Instructions to use microsoft/Florence-2-large-ft with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use microsoft/Florence-2-large-ft with Transformers:

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

pipe = pipeline("image-text-to-text", model="microsoft/Florence-2-large-ft", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large-ft", trust_remote_code=True)
model = AutoModelForMultimodalLM.from_pretrained("microsoft/Florence-2-large-ft", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use microsoft/Florence-2-large-ft with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "microsoft/Florence-2-large-ft"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "microsoft/Florence-2-large-ft",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/microsoft/Florence-2-large-ft

SGLang

How to use microsoft/Florence-2-large-ft 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 "microsoft/Florence-2-large-ft" \
    --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": "microsoft/Florence-2-large-ft",
		"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 "microsoft/Florence-2-large-ft" \
        --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": "microsoft/Florence-2-large-ft",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use microsoft/Florence-2-large-ft with Docker Model Runner:
```
docker model run hf.co/microsoft/Florence-2-large-ft
```

fix: compatibility patches for newer transformers (4.50+ / 5.x)

#44

by TheeRealDude - opened May 1

base: refs/heads/main

←

from: refs/pr/44

Discussion Files changed

+11

-18

TheeRealDude

May 1

Summary

Compatibility patches for newer transformers versions (4.50+ / 5.x) that broke Florence-2's custom code. Each fix is a minimal getattr/wrapper-style change that preserves behavior on older transformers while not crashing on newer ones.

Changes

1. `configuration_florence2.py` — `forced_bos_token_id` access

Newer transformers no longer auto-creates forced_bos_token_id on PretrainedConfig subclasses, causing AttributeError during Florence2LanguageConfig.__init__. Wrap with getattr(self, "forced_bos_token_id", None) for safe access.

2. `modeling_florence2.py` — `torch.linspace().item()` on meta tensors

Newer transformers instantiates the model on the meta device first (lazy loading), then materializes weights. The DaViT vision tower's __init__ calls [x.item() for x in torch.linspace(...)] which fails on meta tensors. Replaced with pure-Python list comprehension that doesn't touch tensors during init.

3. `modeling_florence2.py` — `_supports_sdpa` / `_supports_flash_attn_2` properties

These were @property methods that delegated to self.language_model._supports_sdpa. But during super().__init__() (parent class init), self.language_model doesn't exist yet, causing AttributeError. Replaced with class-level attributes (_supports_sdpa = True, _supports_flash_attn_2 = False) since the support state is known at class definition.

4. `processing_florence2.py` — `tokenizer.additional_special_tokens` access

Newer transformers RobertaTokenizer (and others) no longer expose additional_special_tokens as a direct attribute. Use getattr(tokenizer, 'additional_special_tokens', []) for safe access with empty-list fallback.

Caveat / Known Limitation

These patches resolve the Python-level errors during model loading. There remains a CUDA-level segfault during the SDPA forward pass with transformers 5.x + torch 2.10+ — that's a deeper compatibility issue with the language model's SDPA attention path that this PR does not address. As a workaround, users on transformers 5.x can use attn_implementation="eager" to avoid the SDPA path entirely.

For users on transformers <= 4.49, these 4 patches are sufficient for Florence-2 to work correctly with newer attribute access patterns.

Tested with

transformers 4.49.0 + torch 2.11.0+cu128 + CUDA 12.8 — works
transformers 5.5.0 — Python errors fixed by these patches, but separate CUDA SDPA crash remains (not addressed here)