Installation¶
sweepx.agent is part of sweepx. The install puts the sweep-agent command on
your PATH (also reachable as sweepx agent) and brings the tool registry, the
solver (sweep) and the production runner (sweepx.tasks), so natural-language
forward modelling, shot gathers and FWI / LSRTM all work out of the box —
acoustic and elastic (vp/vs/rho), including bundled benchmark models
(Marmousi, Overthrust). Python 3.10+.
Extras:
pip install "sweepx[agent-ui]" # Gradio web UI
pip install "sweepx[animate]" # GIF export
pip install "sweepx[segy]" # SEG-Y files
You also need a local LLM¶
Chat needs any OpenAI-compatible endpoint with a 32k-token context: the agent's prompt with its tool schemas is 9-25k tokens.
- Ollama (Mac / CPU):
OLLAMA_CONTEXT_LENGTH=32768 ollama serve, thenollama pull qwen2.5:14b, and runsweep-agent chat --model qwen2.5:14b(auto-detected on:11434). Without the context setting Ollama cuts the prompt short without a word. Tool-calling needs a capable model — the default Q4 quants are noticeably weaker; prefer a-q8_0tag, a larger model (qwen2.5:32b), or a full-precision endpoint. -
vLLM (GPU node): vLLM pins its own torch, so install it in an environment of its own; on NVIDIA drivers 570 / 575 use the CUDA 12 build:
pip install "vllm==0.10.2" --extra-index-url https://download.pytorch.org/whl/cu128 pip install "transformers>=4.45,<5.0" vllm serve Qwen/Qwen2.5-14B-Instruct --max-model-len 32768 \ --enable-auto-tool-choice --tool-call-parser hermesexamples/agent/start_vllm_qwen7b.shwraps this (7B by default,MODEL=Qwen/Qwen2.5-14B-Instructfor 14B). Full precision is the most reliable for tool-calling; 14B needs a 48 GB-class GPU.sweep-agent chatfinds the server on:8000/:8001and asks it which model it serves.
macOS (Apple Silicon)
Runs end-to-end on M-series with MPS (CPU 26.7 s → MPS 5.5 s on a
256×384 / 8-shot / 1500-step demo). Use Ollama for the LLM; run_forward_sweep
defaults to device="auto" (MPS → CPU). Do not set SWEEP_BUILD_CUDA —
that's the Linux + NVIDIA path; macOS uses sweep's eager torch.