Spaces:

---

SushantGautam

/

SoccerChat

like 0

Runtime error

Requesting a T4 GPU and persistent storage for an academic project Space:

Project: SoccerChat — Vision-Language Chat for Soccer Clips
https://arxiv.org/html/2505.16630v1

Model (LoRA adapter): https://huggingface.co/SimulaMet/SoccerChat-qwen2-vl-7b
Base model: Qwen/Qwen2-VL-7B-Instruct (loaded via swift app, quantized 4-bit NF4)
Dataset: https://huggingface.co/datasets/SimulaMet/SoccerChat

What it does
SoccerChat lets users drop in short soccer video snippets and ask tactical/analytic questions (e.g., “Who receives the through ball?” “Why was play halted?”). It runs a lightweight Gradio/Swift UI with streaming responses.

T4 is sufficient
We quantize to 4-bit (NF4) with double quant; VRAM fits on a T4 16 GB.
Inference constraints: max_batch_size=1, VIDEO_MAX_PIXELS=100352, FPS_MIN_FRAMES=FPS_MAX_FRAMES=24.
Expected VRAM usage ~12–14 GB; tested on T4/V100 locally with these settings.

Resources requested
GPU: 1× T4 (16 GB)
CPU/RAM: Standard T4 tier is fine
Max session length: less is fine

Tech stack
Runtime: swift app (ms-swift) over the base model with LoRA adapters
Attention: SDPA; quant: bitsandbytes 4-bit NF4
UI: Gradio with streaming

Startup command (inside Space/Docker):
swift app
--adapters "SimulaMet/SoccerChat-qwen2-vl-7b"
--model "Qwen/Qwen2-VL-7B-Instruct"
--use_hf true
--attn_impl sdpa
--quant_method bnb --quant_bits 4
--bnb_4bit_quant_type nf4
--bnb_4bit_use_double_quant true
--bnb_4bit_compute_dtype float16
--max_batch_size 1
--is_multimodal true
--studio_title "SoccerChat"
--stream true
--server_name 0.0.0.0 --server_port 7860 --share false --lang en

Academic context
This is an academic research demo (non-commercial), showcasing multimodal reasoning on soccer content and supporting community benchmarking and teaching.

Safety & content
We’ll keep clips short, avoid copyrighted full-match uploads, and provide a basic usage note in the Space README.

Impact
A T4 will make the demo reliably usable for students and reviewers; CPU fallback is too slow for video-VL.

Thanks a lot for considering! 🙏

—Sushant (SimulaMet)