inference time cpu vs gpu about spanmarkerner HOT 3 OPEN

You may experience improved speed if you use SpanMarkerModel.from_pretrained(..., torch_dtype=torch.float16) or torch.bfloat16. See e.g.:

import time
import torch
from span_marker import SpanMarkerModel

model = SpanMarkerModel.from_pretrained("tomaarsen/span-marker-roberta-large-fewnerd-fine-super", torch_dtype=torch.bfloat16, device_map="cuda")
# model = SpanMarkerModel.from_pretrained("tomaarsen/span-marker-roberta-large-fewnerd-fine-super", device_map="cuda")

text = [
    "Leonardo da Vinci recently published a scientific paper on combatting Mitocromulent disease. Leonardo da Vinci painted the most famous painting in existence: the Mona Lisa.",
    "Leonardo da Vinci scored a critical goal towards the end of the second half. Leonardo da Vinci controversially veto'd a bill regarding public health care last friday. Leonardo da Vinci was promoted to Sergeant after his outstanding work in the war."
]
BS = 64
N = 500
model.predict(text * 50, batch_size=BS)
start_t = time.time()
model.predict(text * N, batch_size=BS)
print(f"{time.time() - start_t:8f}s for {N * 2} samples with batch_size={BS} and torch_dtype={model.dtype}.")

This gave me:

20.745640s for 1000 samples with batch_size=64 and torch_dtype=torch.float16.

16.534876s for 1000 samples with batch_size=64 and torch_dtype=torch.bfloat16.

and

39.655506s for 1000 samples with batch_size=64 and torch_dtype=torch.float32.

Note that float16 is not available on CPU though! Not sure about bfloat16.

If you have a Linux (or Mac?) device, then you can also use load_in_8bit=True and load_in_4bit=True by installing bitsandbytes, but I don't know if that improves inference speed - this is also only for CUDA.

Beyond that the steps to increase the inference speeds become pretty challenging. Hope this helps a bit.

Also, you can process about 8 sentences per second with CPU and about 110 sentences per second in GPU, is that not sufficiently fast yet?

• Tom Aarsen

from spanmarkerner.

thanku @tomaarsen
Using torch.float16 was working for me. It would be excellent if the operation could be completed in less than one second with a batch size of 256.

from spanmarkerner.

@polodealvarado started working on ONNX support here: #26 (comment)
If we can make it work, perhaps then we can improve the speed even further. Until then, it will be hard to get even faster results. Less than a second for a batch size of 256 equals 256 sentences per second, that is already quite efficient.

• Tom Aarsen

from spanmarkerner.