diff --git a/.github/workflows/gpu_tests.yml b/.github/workflows/gpu_tests.yml
index 5094075dafd..ca18afd0e65 100644
--- a/.github/workflows/gpu_tests.yml
+++ b/.github/workflows/gpu_tests.yml
@@ -99,6 +99,9 @@ jobs:
           # sys.monitoring records the same lines as the default C tracer at a fraction of the
           # overhead on import-heavy worker processes. Needs Python 3.12+.
           COVERAGE_CORE: sysmon
+          # Tests marked `manual` (real-weight downloads and the like) run on nightly and on-demand
+          # runs, as in the unit tests, not on pull requests (copy-pr-bot runs those as `push`).
+          PYTEST_ADDOPTS: ${{ github.event_name != 'push' && '--run-manual' || '' }}
         run: |
           # Use `python3` (the vllm image has no `python` on PATH)
           python3 -m pip install nox && nox -s ${{ matrix.example }}
diff --git a/tests/gpu/torch/utils/test_dataset_utils.py b/tests/gpu/torch/utils/test_dataset_utils.py
index bc9451ca03a..35aa3663def 100644
--- a/tests/gpu/torch/utils/test_dataset_utils.py
+++ b/tests/gpu/torch/utils/test_dataset_utils.py
@@ -42,6 +42,9 @@ def qwen3_tokenizer():
     return tokenizer
 
 
+@pytest.mark.manual(
+    reason="Streams the 7 gated Nemotron v3 datasets; runs on the nightly schedule."
+)
 def test_get_dataset_dataloader_nemotron_v3_chat_template(qwen3_tokenizer):
     if not get_token():
         pytest.skip(
diff --git a/tests/gpu_megatron/torch/utils/plugins/test_utils_megatron.py b/tests/gpu_megatron/torch/utils/plugins/test_utils_megatron.py
index 8d3af2d7f3c..acec7515087 100644
--- a/tests/gpu_megatron/torch/utils/plugins/test_utils_megatron.py
+++ b/tests/gpu_megatron/torch/utils/plugins/test_utils_megatron.py
@@ -64,6 +64,9 @@ def _test_megatron_generate_and_mmlu(rank, size, parallelism):
     assert 0.36 < megatron_mmlu(model, tokenizer, fraction=0.1, batch_size=16) < 0.39
 
 
+@pytest.mark.manual(
+    reason="Real Qwen3-0.6B weights and a 1.2-1.5 GB download; run with --run-manual."
+)
 @pytest.mark.parametrize("parallelism", ["tp", "pp", "cp", "dp"])
 def test_megatron_generate_and_mmlu(dist_workers, parallelism, num_gpus):
     if num_gpus == 1 and parallelism != "tp":