=0.7,
|
| 63 |
+
top_p=0.95,
|
| 64 |
+
repetition_penalty=1.15
|
| 65 |
+
)
|
| 66 |
+
|
| 67 |
+
pipe(prompt_template)
|
| 68 |
+
#print(pipe(prompt_template)[0]['generated_text'])
|
| 69 |
+
print(f"Inference time: {time.time() - start_time:.4f} seconds")
|