Instructions to use transformers-community/group-beam-search with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use transformers-community/group-beam-search with Transformers:

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

pipe = pipeline("text-generation", model="transformers-community/group-beam-search")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("transformers-community/group-beam-search")
model = AutoModelForCausalLM.from_pretrained("transformers-community/group-beam-search")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Local Apps Settings

vLLM

How to use transformers-community/group-beam-search with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "transformers-community/group-beam-search"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "transformers-community/group-beam-search",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/transformers-community/group-beam-search

SGLang

How to use transformers-community/group-beam-search 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 "transformers-community/group-beam-search" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "transformers-community/group-beam-search",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "transformers-community/group-beam-search" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "transformers-community/group-beam-search",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use transformers-community/group-beam-search with Docker Model Runner:
```
docker model run hf.co/transformers-community/group-beam-search
```

Support Transformers v5 cache handling

by lavrenko - opened 6 days ago

base: refs/heads/main

←

from: refs/pr/4

Discussion Files changed

+15

-1

lavrenko

6 days ago

•

edited 6 days ago

Fixes cached decoding with recent Transformers versions by passing next_sequence_length to prepare_inputs_for_generation during non-first cached decoding steps.

This addresses the cache-related deterministic-output corruption reported in #3.

A more detailed explanation and local test results are included in the discussion below.

fix(generation): support v5 cached decodingeab40484

lavrenko

6 days ago

Summary

Fix cached decoding with recent Transformers versions.

This updates the custom generation loop to pass next_sequence_length into prepare_inputs_for_generation. In cached decoding, Transformers v5 expects the decode loop to explicitly indicate that only the latest token should be forwarded after the first step. Without this, the cached path may forward the full growing sequence together with past_key_values, which can corrupt deterministic group beam outputs.

What changed

Detect whether the loop is on the first decoding iteration.
Use next_sequence_length=None for the prefill step and for use_cache=False, preserving the existing full-sequence behavior.
Use next_sequence_length=1 for cached non-first decoding steps, matching the expected Transformers v5 cache behavior.
Pass is_first_iteration through to prepare_inputs_for_generation.

This keeps use_cache=True supported instead of forcing use_cache=False.

Motivation

This addresses the issue discussed in #3: deterministic group beam search could produce different and sometimes corrupted outputs depending on whether cache was enabled.

Cache should be an optimization and should not materially change deterministic generation semantics.

Tested

I tested the updated custom generation locally with:

transformers==4.57.1
transformers==5.12.1

using the Qwen/Qwen2.5-0.5B-Instruct reproducer from discussion #3 with deterministic grouped beam search.

Addresses #3

lavrenko changed pull request status to open 6 days ago

RaushanTurganbay

Transformers Community org 6 days ago

Ah nice, thanks for checking it. That PR indeed must have been breaking, we'd ideally check and update other generation methods as well. Feel free to check and PR if you want, otherwise I will look a bit later (low prio 😃 )

RaushanTurganbay changed pull request status to merged 6 days ago

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