test(promote): add training promote e2e regression for MobileNetV1 + CCT

Adds a new CI job `siracusa-training-tiled-promote` that runs
MobileNetV1 and CCT training with PromoteTensorsToL2 enabled. Each
model is tested in two configurations:
  - "off" (baseline, no promotion)
  - "cycle-aware" with includeActivations=True (full promote)

This catches regressions in the promote-for-training code path that
was validated in this branch:
  MobileNetV1: 332M cycles (-19.9% vs 414M baseline), 0/4 PASS
  CCT:         450M cycles (-14.8% vs 528M baseline), 4/4 baseline

pytest marker: `promote and training` (new combination).
CI workflow: ci-platform-siracusa-tiled.yml.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: runwangdl

.github/workflows/ci-platform-siracusa-tiled.yml

@@ -55,3 +55,12 @@ jobs:
       runner: ${{ needs.select-env.outputs.runner }}
       docker-image: ${{ needs.select-env.outputs.image }}
       pytest-marker: "promote and not training"
+
+  # Training L3 + PromoteTensorsToL2 regression (MobileNetV1 + CCT)
+  siracusa-training-tiled-promote:
+    needs: select-env
+    uses: ./.github/workflows/_runner-siracusa-tiled.yml
+    with:
+      runner: ${{ needs.select-env.outputs.runner }}
+      docker-image: ${{ needs.select-env.outputs.image }}
+      pytest-marker: "promote and training"

DeeployTest/test_platforms.py

@@ -41,6 +41,7 @@
 from test_siracusa_tiled_config import L3_DOUBLEBUFFER_MODELS, L3_SINGLEBUFFER_MODELS
 from test_siracusa_tiled_config import L3_SINGLEBUFFER_PROMOTE_MODELS as SIRACUSA_L3_SINGLEBUFFER_PROMOTE_MODELS
 from test_siracusa_tiled_config import L3_SINGLEBUFFER_TRAINING_MODELS as SIRACUSA_L3_SINGLEBUFFER_TRAINING_MODELS
+from test_siracusa_tiled_config import L3_SINGLEBUFFER_TRAINING_PROMOTE_MODELS as SIRACUSA_L3_SINGLEBUFFER_TRAINING_PROMOTE_MODELS
 from test_siracusa_tiled_config import TRAINING_MODEL_OVERRIDES as SIRACUSA_TRAINING_MODEL_OVERRIDES
 from test_snitch_config import DEFAULT_NUM_CORES as SNITCH_DEFAULT_NUM_CORES
 from test_snitch_config import KERNEL_TESTS as SNITCH_KERNEL_TESTS
@@ -415,6 +416,43 @@ def test_siracusa_tiled_training_l3_singlebuffer(test_params, deeploy_test_dir,
     run_and_assert_test(test_name, config, skipgen, skipsim)
 
 
+@pytest.mark.promote
+@pytest.mark.training
+@pytest.mark.siracusa_tiled
+@pytest.mark.l3
+@pytest.mark.singlebuffer
+@pytest.mark.parametrize(
+    "test_params",
+    _generate_promote_test_params(SIRACUSA_L3_SINGLEBUFFER_TRAINING_PROMOTE_MODELS),
+    ids = _promote_param_id,
+)
+def test_siracusa_tiled_training_promote_l3_singlebuffer(test_params, deeploy_test_dir, toolchain, toolchain_dir,
+                                                         cmake_args, skipgen, skipsim) -> None:
+    """L3 training + PromoteTensorsToL2 regression (MobileNetV1 + CCT)."""
+    test_name, l1, strategy, include_acts, promote = test_params
+    overrides = SIRACUSA_TRAINING_MODEL_OVERRIDES.get(test_name, {})
+    config = create_test_config(
+        test_name = test_name,
+        platform = "Siracusa",
+        simulator = "gvsoc",
+        deeploy_test_dir = deeploy_test_dir,
+        toolchain = toolchain,
+        toolchain_dir = toolchain_dir,
+        cmake_args = cmake_args,
+        tiling = True,
+        cores = SIRACUSA_DEFAULT_CORES,
+        l1 = l1,
+        default_mem_level = "L3",
+        double_buffer = False,
+        training = True,
+        training_num_data_inputs = overrides.get("num_data_inputs"),
+        training_tolerance = overrides.get("tolerance"),
+        promote_to_l2 = promote,
+        promote_to_l2_strategy = strategy if promote else "cycle-aware",
+    )
+    run_and_assert_test(test_name, config, skipgen, skipsim)
+
+
 @pytest.mark.siracusa_tiled
 @pytest.mark.kernels
 @pytest.mark.singlebuffer

DeeployTest/test_siracusa_tiled_config.py

@@ -198,6 +198,20 @@
     },
 }
 
+# Training models tested with PromoteTensorsToL2.
+# Same shape as the inference promote dict: test path -> list of
+# (l1, strategy, includeActivations). "off" = baseline (no promotion).
+L3_SINGLEBUFFER_TRAINING_PROMOTE_MODELS = {
+    "Models/Training/MobileNetV1/mobilenetv1_train": [
+        (128000, "off", False),
+        (128000, "cycle-aware", True),
+    ],
+    "Models/Training/CCT/cct_train": [
+        (128000, "off", False),
+        (128000, "cycle-aware", True),
+    ],
+}
+
 # Inference models tested with PromoteTensorsToL2.
 # Each entry maps test path -> list of (l1, strategy, includeActivations).
 # "off" strategy = no promotion (baseline for cycle comparison).