Optimize speculative decoding performance by increasing max_num_batched_tokens to 4096

Fix entrypoint override in Dockerfile

Fix HF space build error by using official vllm-openai:v0.21.0 base image

Remove torchvision pin to allow vllm 0.12.0 to resolve its own torchvision dependency

Upgrade to vllm 0.12.0 and restore original performance arguments for Qwen 3.5 support

Pin transformers to 4.48.2 to satisfy vllm>=4.48.2 dependency and resolve TokenizersBackend AttributeError

Pin transformers to 4.48.0 to resolve TokenizersBackend AttributeError in vLLM 0.8.0

Remove limit-mm-per-prompt argument to bypass vLLM 0.8.0 multimodal registration check

Remove speculative-config argument for vLLM 0.8.0 CLI parser

Fix limit-mm-per-prompt syntax for vLLM 0.8.0 CLI parser

Update torch to 2.6.0+cu124 and torchvision to 0.21.0+cu124 to satisfy vllm 0.8.0 dependency requirements

Downgrade torch to 2.5.1+cu124 and vllm to 0.8.0 to match available CUDA 12.4 pre-compiled wheels

Fix NVIDIA driver mismatch on HF Space by forcing +cu124 torch wheels

Fix: Upgrade to vLLM 0.22.0 and PyTorch 2.11.0 on CUDA 12.4 for native Qwen 3.5 support and host compatibility

Fix: Remove speculative-config argument for vLLM 0.7.2 CLI compliance

Fix: Update --limit-mm-per-prompt format to KEY=VALUE format image=99 for vLLM 0.7.2

Fix: Align vLLM arguments with model card recommendations (chat format, MTP speculative config, limit-mm-per-prompt)

Fix: Change wheel index to cu124 and pin torch to 2.5.1+cu124 for native CUDA 12.4 host driver compatibility

Fix: Remove manual transformers source install to resolve tokenizer AttributeError

Fix: Remove limit-mm-per-prompt and mm-processor-kwargs for vLLM 0.7.2 compatibility

Fix: Update --limit-mm-per-prompt format to KEY=VALUE for vLLM 0.7.2 compatibility

Fix: Pin torch to 2.5.1+cu121 and vllm to 0.7.2 to guarantee CUDA 12.4 host driver compatibility

Fix: Add LD_LIBRARY_PATH forward compatibility for older host GPU drivers

Fix: Switch base to CUDA devel image to provide nvcc for flashinfer JIT, and add .gitignore

Fix: Remove show_copy_button parameter from gr.Textbox for Gradio 6 compatibility

Fix: Remove audioop-lts as Python 3.10 natively provides audioop

Migration: Convert Hugging Face Space to custom Docker Space using CUDA 12.4

Fix: Install transformers from source to support Qwen 3.5 architecture

Initial commit: NuExtract3 Gradio space setup powered by vLLM on A100 GPU