amd-agi/geak/ensure-flydsl

Ensure FlyDSL is importable (single source of truth for the FlyDSL build)

This skill makes import flydsl, kernels.moe_gemm_2stage resolve from ONE consistent tree, so the FlyDSL optimization skills (flydsl_rewrite_quantized_moe for authoring, apply_flydsl_moe_to_vllm for the e2e apply-back) can run. It is a dependency skill: other skills / the System Architect invoke it; it is not selected from the profile Top-N on its own.

When to run

Primary: at Strategize, the moment a FlyDSL expert skill matches — the System Architect runs this skill BLOCKING right then (see roles/system_architect.md §0a), before routing flydsl, and updates the environment (source flydsl_env.sh; add flydsl to available_backends). So flydsl is fully built and importable before any downstream author/apply. Also safe to invoke as an idempotent step 0 of the flydsl author/apply skills as a belt-and-suspenders guard — later calls short-circuit (version-gate reuse), so it never rebuilds.

Procedure

Run the bundled executor (do NOT hand-roll clone/build — it carries the version gate, the ROCm-image fixes, the concurrency lock, and the env file). A from-scratch build is ~15–18 min — longer than the Bash tool's 600 s ceiling — so launch it DETACHED and poll, never as one foreground call:

FR="${FLYDSL_ROOT:-/opt/flydsl/FlyDSL}"; ENV="$FR/flydsl_env.sh"; FAIL="$(dirname "$FR")/.flydsl_build.failed"
SH="perf_knowledge/expert_skills/skills/ensure_flydsl/ensure_flydsl.sh"
nohup bash "$SH" > "$(dirname "$FR")/ensure_flydsl.log" 2>&1 &   # detached
for i in $(seq 1 180); do                                        # deadline 1800s (30 min)
  [ -f "$ENV" ]  && break                                        # success: flydsl_env.sh written
  [ -f "$FAIL" ] && { echo "flydsl build FAILED (see ensure_flydsl.log)"; break; }
  sleep 10
done
[ -f "$ENV" ] && source "$ENV" && \
  python3 -c "import flydsl, kernels.moe_gemm_2stage as k; print(flydsl.__file__, k.__file__)"  # ONE tree

Completion markers: success = `flydsl_env.sh` exists; failure = `.flydsl_build.failed` exists (next to $FLYDSL_ROOT). Idempotent: if flydsl is already present the script writes flydsl_env.sh almost immediately (version-gate reuse), so the poll returns in one tick.

What the executor guarantees (ensure_flydsl.sh)

• Version-only reuse gate: if an ambient flydsl+kernels is importable with `__version__ >=

FLYDSL_MIN_VERSION (default 0.2.2`), REUSE it — never overwrites a newer flydsl, never pip-installs system-wide (portable / does not disturb other workloads). The pinned commit is NOT used to gate reuse.

• Else clone + build the PIN (a35627a2…) into container-internal /opt/flydsl (overlay fs, not a

bind-mounted host dir), guarded by an flock so concurrent Setup/author retries can't corrupt one tree.

• ROCm-image compat fixes applied before the build: hip cmake symlink (cmake refs

libamdhip64.so.7.2.70201 but the runtime is 7.2.53211) and auto-install patchelf (build.sh's CopyFlyPythonSources needs it) — the two failures that otherwise kill the source build late.

• Writes `flydsl_env.sh` (FLYDSL_ROOT / PYTHONPATH / FLYDSL_SHIM_DIR / VLLM_USE_FLYDSL_MOE)

for the workflow + vLLM + subagents to source.

• Exit 0 ⇒ flydsl importable + env written. Non-zero ⇒ install failed (loud); the calling skill must

drop flydsl and record why (do not silently fall back to triton).

Overrides

• FLYDSL_ROOT — build/checkout location (default /opt/flydsl/FlyDSL).
• FLYDSL_MIN_VERSION — reuse floor (default 0.2.2).
• FLYDSL_BUILD_LOCK — lock file path.