|
| 1 | +"""Tests for the aggressive-offload memory management feature. |
| 2 | +
|
| 3 | +These tests validate the Apple Silicon (MPS) memory optimisation path without |
| 4 | +requiring a GPU or actual model weights. Every test mocks the relevant model |
| 5 | +and cache structures so the suite can run in CI on any platform. |
| 6 | +""" |
| 7 | + |
| 8 | +import pytest |
| 9 | +import types |
| 10 | +import torch |
| 11 | +import torch.nn as nn |
| 12 | + |
| 13 | +# --------------------------------------------------------------------------- |
| 14 | +# Fixtures & helpers |
| 15 | +# --------------------------------------------------------------------------- |
| 16 | + |
| 17 | +class FakeLinearModel(nn.Module): |
| 18 | + """Minimal nn.Module whose parameters consume measurable memory.""" |
| 19 | + |
| 20 | + def __init__(self, size_mb: float = 2.0): |
| 21 | + super().__init__() |
| 22 | + # Each float32 param = 4 bytes, so `n` params ≈ size_mb * 1024² / 4 |
| 23 | + n = int(size_mb * 1024 * 1024 / 4) |
| 24 | + self.weight = nn.Parameter(torch.zeros(n, dtype=torch.float32)) |
| 25 | + |
| 26 | + |
| 27 | +class FakeModelPatcher: |
| 28 | + """Mimics the subset of ModelPatcher used by model_management.free_memory.""" |
| 29 | + |
| 30 | + def __init__(self, size_mb: float = 2.0): |
| 31 | + self.model = FakeLinearModel(size_mb) |
| 32 | + self._loaded_size = int(size_mb * 1024 * 1024) |
| 33 | + |
| 34 | + def loaded_size(self): |
| 35 | + return self._loaded_size |
| 36 | + |
| 37 | + def is_dynamic(self): |
| 38 | + return False |
| 39 | + |
| 40 | + |
| 41 | +class FakeLoadedModel: |
| 42 | + """Mimics LoadedModel entries in current_loaded_models.""" |
| 43 | + |
| 44 | + def __init__(self, patcher: FakeModelPatcher, *, currently_used: bool = False): |
| 45 | + self._model = patcher |
| 46 | + self.currently_used = currently_used |
| 47 | + |
| 48 | + @property |
| 49 | + def model(self): |
| 50 | + return self._model |
| 51 | + |
| 52 | + def model_memory(self): |
| 53 | + return self._model.loaded_size() |
| 54 | + |
| 55 | + def model_unload(self, _memory_to_free): |
| 56 | + return True |
| 57 | + |
| 58 | + def model_load(self, _device, _keep_loaded): |
| 59 | + pass |
| 60 | + |
| 61 | + |
| 62 | +# --------------------------------------------------------------------------- |
| 63 | +# 1. BasicCache.clear_all() |
| 64 | +# --------------------------------------------------------------------------- |
| 65 | + |
| 66 | +class TestBasicCacheClearAll: |
| 67 | + """Verify that BasicCache.clear_all() is a proper public API.""" |
| 68 | + |
| 69 | + def test_clear_all_empties_cache_and_subcaches(self): |
| 70 | + """clear_all() must remove every entry in both dicts.""" |
| 71 | + from comfy_execution.caching import BasicCache, CacheKeySetInputSignature |
| 72 | + |
| 73 | + cache = BasicCache(CacheKeySetInputSignature) |
| 74 | + cache.cache["key1"] = "value1" |
| 75 | + cache.cache["key2"] = "value2" |
| 76 | + cache.subcaches["sub1"] = "subvalue1" |
| 77 | + |
| 78 | + cache.clear_all() |
| 79 | + |
| 80 | + assert len(cache.cache) == 0 |
| 81 | + assert len(cache.subcaches) == 0 |
| 82 | + |
| 83 | + def test_clear_all_is_idempotent(self): |
| 84 | + """Calling clear_all() on an already-empty cache must not raise.""" |
