ci: isolate MLX checks on self-hosted runners

Author: apstenku123

.github/workflows/ci-self-hosted.yml

@@ -39,7 +39,7 @@ jobs:
           git diff --check
 
   linux-portable:
-    name: Linux portable conveyor contracts
+    name: Linux portable syntax and conveyor contracts
     runs-on: [self-hosted, Linux, X64, cppmega-mlx]
     timeout-minutes: 20
     steps:
@@ -53,10 +53,13 @@ jobs:
         shell: bash
         run: |
           set -euo pipefail
-          python -m pytest -q \
+          # cppmega.mlx imports Apple's MLX runtime at package import time. The
+          # Linux runner exercises only portable scripts and deliberately skips
+          # the macOS-only root conftest/package initialization.
+          python -m pytest -q --noconftest \
             tests/test_audit_sidecar_parquet.py \
-            tests/test_pack_enriched_rows.py \
             tests/test_streaming_conveyor_progress.py \
             tests/test_process_commits_fail_loud.py \
             tests/test_repair_packed_document_boundaries.py
+          python -m compileall -q scripts tools/clang_indexer
           git diff --check

.github/workflows/cuda-lane.yml

@@ -1,10 +1,9 @@
 name: CUDA Lane
 
 on:
-  push:
-    branches: [main]
-  pull_request:
-    branches: [main]
+  # CUDA capacity is provisioned explicitly (for example on Nebius H200).
+  # Do not leave ordinary PRs queued forever when no self-hosted CUDA runner
+  # is attached to this repository.
   workflow_dispatch:
 
 jobs:

cppmega_mlx/nn/mamba3.py

@@ -523,7 +523,6 @@ def _dispatch_mamba3_scan(
     lowered TileLang DSL fwd+bwd kernel and fails closed if unavailable.
     """
 
-    from cppmega_mlx.nn._tilelang.mamba3 import mamba3_mimo_metal_status
     from cppmega_mlx.runtime.kernel_policy import (
         KernelPath,
         record_dispatch,
@@ -576,6 +575,8 @@ def _dispatch_mamba3_scan(
         )
 
     # AUTO + PATH_B share the Metal-availability check.
+    from cppmega_mlx.nn._tilelang.mamba3 import mamba3_mimo_metal_status
+
     status = mamba3_mimo_metal_status(x)
     if path is KernelPath.PATH_B and not status.available:
         # Device-aware Path B: on a CUDA host (Metal unavailable, default

tests/test_mamba3_dispatch.py

@@ -9,6 +9,8 @@
 
 from __future__ import annotations
 
+import sys
+
 import numpy as np
 import pytest
 
@@ -85,6 +87,20 @@ def test_reference_policy_forces_pure_mlx(monkeypatch: pytest.MonkeyPatch) -> No
     assert matches[-1]["kernel_used"] == "reference_pure_mlx"
 
 
+def test_reference_policy_does_not_import_tilelang_dispatch(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("CPPMEGA_KERNEL_PATH", "ref")
+    monkeypatch.setitem(sys.modules, "cppmega_mlx.nn._tilelang.mamba3", None)
+
+    block, hidden = _make_block()
+    out, _ = block(hidden)
+    mx.eval(out)
+
+    matches = [e for e in get_dispatch_log() if e["op_name"] == "mamba3_mimo"]
+    assert matches[-1]["kernel_used"] == "reference_pure_mlx"
+
+
 def test_path_b_policy_forces_metal(monkeypatch: pytest.MonkeyPatch) -> None:
     if not _METAL_AVAILABLE:
         pytest.skip("Metal not available")

tests/test_megatron_indexed.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import os
 import subprocess
 import sys
 import struct
@@ -255,9 +256,15 @@ def _write_structured_multishard_fixture(
 
 
 def _run_train_hybrid_tiny(*args: str) -> subprocess.CompletedProcess[str]:
+    env = os.environ.copy()
+    # This suite validates indexed-dataset ingress, not optional kernel discovery.
+    # An explicit reference route keeps the subprocess independent of any shared
+    # TileLang/TVM development checkout configured by the parent pytest process.
+    env["CPPMEGA_KERNEL_PATH"] = "ref"
     return subprocess.run(
         [sys.executable, str(TRAIN_HYBRID_TINY), *args],
         cwd=ROOT,
+        env=env,
         text=True,
         capture_output=True,
         timeout=45,