Merge remote-tracking branch 'origin/master' into ralph/proxy-stream-cancellation

tests/test_tracker_routing.py

@@ -2874,6 +2874,43 @@ def test_same_endpoint_can_register_multiple_models():
         tracker.stop()
 
 
+def test_explicit_model_placement_targets_only_the_selected_node():
+    "An admin can add and release a model on one chosen multi-model node.\n\nTags: http, routing, tracker"
+    from meshnet_tracker.server import _release_model_locked, _request_model_load_locked
+
+    tracker = _tracker(model_presets={
+        "model-a": {"total_layers": 4, "bytes_per_layer": {"bfloat16": 1_000}, "hf_repo": "org/ModelA"},
+        "model-b": {"total_layers": 4, "bytes_per_layer": {"bfloat16": 1_000}, "hf_repo": "org/ModelB"},
+    })
+    tracker_port = tracker.start()
+    try:
+        registrations = []
+        for port in (9062, 9063):
+            registrations.append(_post_json(
+                f"http://127.0.0.1:{tracker_port}/v1/nodes/register",
+                {"endpoint": f"http://127.0.0.1:{port}", "model": "model-a", "hf_repo": "org/ModelA",
+                 "num_layers": 4, "shard_start": 0, "shard_end": 3, "max_loaded_shards": 2,
+                 "vram_bytes": 20_000, "ram_bytes": 20_000, "quantizations": ["bfloat16"],
+                 "benchmark_tokens_per_sec": 1.0, "hardware_profile": {}, "score": 1.0},
+            ))
+        selected, other = (item["node_id"] for item in registrations)
+
+        with tracker._lock:
+            assignment = _request_model_load_locked(tracker._server, "model-b", selected)  # type: ignore[arg-type]
+        assert assignment is not None
+        assert assignment["node_id"] == selected
+        assert tracker._registry[selected].pending_new_assignment is not None
+        assert tracker._registry[other].pending_new_assignment is None
+
+        with tracker._lock:
+            released = _release_model_locked(tracker._server, "model-a", selected)  # type: ignore[arg-type]
+        assert released == 1
+        assert len(tracker._registry[selected].pending_directives) == 2
+        assert tracker._registry[other].pending_directives == []
+    finally:
+        tracker.stop()
+
+
 def test_scale_demanded_models_queues_add_shard_on_spare_host():
     "Scale demanded models queues add shard on spare host\n\nTags: http, routing, tracker"
     tracker = _tracker(model_presets={
@@ -3010,6 +3047,13 @@ def test_a_node_declaring_an_unsupported_quantization_is_never_routed():
     assert status == 503
 
 
+def test_a_node_declaring_auto_quantization_serves_a_default_precision_request():
+    "'auto' is the CLI default that delegates the choice — it is not a refusal, so the node must resolve to its best advertised precision and route.\n\nTags: http, routing, tracker"
+    status, response = _proxy_chat_status(POLICY_COMPAT, quantization="auto")
+    assert status == 200
+    assert response["choices"][0]["message"]["content"] == "ok"
+
+
 def test_a_node_declaring_a_null_quantization_is_never_routed():
     "An explicit null states 'no usable precision' -- only an absent field is legacy.\n\nTags: http, routing, tracker"
     node = http.server.HTTPServer(("127.0.0.1", 0), _EchoChatHandler)