| 85 | + from comfy_execution.caching import BasicCache, CacheKeySetInputSignature |
| 86 | + |
| 87 | + cache = BasicCache(CacheKeySetInputSignature) |
| 88 | + cache.clear_all() # should be a no-op |
| 89 | + cache.clear_all() # still a no-op |
| 90 | + |
| 91 | + assert len(cache.cache) == 0 |
| 92 | + |
| 93 | + def test_null_cache_clear_all_is_noop(self): |
| 94 | + """NullCache.clear_all() must not raise — it's the null backend.""" |
| 95 | + from comfy_execution.caching import NullCache |
| 96 | + |
| 97 | + null = NullCache() |
| 98 | + null.clear_all() # must not raise AttributeError |
| 99 | + |
| 100 | + def test_lru_cache_clear_all_resets_metadata(self): |
| 101 | + """LRUCache.clear_all() must also reset used_generation, children, min_generation.""" |
| 102 | + from comfy_execution.caching import LRUCache, CacheKeySetInputSignature |
| 103 | + |
| 104 | + cache = LRUCache(CacheKeySetInputSignature, max_size=10) |
| 105 | + # Simulate some entries |
| 106 | + cache.cache["k1"] = "v1" |
| 107 | + cache.used_generation["k1"] = 5 |
| 108 | + cache.children["k1"] = ["child1"] |
| 109 | + cache.min_generation = 3 |
| 110 | + cache.generation = 5 |
| 111 | + |
| 112 | + cache.clear_all() |
| 113 | + |
| 114 | + assert len(cache.cache) == 0 |
| 115 | + assert len(cache.used_generation) == 0 |
| 116 | + assert len(cache.children) == 0 |
| 117 | + assert cache.min_generation == 0 |
| 118 | + # generation counter should NOT be reset (it's a monotonic counter) |
| 119 | + assert cache.generation == 5 |
| 120 | + |
| 121 | + def test_ram_pressure_cache_clear_all_resets_timestamps(self): |
| 122 | + """RAMPressureCache.clear_all() must also reset timestamps.""" |
| 123 | + from comfy_execution.caching import RAMPressureCache, CacheKeySetInputSignature |
| 124 | + |
| 125 | + cache = RAMPressureCache(CacheKeySetInputSignature) |
| 126 | + cache.cache["k1"] = "v1" |
| 127 | + cache.used_generation["k1"] = 2 |
| 128 | + cache.timestamps["k1"] = 1234567890.0 |
| 129 | + |
| 130 | + cache.clear_all() |
| 131 | + |
| 132 | + assert len(cache.cache) == 0 |
| 133 | + assert len(cache.used_generation) == 0 |
| 134 | + assert len(cache.timestamps) == 0 |
| 135 | + |
| 136 | + |
| 137 | +# --------------------------------------------------------------------------- |
| 138 | +# 2. Callback registration & dispatch |
| 139 | +# --------------------------------------------------------------------------- |
| 140 | + |
| 141 | +class TestModelDestroyedCallbacks: |
| 142 | + """Validate the on_model_destroyed lifecycle callback system.""" |
| 143 | + |
| 144 | + def setup_method(self): |
| 145 | + """Reset the callback list before every test.""" |
| 146 | + import comfy.model_management as mm |
| 147 | + self._original = mm._on_model_destroyed_callbacks.copy() |
| 148 | + mm._on_model_destroyed_callbacks.clear() |
| 149 | + |
| 150 | + def teardown_method(self): |
| 151 | + """Restore the original callback list.""" |
| 152 | + import comfy.model_management as mm |
| 153 | + mm._on_model_destroyed_callbacks.clear() |
| 154 | + mm._on_model_destroyed_callbacks.extend(self._original) |
| 155 | + |
| 156 | + def test_register_single_callback(self): |
| 157 | + import comfy.model_management as mm |
| 158 | + |
| 159 | + invocations = [] |
