Fix Mammoth’s CLI selecting the wrong engine

serve.patch

@@ -783,3 +783,22 @@ index 4f6f7bd..ae428d9 100644
              try:
                  opt.step()
              except torch.OutOfMemoryError:
+diff --git a/qev/cli.py b/qev/cli.py
+--- a/qev/cli.py
++++ b/qev/cli.py
+@@ -115,7 +115,14 @@
+     if calib_path:
+         logging.getLogger("qev.cli").info("calibration: %s", calib_path)
+     config = json.loads((Path(model_path) / "config.json").read_text())
+-    if args.multimodal or config.get("vision_config") is not None:
++    from .readout import ReadoutConfig
++
++    # Only the patched text engine implements block-bidirectional attention.
++    noncausal = ReadoutConfig.load(model_path).noncausal
++    if args.multimodal and noncausal:
++        raise SystemExit("Block-bidirectional attention is unsupported by the multimodal engine; "
++                         "omit --multimodal to serve text requests with DecisionEngine.")
++    if args.multimodal or (config.get("vision_config") is not None and not noncausal):
+         from .mm_engine import MMDecisionEngine
+ 
+         engine = MMDecisionEngine(model_path, device=args.device, dtype=args.dtype, calibration=calib,

test_cli_routing.py

@@ -0,0 +1,68 @@
+"""CPU-only routing regression: python test_cli_routing.py /path/to/patched/OneJev."""
+import argparse
+import importlib.util
+import contextlib
+import json
+from pathlib import Path
+import sys
+import tempfile
+import types
+import unittest
+from unittest.mock import patch
+
+ROOT = Path(sys.argv.pop(1))
+
+class RoutingTest(unittest.TestCase):
+    def route(self, config, readout=None, multimodal=False):
+        # Import the actual readout module and execute the CLI; stub GPU/server boundaries.
+        called = []
+        def engine(kind):
+            return lambda *a, **kw: called.append(kind) or object()
+        mods = {}
+        for name, attrs in {
+            'uvicorn': {'run': lambda *a, **kw: None},
+            'qev': {}, 'qev.calibrate': {'Calibration': object},
+            'qev.engine': {'DecisionEngine': engine('text')},
+            'qev.mm_engine': {'MMDecisionEngine': engine('multimodal'), 'local_model_dir': lambda x: x},
+            'torch': {'nn': types.SimpleNamespace(Module=object)},
+            'qev.prompt': {'MAX_SLOTS': 256},
+            'qev.server': {'create_app': lambda *a, **kw: object()},
+        }.items():
+            mods[name] = types.ModuleType(name)
+            mods[name].__dict__.update(attrs)
+        with tempfile.TemporaryDirectory() as d, patch.dict(sys.modules, mods), patch('importlib.util.find_spec', return_value=object()):
+            Path(d, 'config.json').write_text(json.dumps(config))
+            if readout is not None:
+                Path(d, 'readout.json').write_text(json.dumps(readout))
+            spec = importlib.util.spec_from_file_location('qev.readout', ROOT / 'qev/readout.py')
+            module = importlib.util.module_from_spec(spec)
+            sys.modules['qev.readout'] = module
+            spec.loader.exec_module(module)
+            cli = {'__name__': 'qev.cli', '__package__': 'qev'}
+            exec(compile((ROOT / 'qev/cli.py').read_text(), 'cli.py', 'exec'), cli)
+            args = argparse.Namespace(model=d, calibration=None, multimodal=multimodal,
+                gguf=None, llama_url=None, device='cpu', dtype='bfloat16', fork_mode='sequential',
+                head_dtype='float32', media_root=None, no_gpu_preprocess=False, no_cuda_graphs=False,
+                quantize=None, max_branch_tokens=16384, max_request_tokens=65536,
+                name=None, host='127.0.0.1', port=8000)
+            cli['cmd_serve'](args)
+        return called
+
+    def test_routes(self):
+        layout = {'vision_config': {}, 'modeling_overrides': {'attention': {'full_attention': 'block_bidirectional'}}}
+        for config, readout, explicit, expected in [
+            (layout, None, False, 'text'), ({}, None, False, 'text'),
+            ({'vision_config': {}}, None, False, 'multimodal'),
+            ({'vision_config': {}}, None, True, 'multimodal'),
+            ({'vision_config': {}}, {'attention': 'noncausal'}, False, 'text'),
+            ({'vision_config': {}}, {'readout': 'head'}, False, 'multimodal'),
+        ]:
+            with self.subTest(config=config, readout=readout, explicit=explicit):
+                self.assertEqual(self.route(config, readout, explicit), [expected])
+        with self.assertRaisesRegex(SystemExit, 'omit --multimodal'):
+            self.route(layout, multimodal=True)
+        with self.assertRaisesRegex(ValueError, 'unsupported full_attention'):
+            self.route({'modeling_overrides': {'attention': {'full_attention': 'unknown'}}})
+
+if __name__ == '__main__':
+    unittest.main()