NVIDIA Warp (APIC): Triton deploy runtime [2/2, stacked on export]

[lgulich]

Stacked MR — 2 of 2: Triton deploy runtime. Built on top of the Warp-export MR (nvidia-isaac/leapp#3). Review nvidia-isaac#3 first.

This is a fork-internal PR whose base is the export branch (lgulich/warp-node-prototype), so the diff below is exactly this MR's net-new change — the leapp_runtimes/ Triton runtime + its tests/example. It will be retargeted to nvidia-isaac/leapp:main once nvidia-isaac#3 merges.

Summary

Adds the Triton deploy runtime for Warp .wrp nodes, on top of the Warp export work in nvidia-isaac#3. nvidia-isaac#3 captures a mixed PyTorch + Warp graph and emits per-node artifacts (torch → ONNX/JIT, warp → native APIC .wrp) runnable via the in-process Python InferenceManager. This MR adds the path to deploy those same artifacts under Triton.

What it adds (leapp_runtimes/)

• triton/create_triton_model_repo.py — turns a LEAPP YAML + per-node artifacts into a Triton model repository (ONNX/JIT + warp branch; emits a python-backend model dir per warp node and an ensemble wiring the data_flow edges).
• triton/warp_node/warp_apic_runtime.py — WarpApicRunner: loads a warp node's .wrp, binds inputs via wp.from_torch (using the _in_/_out_ APIC binding names from the sidecar), replays the captured CUDA graph with capture_launch, reads outputs back — the deploy-side analogue of the export backend's _WarpGraphCallable.
• triton/warp_node/model.py — the Triton python-backend model template copied verbatim into each generated warp model dir.
• Tests: test_triton_model_repo.py (generator: ONNX-only, warp-only, onnx→warp→onnx ensemble) and test_triton_warp_node.py (live (py)triton replay — skips cleanly when pytriton is absent). Example helper examples/triton_warp_node/make_warp_model_repo.py.

The runtime is self-contained (it imports nothing from leapp/; it reads the YAML + .warpmeta.json the export side writes), so it sits cleanly on top of the export MR.

Test plan

• test_triton_model_repo.py + test_triton_warp_node.py — 10 passed, 1 skipped (the skip is the live (py)triton-server tier, absent in this env).
• The export MR's test_warp_autocapture.py validates the captured .wrp via the export backend (WarpExportBackend), so NVIDIA Warp (APIC) export: non-invasive auto-split of mixed PyTorch + Warp graphs [1/2] nvidia-isaac/leapp#3 stands alone; this MR's WarpApicRunner is the deploy-side runner.

Deliberately out of scope (left for a follow-up)

Consolidating both runtimes into one leapp_runtime*/ package is intentionally NOT done here. Today InferenceManager lives at leapp/inference_manager.py (the existing in-process runtime, on main) and the Triton runtime lives at leapp_runtimes/. Unifying them would move the public from leapp import InferenceManager import path — a repo-wide, downstream-visible change — so it's deferred to a separate, focused refactor MR rather than bundled into this deploy-runtime MR.

🤖 Generated with Claude Code