| 160 | + mm.register_model_destroyed_callback(lambda reason: invocations.append(reason)) |
| 161 | + |
| 162 | + assert len(mm._on_model_destroyed_callbacks) == 1 |
| 163 | + |
| 164 | + # Simulate dispatch |
| 165 | + for cb in mm._on_model_destroyed_callbacks: |
| 166 | + cb("test") |
| 167 | + assert invocations == ["test"] |
| 168 | + |
| 169 | + def test_register_multiple_callbacks(self): |
| 170 | + """Multiple registrants must all fire — no silent overwrites.""" |
| 171 | + import comfy.model_management as mm |
| 172 | + |
| 173 | + results_a, results_b = [], [] |
| 174 | + mm.register_model_destroyed_callback(lambda r: results_a.append(r)) |
| 175 | + mm.register_model_destroyed_callback(lambda r: results_b.append(r)) |
| 176 | + |
| 177 | + for cb in mm._on_model_destroyed_callbacks: |
| 178 | + cb("batch") |
| 179 | + |
| 180 | + assert results_a == ["batch"] |
| 181 | + assert results_b == ["batch"] |
| 182 | + |
| 183 | + def test_callback_receives_reason_string(self): |
| 184 | + """The callback signature is (reason: str) -> None.""" |
| 185 | + import comfy.model_management as mm |
| 186 | + |
| 187 | + captured = {} |
| 188 | + def _cb(reason): |
| 189 | + captured["reason"] = reason |
| 190 | + captured["type"] = type(reason).__name__ |
| 191 | + |
| 192 | + mm.register_model_destroyed_callback(_cb) |
| 193 | + for cb in mm._on_model_destroyed_callbacks: |
| 194 | + cb("batch") |
| 195 | + |
| 196 | + assert captured["reason"] == "batch" |
| 197 | + assert captured["type"] == "str" |
| 198 | + |
| 199 | + |
| 200 | +# --------------------------------------------------------------------------- |
| 201 | +# 3. Meta-device destruction threshold |
| 202 | +# --------------------------------------------------------------------------- |
| 203 | + |
| 204 | +class TestMetaDeviceThreshold: |
| 205 | + """Verify that only models > 1 GB are queued for meta-device destruction.""" |
| 206 | + |
| 207 | + def test_small_model_not_destroyed(self): |
| 208 | + """A 160 MB model (VAE-sized) must NOT be moved to meta device.""" |
| 209 | + model = FakeLinearModel(size_mb=160) |
| 210 | + |
| 211 | + # Simulate the threshold check from free_memory |
| 212 | + model_size = sum(p.numel() * p.element_size() for p in model.parameters()) |
| 213 | + threshold = 1024 * 1024 * 1024 # 1 GB |
| 214 | + |
| 215 | + assert model_size < threshold, ( |
| 216 | + f"160 MB model should be below 1 GB threshold, got {model_size / (1024**2):.0f} MB" |
| 217 | + ) |
| 218 | + # Confirm parameters are still on a real device |
| 219 | + assert model.weight.device.type != "meta" |
| 220 | + |
| 221 | + def test_large_model_above_threshold(self): |
| 222 | + """A 2 GB model (UNET/CLIP-sized) must BE above the destruction threshold.""" |
| 223 | + # Use a meta-device tensor to avoid allocating 2 GB of real memory. |
| 224 | + # Meta tensors report correct numel/element_size but use zero storage. |
| 225 | + n = int(2048 * 1024 * 1024 / 4) # 2 GB in float32 params |
| 226 | + meta_weight = torch.empty(n, dtype=torch.float32, device="meta") |
| 227 | + |
| 228 | + model_size = meta_weight.numel() * meta_weight.element_size() |
| 229 | + threshold = 1024 * 1024 * 1024 # 1 GB |
| 230 | + |
| 231 | + assert model_size > threshold, ( |
| 232 | + f"2 GB model should be above 1 GB threshold, got {model_size / (1024**2):.0f} MB" |
