Upgrade to Llama 3.1 8B-Instruct for better long-form content

Fix build error: Remove flash-attn dependency

Speed optimizations: Switch to Mistral-7B + optimize generation params

Major update: Add NFL training data generation and improve model handling

Fix syntax errors: correct comma placement and indentation

Force all CUDA operations to cuda:0 and use device_map to prevent multi-GPU distribution

Fix multi-GPU device placement error: disable device_map auto and ensure tensors on same device

Fix duplicate import in app.py

Switch to Llama-3.1-8B-Instruct: update model loading, prompts, and generation parameters

Switch to FLAN-T5-Large: uses standard HF storage, excellent for question generation

Fix XetHub issue: use Meta-Llama-3.1-8B-Instruct (official HF storage)

Fix download errors and warnings: retry logic + clean startup logs

Fix version conflict: use PyTorch 2.5.0 + TorchVision 0.20.0 exact match

Use PyTorch 2.5.1 with safetensors to avoid CVE-2025-32434 - practical fix

Fix PyTorch CVE-2025-32434: upgrade to v2.6+, use safetensors, restore Llama 3.1

Deploy temporary working model: DialoGPT while awaiting Llama 3.1 access

Fix permissions error: proper cache directory and HF token auth for Llama 3.1

Major optimization: Replace llama-cpp-python with transformers to eliminate compilation

Simplify build: use CPU llama-cpp-python for reliability

Initial setup: Question Generation API with DeepHermes reasoning model