Fix: use_cache=True, greedy decoding, model.eval() — was producing 0-char output

Switch base image to python:3.11-slim for fresh Docker image build

Force new commit SHA with clean FP16 app.py

Upload Dockerfile with huggingface_hub

Remove load_in_8bit flag to use pure FP16 model loading

Upload requirements.txt with huggingface_hub

Use pure FP16 precision (~5.8 GB VRAM) for clean loading on Nvidia T4 GPU

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Upload requirements.txt with huggingface_hub

Use native float16 precision for Nvidia T4 GPU (Turing architecture)

Use standard instruction prompt and robust extraction to prevent 0-byte outputs

Fix prompt truncation to max 1200 tokens and add decoding fallback

Set pad_token_id and return clean JSON error messages

Add detailed error reporting inside summarize endpoint

Upload app.py with huggingface_hub

Upload Dockerfile with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload README.md with huggingface_hub

Pin transformers==4.45.2 and tokenizers==0.20.0 for PyPreTokenizerTypeWrapper compatibility

Use official BharatGen snippet: tokenizer trust_remote_code=False & torch_dtype=bfloat16

Use PreTrainedTokenizerFast(tokenizer_file=tokenizer.json) directly

Use LlamaTokenizer directly to force pure Python tokenizer loading

Set use_fast=False for AutoTokenizer in app.py

Upload clean FastAPI app.py for transformers 4.38.2

Pin exact transformers==4.38.2 and tokenizers==0.15.2 wheel pair

Upload README.md with huggingface_hub

Upload requirements.txt with huggingface_hub

Upload app.py with huggingface_hub

Pin transformers==4.42.4 and accelerate==0.31.0

Set use_safetensors=True for sharded model weights loading

Add safetensors>=0.4.3 to requirements.txt

Pin transformers==4.40.2 matching model config.json transformers_version

Use use_fast=False for AutoTokenizer

Pin transformers==4.35.2 with sentencepiece & fast-api

Convert app.py to FastAPI high-performance REST service

Add fastapi, uvicorn, pydantic to requirements.txt

Upload Gradio 3.50.2 app.py

Set sdk_version to 3.50.2 to fix Gradio 4 localhost check & jinja2 template dict error

Pin transformers==4.40.2 to fix PyPreTokenizerTypeWrapper enum parsing in tokenizer.json

Use gr.Blocks with explicit api_name=summarize_legal_case and share=True

Use default demo.launch() for HuggingFace Space environment

Add server_name=0.0.0.0 and server_port=7860 to demo.launch for HF Space docker container

Add api_name=summarize_legal_case to gr.Interface

Upload README.md with huggingface_hub

Pin transformers==4.35.2 for compatibility with ParamBharatGen custom model code

Set use_cache=False to completely bypass DynamicCache indexing in custom model code

Add return_legacy_cache=True to fix DynamicCache tuple indexing in custom model code

Upload clean scratch/app.py without shell string escaping

Fix syntax error in app.py