| 233 | + ) |
| 234 | + |
| 235 | + def test_meta_device_move_releases_storage(self): |
| 236 | + """Moving parameters to 'meta' must place them on the meta device.""" |
| 237 | + model = FakeLinearModel(size_mb=2) |
| 238 | + assert model.weight.device.type != "meta" |
| 239 | + |
| 240 | + model.to(device="meta") |
| 241 | + |
| 242 | + assert model.weight.device.type == "meta" |
| 243 | + # Meta tensors retain their logical shape but live on a virtual device |
| 244 | + # with no physical backing — this is what releases RAM. |
| 245 | + assert model.weight.nelement() > 0 # still has logical shape |
| 246 | + assert model.weight.untyped_storage().device.type == "meta" |
| 247 | + |
| 248 | + |
| 249 | +# --------------------------------------------------------------------------- |
| 250 | +# 4. MPS flush conditionality |
| 251 | +# --------------------------------------------------------------------------- |
| 252 | + |
| 253 | +class TestMpsFlushConditionality: |
| 254 | + """Verify the MPS flush only activates under correct conditions.""" |
| 255 | + |
| 256 | + def test_flush_requires_aggressive_offload_flag(self): |
| 257 | + """The MPS flush in samplers is gated on AGGRESSIVE_OFFLOAD.""" |
| 258 | + import comfy.model_management as mm |
| 259 | + |
| 260 | + # When False, flush should NOT be injected |
| 261 | + original = getattr(mm, "AGGRESSIVE_OFFLOAD", False) |
| 262 | + try: |
| 263 | + mm.AGGRESSIVE_OFFLOAD = False |
| 264 | + assert not (True and getattr(mm, "AGGRESSIVE_OFFLOAD", False)) |
| 265 | + |
| 266 | + mm.AGGRESSIVE_OFFLOAD = True |
| 267 | + assert (True and getattr(mm, "AGGRESSIVE_OFFLOAD", False)) |
| 268 | + finally: |
| 269 | + mm.AGGRESSIVE_OFFLOAD = original |
| 270 | + |
| 271 | + def test_flush_requires_mps_device(self): |
| 272 | + """The flush condition checks device.type == 'mps'.""" |
| 273 | + # Simulate CPU device — flush should not activate |
| 274 | + cpu_device = torch.device("cpu") |
| 275 | + assert cpu_device.type != "mps" |
| 276 | + |
| 277 | + # Simulate MPS device string check |
| 278 | + if torch.backends.mps.is_available(): |
| 279 | + mps_device = torch.device("mps") |
| 280 | + assert mps_device.type == "mps" |
| 281 | + |
| 282 | + |
| 283 | +# --------------------------------------------------------------------------- |
| 284 | +# 5. AGGRESSIVE_OFFLOAD flag integration |
| 285 | +# --------------------------------------------------------------------------- |
| 286 | + |
| 287 | +class TestAggressiveOffloadFlag: |
| 288 | + """Verify the CLI flag is correctly exposed.""" |
| 289 | + |
| 290 | + def test_flag_exists_in_model_management(self): |
| 291 | + """AGGRESSIVE_OFFLOAD must be importable from model_management.""" |
| 292 | + import comfy.model_management as mm |
| 293 | + assert hasattr(mm, "AGGRESSIVE_OFFLOAD") |
| 294 | + assert isinstance(mm.AGGRESSIVE_OFFLOAD, bool) |
| 295 | + |
| 296 | + def test_flag_defaults_from_cli_args(self): |
| 297 | + """The flag should be wired from cli_args to model_management.""" |
| 298 | + import comfy.cli_args as cli_args |
| 299 | + import comfy.model_management as mm |
| 300 | + assert hasattr(cli_args.args, "aggressive_offload") |
| 301 | + assert mm.AGGRESSIVE_OFFLOAD == cli_args.args.aggressive_offload |
0 commit comments