Agentic/vibe coding configuration? Particularly - context window

[haidij]

Does anyone have any guidance on the correct size of context window to set in my coding assistant?

I’m using Roo Code in VSCode with Qwen3-Coder 30B 4 bit quantised version and llama.cpp. When setting up the model to be used in Roo Code you can set a context window maximum size. Qwen3-Coder supports 256k, I initially set it to 70k but was getting occasional crashes. I’m now using 20k, which seems to work, but it would be good to understand how to size this optimally. If anyone can point me to a relevant resource or share any rules of thumb on this, it would be appreciated.

My llama.cpp settings are -c 0 -b 4096 -ub 4096 -ngl 999 --flash-attn on --no-mmap.

[raphael.amorim]

Best scenario for “vibe coding” is to use vLLM or SGLang with 128k context-window max, primarily because of the concurrency support, which is what code assistants will rely on.

I would definitely would use the great work of @eugr, that already works for 1 -2 sparks: GitHub - eugr/spark-vllm-docker: Docker configuration for running VLLM on dual DGX Sparks

Please check also this discussion to get more background:

Run VLLM in Spark

Run VLLM in Spark Install uv curl -LsSf https://astral.sh/uv/install.sh | sh Create environment uv venv .vllm --python 3.12 source .vllm/bin/activate Install Pytorch uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130 Install flashinfer and triton uv pip install xgrammar triton flashinfer-python --prerelease=allow git clone --recursive https://github.com/vllm-project/vllm.git cd vllm python3 use_existing_torch.py uv pip install -r requiremen…

[christopher_owen]

I use 128k, personally. I think it will depend on how many files you load and their sizes during your workflow. I found that the orchestrator mode in Roo helped a bunch with context window management.

[eugr]

I use 128K window too, even if I can fit more.