[Draft][Reload] Add manifest-driven load receipts and update scopes

docs/serving/offline_inference.md

@@ -62,6 +62,7 @@ For further details on metrics, please refer to [this page](../design/metrics.md
 For further details on Weight Transfer, please refer to [this page](../training/weight_transfer/README.md).
 
 - `LLM.init_weight_transfer_engine` - Initializes the weight transfer engine for RL training.
+- `LLM.get_weight_update_manifest` - Lists updatable model weights and LoRA adapters.
 - `LLM.start_weight_update` - Starts a new weight update cycle.
 - `LLM.update_weights` - Updates the model weights.
 - `LLM.finish_weight_update` - Finishes the current weight update cycle.

docs/serving/online_serving/README.md

@@ -171,6 +171,7 @@ For further details on Weight Transfer, please refer to [this page](../../traini
 - `/resume` - Resume generation
 - `/is_paused` - Check if generation is paused
 - `/init_weight_transfer_engine` - Initialize weight transfer engine for RLHF
+- `/weight_update_manifest` - List updatable model weights and LoRA adapters
 - `/start_weight_update` - Prepares the inference engine for a weight update.
 - `/update_weights` - Update model weights (can alter model behavior)
 - `/finish_weight_update` - Finalizes the weight update

docs/training/weight_transfer/README.md

@@ -51,6 +51,7 @@ When running vLLM as an HTTP server, the following endpoints are available for w
 | Endpoint | Method | Description |
 | -------- | ------ | ----------- |
 | `/init_weight_transfer_engine` | POST | Initialize the weight transfer engine with backend-specific info |
+| `/weight_update_manifest` | GET | Get updatable model-weight and LoRA manifests |
 | `/start_weight_update` | POST | Start a weight update |
 | `/update_weights` | POST | Transfer a batch of weights with backend-specific metadata |
 | `/finish_weight_update` | POST | Finish the weight update and run post-processing |
@@ -61,6 +62,37 @@ When running vLLM as an HTTP server, the following endpoints are available for w
 !!! note
     The HTTP weight transfer endpoints require `VLLM_SERVER_DEV_MODE=1` to be set.
 
+### Discovering legal partial-update scopes
+
+Call `GET /weight_update_manifest` after model loading. The response includes
+separate `model_weights` and `lora_adapters` fields. `model_weights` contains
+all source names observed during the initial checkpoint load and
+`atomic_source_groups`. The corresponding `atomic_update_scopes` entries can be
+sent directly to `/start_weight_update`. A larger legal partial scope is a union
+of complete atomic groups:
+
+```json
+{
+  "kind": "base_checkpoint",
+  "mode": "partial",
+  "source_names": [
+    "model.layers.0.input_layernorm.weight",
+    "model.layers.0.post_attention_layernorm.weight"
+  ]
+}
+```
+
+The groups are merged across workers, so TP, PP, and EP closure constraints
+are preserved. Use the scope only when `model_weights.ready` is `true`.
+A dummy model without metadata probing returns a provisional baseline until it
+has completed one real full base-weight update.
+
+Each entry in `lora_adapters` includes adapter identity, generation, the union
+of rank-local runtime module names, and templates for replacement, partial
+patch, and removal. A patch must use the current `base_generation`, select
+complete runtime modules, and provide complete A/B pairs for every selected
+module. Unselected modules retain their current weights.
+
 ## Trainer-Side API
 
 Both backends provide static methods that the trainer calls to send weights. The general pattern is:

tests/distributed/test_weight_transfer.py

@@ -7,6 +7,7 @@
 """
 
 import pickle
+from types import SimpleNamespace
 from unittest.mock import MagicMock
 
 import pybase64 as base64
@@ -18,6 +19,7 @@
 from vllm.config.parallel import ParallelConfig
 from vllm.config.weight_transfer import WeightTransferConfig
 from vllm.distributed.weight_transfer import WeightTransferEngineFactory
+from vllm.distributed.weight_transfer.base import WeightTransferEngine
 from vllm.distributed.weight_transfer.ipc_engine import (
     IPCWeightTransferEngine,
     IPCWeightTransferInitInfo,
@@ -91,6 +93,50 @@ def create_mock_vllm_config(
     return vllm_config
 
 
+class TestDeclaredSourceManifest:
+    def _state(self, *, required: bool = True):
+        return SimpleNamespace(
+            _requires_declared_source_manifest=required,
+            _expected_source_names=None,
+            _received_source_names=set(),
+        )
+
+    def test_complete_chunked_manifest(self):
+        state = self._state()
+        WeightTransferEngine.observe_source_manifest(
+            state, ["q_proj.weight"], ["q_proj.weight", "k_proj.weight"]
+        )
+        WeightTransferEngine.observe_source_manifest(
+            state, ["k_proj.weight"], None
+        )
+        WeightTransferEngine.finish_source_manifest_validation(state)
+
+    def test_missing_packed_fragment_fails(self):
+        state = self._state()
+        WeightTransferEngine.observe_source_manifest(
+            state,
+            ["q_proj.weight", "v_proj.weight"],
+            ["q_proj.weight", "k_proj.weight", "v_proj.weight"],
+        )
+        with pytest.raises(RuntimeError, match=r"k_proj\.weight"):
+            WeightTransferEngine.finish_source_manifest_validation(state)
+
+    def test_dummy_first_transfer_requires_declaration(self):
+        state = self._state()
+        WeightTransferEngine.observe_source_manifest(
+            state, ["q_proj.weight"], None
+        )
+        with pytest.raises(RuntimeError, match="authoritative expected"):
+            WeightTransferEngine.finish_source_manifest_validation(state)
+
+    def test_later_transfer_keeps_backward_compatibility(self):
+        state = self._state(required=False)
+        WeightTransferEngine.observe_source_manifest(
+            state, ["q_proj.weight"], None
+        )
+        WeightTransferEngine.finish_source_manifest_validation(state)
+
+
 # --- Unit Tests: NCCLWeightTransferUpdateInfo Validation ---
 
 

tests/entrypoints/weight_transfer/test_weight_update_baseline_api.py

@@ -0,0 +1,41 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import asyncio
+import json
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+from vllm.entrypoints.serve.dev.rlhf.api_router import (
+    get_weight_update_manifest,
+)
+
+
+def test_weight_update_manifest_endpoint_returns_engine_manifest() -> None:
+    manifest = {
+        "model_weights": {
+            "ready": True,
+            "scope_template": {
+                "kind": "base_checkpoint",
+                "mode": "partial",
+                "source_names": [],
+            },
+            "source_names": ["layer.weight"],
+            "atomic_source_groups": [["layer.weight"]],
+            "workers": [],
+            "reason": None,
+        },
+        "lora_adapters": [],
+    }
+    engine = SimpleNamespace(
+        get_weight_update_manifest=AsyncMock(return_value=manifest)
+    )
+    request = SimpleNamespace(
+        app=SimpleNamespace(state=SimpleNamespace(engine_client=engine))
+    )
+
+    response = asyncio.run(get_weight_update_manifest(request))
+
+    assert response.status_code == 200
+    assert json.loads(response.body) == manifest
+    engine.get_weight_update_manifest.assert_awaited_once_with()