-
Notifications
You must be signed in to change notification settings - Fork 957
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
T5 support more model types #127
Comments
Hi @VictorM-PS , Before that, please kindly make the two changes below to make t5-large working:
As for the triton server question, are you aware of the corresponding TRT-LLM backend released at the same time as the main repo: https://github.com/triton-inference-server/tensorrtllm_backend? |
Hi @VictorM-PS , more general T5 and Flan-T5 support has been done and scheduled for TRT-LLM's 0.6.0 monthly release soon. Please stay tuned. |
Hi @symphonylyh, thank you very much for the heads-up and the amazing work! :) |
@VictorM-PS |
Let me close this issue as the support has been released. Feel free to reopen if needed. |
@symphonylyh Thank you for your efforts to support T5. Traceback (most recent call last):
File "/usr/lib/python3.10/configparser.py", line 791, in get
value = d[option]
File "/usr/lib/python3.10/collections/__init__.py", line 986, in __getitem__
return self.__missing__(key) # support subclasses that define __missing__
File "/usr/lib/python3.10/collections/__init__.py", line 978, in __missing__
raise KeyError(key)
KeyError: 'n_positions'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/code/TensorRT-LLM/examples/enc_dec/build.py", line 515, in <module>
run_build(component='encoder')
File "/code/TensorRT-LLM/examples/enc_dec/build.py", line 489, in run_build
args = parse_arguments(component)
File "/code/TensorRT-LLM/examples/enc_dec/build.py", line 235, in parse_arguments
args = parse_config(
File "/code/TensorRT-LLM/examples/enc_dec/build.py", line 40, in parse_config
args = globals()[f'parse_{model_type}_config'](config, component, args)
File "/code/TensorRT-LLM/examples/enc_dec/t5/weight.py", line 31, in parse_t5_config
args.n_positions = config.getint(component, 'n_positions')
File "/usr/lib/python3.10/configparser.py", line 820, in getint
return self._get_conv(section, option, int, raw=raw, vars=vars,
File "/usr/lib/python3.10/configparser.py", line 810, in _get_conv
return self._get(section, conv, option, raw=raw, vars=vars,
File "/usr/lib/python3.10/configparser.py", line 805, in _get
return conv(self.get(section, option, **kwargs))
File "/usr/lib/python3.10/configparser.py", line 794, in get
raise NoOptionError(option, section)
configparser.NoOptionError: No option 'n_positions' in section: 'encoder' According to huggingface/transformers#8047, the parameter 'n_positions' seems have been deleted long time ago. So I just add a |
Hi @yuanze1024 , thanks, yes it seems other Flan-T5 variants all have a |
Hi,
I am following the enc_dec example and after changing a bit your code and adapting it to my case, I am hitting a wall when defining the weights. I see the example is running t5-small, my version is a t5-large, but after changing the config.ini with my values I expected it to run.
The error:
I currently use fastertransformer_backend to optimize T5, but I see maintenance effort is dropped. Plus, it is not great not to be able to select GPU in the instance group config for triton server...
So I have two questions:
Thank you for the amazing work!
The text was updated successfully, but these errors were encountered: