[https://nvbugs/6631019][test] reset compile mode in reused MPI workers + unwaive Nemotron Nano FP8 CUDA graph test on DGX B200

tests/integration/test_lists/waives.txt

@@ -173,7 +173,6 @@ full:B300/disaggregated/test_disaggregated.py::test_disaggregated_logprobs_servi
 full:B300/llmapi/test_llm_api_pytorch_moe_lora.py::test_qwen_moe_routed_expert_multi_lora_varying_ranks[cudagraph] SKIP (https://nvbugs/6475623)
 full:B300/unittest/_torch/attention/test_attention_backends.py::test_attention_backend[qwen2_0_5b_gqa_hd64-ctx-bf16-HND-p32-v1] SKIP (https://nvbugs/6610548)
 full:DGX_B200/accuracy/test_llm_api_pytorch.py::TestDeepSeekV4Pro::test_gsm8k_full_accuracy SKIP (https://nvbugs/6571418)
-full:DGX_B200/accuracy/test_llm_api_pytorch_multimodal.py::TestNanoV3Omni::test_auto_dtype[fp8_mmmu_encoder_cuda_graph] SKIP (https://nvbugs/6631019)
 full:DGX_B200/disaggregated/test_disaggregated.py::test_disaggregated_gpt_oss_120b_harmony[gpt_oss/gpt-oss-120b] SKIP (https://nvbugs/6594241)
 full:DGX_B200/perf/test_perf_sanity.py::test_e2e[aggr_upload-gemma4_26b_a4b_nvfp4_blackwell-gemma4_26b_a4b_nvfp4_tp1_1k1k] SKIP (https://nvbugs/6571410)
 full:DGX_B200/perf/test_perf_sanity.py::test_e2e[aggr_upload-host_perf_llama8b_spec_decode-llama8b_spec_bs1_128_128] SKIP (https://nvbugs/6571408)

tests/test_common/grouped_test_utils.py

@@ -59,17 +59,17 @@ def submit_sync_per_worker(mpi_session, fn, timeout: float = 60.0) -> list:
 
 
 def reset_worker_torch_compile_state() -> None:
-    """Reset per-worker torch.compile / Dynamo state (runs inside each worker).
+    """Reset per-worker torch.compile state (runs inside each worker).
 
-    Dynamo's recompile counter is process-global and per-code-object. When
-    worker processes are reused across LLMs (shared ``MpiPoolSession``), each
-    ``torch_compile`` case recompiles the same ``model.forward`` code object
-    under new guards; the count accumulates and eventually trips
-    ``recompile_limit`` (16), which is a HARD failure under ``fullgraph=True``
-    (``FailOnRecompileLimitHit``) and aborts the whole MPI job. Resetting
-    between cases makes each LLM start from a clean compile cache, like a fresh
-    process. Run on every worker via ``submit_sync_per_worker``.
+    Both Dynamo's recompile counter and TensorRT-LLM's compile-mode flag are
+    process-global. When worker processes are reused across LLMs (shared
+    ``MpiPoolSession``), each case must start with a clean compile cache and
+    compile mode disabled, like a fresh process. Run on every worker via
+    ``submit_sync_per_worker``.
     """
     import torch
 
+    from tensorrt_llm._torch.utils import set_torch_compiling
+
     torch._dynamo.reset()
+    set_torch_compiling(False)

tests/unittest/llmapi/test_session_reuse.py

@@ -83,6 +83,22 @@ def test_reuse_hands_back_same_pool(reuse_cache):
     assert len(reuse_cache.prefetch.restocks) == 1  # reuse does not create a shadow
 
 
+def test_worker_reset_clears_tensorrt_llm_compile_mode(monkeypatch):
+    import torch
+    from test_common.grouped_test_utils import reset_worker_torch_compile_state
+
+    from tensorrt_llm._torch.utils import is_torch_compiling, set_torch_compiling
+
+    dynamo_resets = []
+    monkeypatch.setattr(torch._dynamo, "reset", lambda: dynamo_resets.append(True))
+    set_torch_compiling(True)
+
+    reset_worker_torch_compile_state()
+
+    assert dynamo_resets == [True]
+    assert not is_torch_compiling()
+
+
 def test_cached_handover_reaps_in_flight_retires(reuse_cache):
     # Two same-size pools released back-to-back (concurrent LLMs in one
     # test): the duplicate is retired in a BACKGROUND thread while holding