Add llamatune checks with quantized values

mlos_bench/mlos_bench/tests/optimizers/toy_optimization_loop_test.py

@@ -113,7 +113,7 @@ def test_smac_optimization_loop(mock_env_no_noise: MockEnv, smac_opt: MlosCoreOp
  "vmSize": "Standard_B2s",
  "idle": "mwait",
  "kernel_sched_migration_cost_ns": 297669,
- "kernel_sched_latency_ns": 290365137,
+ "kernel_sched_latency_ns": 290365100,
  }
  assert score == pytest.approx(expected_score, 0.01)
  assert tunables.get_param_values() == expected_tunable_values

mlos_bench/mlos_bench/tests/tunable_groups_fixtures.py

@@ -62,7 +62,8 @@
  "type": "int",
  "default": 2000000,
  "range": [0, 1000000000],
- "log": false
+ "log": false,
+ "quantization": 100
  }
  }
  }

mlos_core/mlos_core/tests/spaces/adapters/llamatune_test.py

@@ -15,9 +15,16 @@
 from mlos_core.spaces.adapters import LlamaTuneAdapter
 
 
-def construct_parameter_space(
+# Explicitly test quantized values with llamatune space adapter.
+# TODO: Add log scale sampling tests as well.
+
+
+def construct_parameter_space( # pylint: disable=too-many-arguments
+ *,
  n_continuous_params: int = 0,
+ n_quantized_continuous_params: int = 0,
  n_integer_params: int = 0,
+ n_quantized_integer_params: int = 0,
  n_categorical_params: int = 0,
  seed: int = 1234,
 ) -> CS.ConfigurationSpace:
@@ -26,8 +33,12 @@ def construct_parameter_space(
 
  for idx in range(n_continuous_params):
  input_space.add(CS.UniformFloatHyperparameter(name=f"cont_{idx}", lower=0, upper=64))
+ for idx in range(n_quantized_continuous_params):
+ input_space.add(CS.UniformFloatHyperparameter(name=f"cont_{idx}", lower=0, upper=64, q=12.8))
  for idx in range(n_integer_params):
  input_space.add(CS.UniformIntegerHyperparameter(name=f"int_{idx}", lower=-1, upper=256))
+ for idx in range(n_quantized_integer_params):
+ input_space.add(CS.UniformIntegerHyperparameter(name=f"int_{idx}", lower=0, upper=256, q=16))
  for idx in range(n_categorical_params):
  input_space.add(
  CS.CategoricalHyperparameter(
@@ -49,6 +60,13 @@ def construct_parameter_space(
  {"n_continuous_params": int(num_target_space_dims * num_orig_space_factor)},
  {"n_integer_params": int(num_target_space_dims * num_orig_space_factor)},
  {"n_categorical_params": int(num_target_space_dims * num_orig_space_factor)},
+ {"n_categorical_params": int(num_target_space_dims * num_orig_space_factor)},
+ {"n_quantized_integer_params": int(num_target_space_dims * num_orig_space_factor)},
+ {
+ "n_quantized_continuous_params": int(
+ num_target_space_dims * num_orig_space_factor
+ )
+ },
  # Mix of all three types
  {
  "n_continuous_params": int(num_target_space_dims * num_orig_space_factor / 3),
@@ -358,6 +376,12 @@ def test_max_unique_values_per_param() -> None:
  {"n_continuous_params": int(num_target_space_dims * num_orig_space_factor)},
  {"n_integer_params": int(num_target_space_dims * num_orig_space_factor)},
  {"n_categorical_params": int(num_target_space_dims * num_orig_space_factor)},
+ {"n_quantized_integer_params": int(num_target_space_dims * num_orig_space_factor)},
+ {
+ "n_quantized_continuous_params": int(
+ num_target_space_dims * num_orig_space_factor
+ )
+ },
  # Mix of all three types
  {
  "n_continuous_params": int(num_target_space_dims * num_orig_space_factor / 3),