test(trtllm): pin two tests to a deterministic TRT-LLM capability state

Both passed only by accident against the sandboxed/old-pin CI environment
and failed when actually run against an installed TRT-LLM carrying the full
RL refit lifecycle:

- test_collective_refit_runs_at_async_engine_boundary asserted a call order
  that only holds when WorkerExtension.finalize_weight_update is absent, but
  never forced that fallback path (unlike the sibling
  test_refit_finalization_falls_back_for_older_trtllm, which does). Force it
  explicitly.
- test_collective_refit_always_resets_prefix_cache used a bare MagicMock
  engine, which auto-creates recompute_active_requests and masks the
  reset_prefix_cache fallback its name asserts. Delete the attribute so
  absence is real.

Verified: all 72 trtllm-marked tests in tests/unit/models/generation/trtllm/
pass against a real tekit-backed TRT-LLM install (job 3074375).

Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com>

Author: hchings

tests/unit/models/generation/trtllm/test_trtllm_backend.py

@@ -127,6 +127,13 @@ def test_collective_refit_runs_at_async_engine_boundary(
     from nemo_rl.models.generation.trtllm import trtllm_backend as backend
 
     extension, module, model, model_loader, engine = _extension(backend)
+    # Force the manual finalize fallback so the asserted call order is
+    # deterministic. On an installed TRT-LLM that actually has
+    # WorkerExtension.finalize_weight_update, _finalize_weight_update takes
+    # that fast path instead and never drives process/post individually.
+    monkeypatch.setattr(
+        backend.WorkerExtension, "finalize_weight_update", None, raising=False
+    )
     call_order = []
     incoming = [("model.weight", torch.tensor([1.0]))]
     converted = {"converted.weight": torch.tensor([2.0])}

tests/unit/models/generation/trtllm/test_trtllm_refit.py

@@ -19,6 +19,9 @@ def test_collective_refit_always_resets_prefix_cache():
     engine = MagicMock()
     engine.model_engine = SimpleNamespace(model=model, model_loader=model_loader)
     engine.control_action.return_value = nullcontext()
+    # Force the reset_prefix_cache fallback: a bare MagicMock auto-creates
+    # recompute_active_requests, which would mask it entirely.
+    del engine.recompute_active_requests
 
     extension.engine = engine
     extension.device_id = 0