Instructions to use openeurollm/oellm-9b-256k-theta64m-prelude with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use openeurollm/oellm-9b-256k-theta64m-prelude with Transformers:

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

pipe = pipeline("text-generation", model="openeurollm/oellm-9b-256k-theta64m-prelude")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("openeurollm/oellm-9b-256k-theta64m-prelude")
model = AutoModelForCausalLM.from_pretrained("openeurollm/oellm-9b-256k-theta64m-prelude", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use openeurollm/oellm-9b-256k-theta64m-prelude with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "openeurollm/oellm-9b-256k-theta64m-prelude"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "openeurollm/oellm-9b-256k-theta64m-prelude",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/openeurollm/oellm-9b-256k-theta64m-prelude

SGLang

How to use openeurollm/oellm-9b-256k-theta64m-prelude 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 "openeurollm/oellm-9b-256k-theta64m-prelude" \
    --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": "openeurollm/oellm-9b-256k-theta64m-prelude",
		"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 "openeurollm/oellm-9b-256k-theta64m-prelude" \
        --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": "openeurollm/oellm-9b-256k-theta64m-prelude",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use openeurollm/oellm-9b-256k-theta64m-prelude with Docker Model Runner:
```
docker model run hf.co/openeurollm/oellm-9b-256k-theta64m-prelude
```

birgermoell commited on 9 days ago

Commit

55e6060

verified ·

1 Parent(s): 6dad799

Upload README.md with huggingface_hub

Browse files

Files changed (1) hide show

README.md +5 -3

README.md CHANGED Viewed

@@ -33,12 +33,14 @@ multilingual portion prevents non-English regression during the 256K stage.
 [`scripts/eval_base_lm_niah.py`](https://github.com/BirgerMoell/openeuro-longctx-datamix). The 256K
 forward needs multi-GPU (`device_map="auto"` over a full 8-GCD node; one GCD OOMs at 256K).
-**@256K** (multilingual sweep in progress; ~92% so far, **depth-0 = 100%**):
 | depth | 0.0 | 0.25 | 0.5 | 0.75 | 1.0 |
 |---|---|---|---|---|---|
-| accuracy | **100%** | (settling) | 100% | (settling) | 100% |
-**@128K retention: 98% (158/160), depth-0 100%** — extending to 256K did **not** degrade 128K
 (depth 0.5 = 93%, the mild mid-depth softness carried from the 128K model; see that model's card).
 ## Architecture

 [`scripts/eval_base_lm_niah.py`](https://github.com/BirgerMoell/openeuro-longctx-datamix). The 256K
 forward needs multi-GPU (`device_map="auto"` over a full 8-GCD node; one GCD OOMs at 256K).
+**@256K** (7 of 15 languages, 185 trials; full sweep in progress). **Overall 90% (168/185).**
 | depth | 0.0 | 0.25 | 0.5 | 0.75 | 1.0 |
 |---|---|---|---|---|---|
+| accuracy | 93% | 86% | 87% | 88% | 100% |
+By language: fr 100%, en 98%, pt 96%, pl 93%, de 83%, el 81%, sv 76%. **depth-0 = 93%** confirms
+theta=64M solves far-position retrieval at 256K. Mid-depths (0.25/0.5/0.75) in the mid-80s reflect
+the mild lost-in-the-middle inherited from the 1T base (see 128K-prelude card). 128K retention: 98%.
 (depth 0.5 = 93%, the mild mid-depth softness carried from the 128K model; see that model's card).
 ## Architecture