PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
# Copyright 2025 Intel Corporation
# SPDX-License-Identifier: Apache-2.0
"""Tests for the in-place calibration-pool release (composer)."""
import unittest
import torch
from auto_round.algorithms.composer import _release_pool_inplace
class TestReleasePoolInplace(unittest.TestCase):
def test_list_pool_entries_nulled_and_shared_ref_sees_it(self):
pool = [torch.zeros(2), torch.zeros(2)]
shared = pool # the orchestrator's input_ids view
_release_pool_inplace(pool)
self.assertEqual(pool, [None, None])
self.assertIs(shared, pool)
self.assertEqual(shared, [None, None])
def test_dict_of_lists_recurces(self):
pool = {"hidden_states": [torch.zeros(2)], "other": {"nested": [torch.zeros(2)]}}
_release_pool_inplace(pool)
self.assertEqual(pool["hidden_states"], [None])
self.assertEqual(pool["other"]["nested"], [None])
self.assertEqual(list(pool.keys()), ["hidden_states", "other"]) # structure preserved
def test_tuple_skipped_immutable(self):
pool = (torch.zeros(2), torch.zeros(2))
_release_pool_inplace(pool)
# tuples cannot be mutated in place; tensors must remain untouched
self.assertIsInstance(pool[0], torch.Tensor)
self.assertIsInstance(pool[1], torch.Tensor)
def test_mixed_structure(self):
pool = {"a": [torch.zeros(1), (torch.zeros(1),)], "b": "scalar"}
_release_pool_inplace(pool)
self.assertEqual(pool["a"][0], None)
self.assertIsNone(pool["a"][1]) # slot dropped wholesale; tuple internals moot
self.assertEqual(pool["b"], "scalar")