Instructions to use OpenNLPLab/TransNormerLLM-7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use OpenNLPLab/TransNormerLLM-7B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="OpenNLPLab/TransNormerLLM-7B", trust_remote_code=True)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("OpenNLPLab/TransNormerLLM-7B", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use OpenNLPLab/TransNormerLLM-7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "OpenNLPLab/TransNormerLLM-7B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenNLPLab/TransNormerLLM-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/OpenNLPLab/TransNormerLLM-7B

SGLang

How to use OpenNLPLab/TransNormerLLM-7B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "OpenNLPLab/TransNormerLLM-7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenNLPLab/TransNormerLLM-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "OpenNLPLab/TransNormerLLM-7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "OpenNLPLab/TransNormerLLM-7B",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use OpenNLPLab/TransNormerLLM-7B with Docker Model Runner:
```
docker model run hf.co/OpenNLPLab/TransNormerLLM-7B
```

Unsafe use of eval

by Daniel-P-Gonzalez - opened Nov 13, 2023

Discussion

Daniel-P-Gonzalez

Nov 13, 2023

In modeling_transnormer.py and utils.py, eval is used to parse environment variables.
Instead of:

some_option = eval(os.environ.get("some_option", default="False"))

I would recommend using something like:

some_option = os.environ.get("some_option", default="False").lower() in ["true", "yes", "y", "1"]

Also, do_eval in particular is evaluated on every forward call for each NormLinearAttention layer. Is there a particular reason for this, or should it instead be a global?

Yhyu13

Jan 12, 2024

•

edited Jan 12, 2024

They are just poor codes, there should be a pr to improve these code.

The reason for eval in every forward is probably bc the authors has some testing code that can be made easy to switch between evaluation and training for their own convinence.
It shouldn't be this way though. In practice, we should replace the do_eval assignment with checks derived from user's previous call to pytorch model.eval(), ie. the model.training bool https://discuss.pytorch.org/t/check-if-model-is-eval-or-train/9395

Doreamonzzz

Jan 12, 2024

Hello, thank you for your suggestion. We will optimize the code in the future. The "do_eval" is related to attention calculation, and we will also update this in the future.

stereoplegic

Dec 2, 2024

Could we not just do is_cuda_available and possibly check torch.version.hip (for new MLIR Triton with ROCM support) in a general transformers.utils.import_utils.is_triton_available that would benefit every model author relying on Triton?

As for the other evals, wouldn't it make more sense just to set them as hyperparameters and/or args/kwargs?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment