feat(domino): add Qwen3.5-4B managed-local NPU disaggregated recipe

examples/README.md

@@ -38,6 +38,7 @@ NPU, offline, and managed/external-service variants, is in
 | `examples/configs/qwen3.5-4b-dflash-online-npu.yaml` | Disaggregated NPU SGLang capture | DFlash |
 | `examples/configs/qwen3.5-4b-dflash-disaggregated-npu.yaml` | Managed local NPU full stack | DFlash |
 | `examples/configs/qwen3.5-4b-domino-online-npu.yaml` | Disaggregated NPU SGLang capture | Domino |
+| `examples/configs/qwen3.5-4b-domino-disaggregated-npu.yaml` | Managed local NPU full stack | Domino |
 
 Online configs point `data.train_data_path` at raw conversation data. Offline
 configs expect strategy-specific feature checkpoints in

examples/configs/qwen3.5-4b-domino-disaggregated-npu.yaml

@@ -0,0 +1,62 @@
+# Single-node managed full stack for Ascend NPU: one `specforge train` command
+# owns Mooncake, one patched SGLang capture server (device 0), and a 14-rank
+# trainer (devices 2-15). Device ordinals are injected via
+# ASCEND_RT_VISIBLE_DEVICES on Ascend hosts.
+model:
+  target_model_path: "Qwen/Qwen3.5-4B"
+  draft_model_config: "configs/qwen3.5-4b-domino.json"
+  target_backend: sglang
+  trust_remote_code: true
+  embedding_key: "model.language_model.embed_tokens.weight"
+  torch_dtype: "bfloat16"
+  sglang_attention_backend: ascend
+data:
+  train_data_path: "./cache/dataset/train_regen.jsonl"
+  max_length: 1024
+  chat_template: "qwen3.5"
+  build_dataset_num_proc: 32
+  cache_dir: "./cache"
+training:
+  strategy: "domino"
+  num_epochs: 10
+  batch_size: 1
+  accumulation_steps: 4
+  learning_rate: 0.0006
+  warmup_ratio: 0.04
+  max_grad_norm: 1
+  attention_backend: "sdpa"
+  num_anchors: 16
+  loss_decay_gamma: 7
+  lambda_base_start: 1
+  lambda_base_decay_ratio: 1
+  save_interval: 3000
+  log_interval: 50
+  dist_timeout: 30
+  seed: 42
+tracking:
+  report_to: "tensorboard"
+run_id: "qwen3.5-4b-domino-npu-managed"
+output_dir: "./outputs/qwen3.5-4b-domino-npu-managed"
+
+deployment:
+  mode: disaggregated
+  trainer:
+    nnodes: 1
+    nproc_per_node: 14
+  disaggregated:
+    control_dir: outputs/qwen3.5-4b-domino-npu-managed/control
+    consumer_state_dir: outputs/qwen3.5-4b-domino-npu-managed/consumer-state
+    backend: mooncake
+    managed_local:
+      # 16-card A3 layout: capture server owns device 0, trainer owns 2-15.
+      # On 8-card hosts use devices 1-7 with trainer.nproc_per_node: 7.
+      trainer_cuda_visible_devices:
+        ["2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15"]
+      mooncake:
+        # Host segment mounted by the capture-server sink for feature storage.
+        # 16 GiB fits a 64GB-card host next to the target model and KV cache.
+        global_segment_size_bytes: 17179869184
+      capture_servers:
+        - port: 30000
+          cuda_visible_devices: ["0"]
+          tp_size: 1

tests/test_config/test_launch_topology.py

@@ -68,6 +68,7 @@
     "qwen3.5-35b-a3b-eagle3-online.yaml": 2,
     "qwen3.5-4b-dflash-online-npu.yaml": 8,
     "qwen3.5-4b-dflash-disaggregated-npu.yaml": 14,
+    "qwen3.5-4b-domino-disaggregated-npu.yaml": 14,
     "qwen3.5-4b-domino-online-npu.yaml": 8,
     "qwen3.6-27b-dflash-disaggregated.yaml": 2,
     "qwen3.6-27b-dflash-1server-dp2-disaggregated.yaml": 2,
@@ -261,6 +262,39 @@
             ],
         },
     },
+    "qwen3.5-4b-domino-disaggregated-npu.yaml": {
+        "control_dir": "outputs/qwen3.5-4b-domino-npu-managed/control",
+        "consumer_state_dir": "outputs/qwen3.5-4b-domino-npu-managed/consumer-state",
+        "backend": "mooncake",
+        "managed_local": {
+            "trainer_cuda_visible_devices": [
+                "2",
+                "3",
+                "4",
+                "5",
+                "6",
+                "7",
+                "8",
+                "9",
+                "10",
+                "11",
+                "12",
+                "13",
+                "14",
+                "15",
+            ],
+            "mooncake": {
+                "global_segment_size_bytes": 17179869184,
+            },
+            "capture_servers": [
+                {
+                    "port": 30000,
+                    "cuda_visible_devices": ["0"],
+                    "tp_size": 1,
+                }
+            ],
+        },
+    },
     "qwen3.6-27b-dflash-disaggregated.yaml": {
         "control_dir": "outputs/qwen3.6-27b-dflash-disaggregated/control",
         "backend": "mooncake",
@@ -347,7 +381,7 @@ def _recipes() -> dict[str, Path]:
 class ExampleLaunchTopologyTest(unittest.TestCase):
     def test_every_recipe_has_the_explicit_golden_topology(self):
         recipes = _recipes()
-        self.assertEqual(len(EXPECTED_NPROC_PER_NODE), 65)
+        self.assertEqual(len(EXPECTED_NPROC_PER_NODE), 66)
         self.assertEqual(set(recipes), set(EXPECTED_NPROC_PER_NODE))
 
         for filename, nproc_per_node in EXPECTED_NPROC_PER_NODE.items():

tests/test_config/test_unified_feature_reachability.py

@@ -146,7 +146,7 @@ def test_all_example_configs_validate_through_the_typed_entry(self):
             for path in EXAMPLE_CONFIG_DIR.glob("*.yaml")
             if not path.name.startswith(".")
         )
-        self.assertEqual(len(paths), 65)
+        self.assertEqual(len(paths), 66)
 
         resolved_runs = {
             path.name: resolve_run(Config.from_file(str(path))) for path in paths