fix(servicer): never fail a served request on a media identity it cannot build

grpc_servicer/smg_grpc_servicer/vllm/servicer.py

@@ -323,10 +323,21 @@ async def Generate(
                 finally:
                     self._mm_inflight.release()
                 # A PD prefill leg answers with the identity so decode is
-                # served without pixels or references.
+                # served without pixels or references. The identity is an
+                # optimisation with a fallback (decode reprocesses), so a
+                # shape it cannot read must not fail a served request.
                 if kv_transfer_params is not None:
                     if media_identity_supported():
-                        media_identity = build_media_identity(prompt)
+                        try:
+                            media_identity = build_media_identity(prompt)
+                        except Exception as e:  # noqa: BLE001 - any failure falls back
+                            logger.warning(
+                                "Request %s: media identity not built (%s); the decode leg "
+                                "will reprocess the media",
+                                request_id,
+                                e,
+                            )
+                            media_identity = None
                     else:
                         logger.warning(
                             "Request %s: the installed smg-grpc-proto has no media_identity; "

grpc_servicer/tests/test_vllm_pd_media_identity.py

@@ -154,3 +154,27 @@ async def test_decode_leg_from_identity_never_touches_the_processor(monkeypatch)
     decode = await _complete(servicer, _request(kv=True, identity=prefill.media_identity))
     assert processor.calls == [], "the decode leg carried identity, not references"
     assert not decode.HasField("media_identity")
+
+
+@pytest.mark.asyncio
+async def test_an_identity_that_cannot_be_built_does_not_fail_the_request(monkeypatch, caplog):
+    # The identity is an optimisation with a fallback (decode reprocesses the
+    # media), so a shape it did not anticipate must not turn a served request
+    # into an error.
+    from smg_grpc_servicer.vllm import servicer as servicer_module
+
+    def explode(prompt):
+        raise RuntimeError("stack expects each tensor to be equal size")
+
+    monkeypatch.setattr(servicer_module, "build_media_identity", explode)
+    processor = _Processor()
+    servicer = _servicer(monkeypatch, processor)
+    with caplog.at_level("WARNING", logger="smg_grpc_servicer.vllm.servicer"):
+        complete = await _complete(servicer, _request(refs=True, kv=True))
+    assert processor.calls == ["req-pd"]
+    assert list(complete.output_ids) == [42]
+    assert not complete.HasField("media_identity")
+    assert any(
+        "media identity not built" in r.getMessage() and "equal size" in r.getMessage()
+        for r in caplog.records
+    )