See Qwen3 Decode as a CUDA Graph

OpenInfer can now export the live rank-0, batch-1 SplitKv decode CUDA Graph at startup. A single flag writes an unfolded .dot sidecar with full kernel symbols and launch metadata for LLMs or scripts, plus a folded 192-DPI .png for people.

cargo run --release -- \
  --model-path models/Qwen3-4B \
  --dump-graph-png qwen3-decode.png

The Qwen3-4B example below contains 507 kernels and 506 dependency edges, with its 14-kernel layer body collapsed as a ×36 repeated block.

Qwen3-4B batch-1 decode CUDA Graph with 36 repeated layers folded into one block

Qwen3-4B decode · batch 1 · SplitKv · 507 kernels / 506 edges