[coderfeli]

Mono decode runners retain JIT launchers after compile-only preparation, so every step still enters FlyDSL's JIT dispatch. Route execution through AITER's existing _run_compiled: the first real warmup launch compiles and executes once, then subsequent calls use its retained CompiledFunction with fresh runtime addresses, scalars and stream.

Reuse aiter.ops.flydsl.kernels.tensor_shim directly, without adding a local cache class or changing atom/mono/runtime/compile.py. Keep the original JIT objects for ABI and binary-resource inspection. DeepSeek V4's deferred compile hints stay unchanged. Kimi debug launchers are retained by build key so they also reuse compiled dispatch.

This covers DeepSeek V4/V4.1, GLM-5.2, Kimi-K3, MiniMax M3 sparse/dense paths and step-begin. The stream is passed positionally, matching AITER and CompiledFunction. Warm-cache compile-only handling comes from current AITER main (ROCm/aiter#6946, commit 2bafed472).

Validation:

• 280 existing runtime, plan and model regression tests passed. No test files are added or changed.
• Local GPU validation passed for no-dispatch preload, single warmup execution, changed addresses/scalars/streams, warm compile-only, JIT bypass, and graph capture/replay.
• Ruff, Black and git diff --check passed; the vendored GLM op retains its existing formatter exemption.

Launch overhead was measured in eager mode with CUDA Graph disabled (no capture or replay), using a warmed no-op FlyDSL launcher and the shared AITER dispatch path.

Dispatch path	Median host launch overhead (us/launch)
JIT dispatch via JitFunction	15.31
Cached dispatch via AITER _run_compiled	2.75

Cached dispatch saves 12.56 us per launch (82.0% lower host launch overhead).

Method: 40 alternating batches of 100 warm launches per path; GPU synchronization occurs outside the timed batches. Compilation and warmup are excluded. These figures measure steady-state host dispatch overhead; full-model latency and serving throughput were not measured.

Validation used the local FlyDSL MLIR build and a clean source snapshot of AITER main at 2bafed472. Full-model accuracy and serving throughput comparisons were not run locally.

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2741 --add-label <label>