File size: 14,302 Bytes
2415c4c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 | # SPDX-FileCopyrightText: © 2026 Tenstorrent AI ULC
# SPDX-License-Identifier: Apache-2.0
from dataclasses import replace
from types import SimpleNamespace
import pytest
import torch
import ttnn
from models.common.llm_runtime.config import PagedKVCacheConfig
from models.common.llm_runtime.paged_kv_cache import PagedKVCacheManager, torch_dtype_for_ttnn
class FakeMesh:
def __init__(self, num_devices=2):
self._num_devices = num_devices
def get_num_devices(self):
return self._num_devices
class FakeTensor:
def __init__(self, shape, dtype):
self.shape = tuple(shape)
self.dtype = dtype
def volume(self):
result = 1
for dimension in self.shape:
result *= dimension
return result
def element_size(self):
return 1 if self.dtype in (ttnn.bfloat8_b, ttnn.bfloat4_b) else 2
class FakeModel:
def __init__(self, dtypes=(ttnn.bfloat8_b, ttnn.bfloat8_b), *, bind_fails=False):
mesh = FakeMesh(2)
blocks = []
for dtype in dtypes:
attention = SimpleNamespace(
n_kv_heads=8,
head_dim=16,
kv_cache_dtype=dtype,
paged_attention_config=SimpleNamespace(block_size=32, max_num_blocks=8),
)
blocks.append(SimpleNamespace(attention_config=attention))
self.config = SimpleNamespace(
block_configs=blocks,
n_layers=len(blocks),
num_devices=2,
mesh_device=mesh,
)
self.set_calls = []
self.bound_cache = None
self.bind_fails = bind_fails
def set_kv_cache(self, cache):
self.set_calls.append(cache)
if cache is not None and self.bind_fails:
raise RuntimeError("bind failed")
self.bound_cache = cache
def cache_config(**overrides):
values = {
"block_size": 32,
"max_num_blocks": 8,
"dtype": ttnn.bfloat8_b,
}
values.update(overrides)
return PagedKVCacheConfig(**values)
@pytest.fixture
def fake_allocator(monkeypatch):
allocated = []
deallocated = []
# Non-failure TTNN fakes retain overloaded backend keyword options for assertions.
def as_tensor(host_tensor, **kwargs):
tensor = FakeTensor(host_tensor.shape, kwargs["dtype"])
allocated.append((tensor, host_tensor, kwargs))
return tensor
monkeypatch.setattr(ttnn, "as_tensor", as_tensor)
monkeypatch.setattr(ttnn, "ReplicateTensorToMesh", lambda mesh: ("replicate", mesh))
monkeypatch.setattr(ttnn, "deallocate", lambda tensor: deallocated.append(tensor))
return allocated, deallocated
def test_unresolved_config_accepts_one_resolved_replacement(expect_error):
manager = PagedKVCacheManager(FakeModel(), cache_config())
resolved = replace(manager.config, num_blocks=4)
manager.configure(resolved)
assert manager.config is resolved
assert manager.config.num_blocks == 4
with expect_error(RuntimeError, "only once"):
manager.configure(replace(resolved, num_blocks=5))
def test_resolved_replacement_accepts_model_aligned_external_geometry(expect_error):
model = FakeModel()
manager = PagedKVCacheManager(model, cache_config())
for block in model.config.block_configs:
block.attention_config.paged_attention_config.block_size = 16
block.attention_config.paged_attention_config.max_num_blocks = 12
manager.configure(cache_config(block_size=16, max_num_blocks=12, num_blocks=12))
assert manager.config.block_size == 16
assert manager.config.max_num_blocks == manager.config.num_blocks == 12
other = PagedKVCacheManager(FakeModel(), cache_config())
with expect_error(ValueError, "dtype"):
other.configure(cache_config(dtype=ttnn.bfloat16, num_blocks=4))
def test_pre_resolved_config_starts_configured_and_cannot_be_replaced(expect_error):
manager = PagedKVCacheManager(FakeModel(), cache_config(num_blocks=4))
with expect_error(RuntimeError, "only once"):
manager.configure(cache_config(num_blocks=5))
def test_model_paged_attention_policy_and_uniform_dtype_are_validated(expect_error):
model = FakeModel()
model.config.block_configs[1].attention_config.paged_attention_config.block_size = 16
with expect_error(ValueError, "block_size"):
PagedKVCacheManager(model, cache_config())
with expect_error(ValueError, "model-owned dtype"):
PagedKVCacheManager(FakeModel(), cache_config(dtype=ttnn.bfloat16))
def test_vllm_torch_dtype_mapping_includes_quantized_surrogate(expect_error):
manager = PagedKVCacheManager(FakeModel(), cache_config())
assert torch_dtype_for_ttnn(ttnn.bfloat8_b) == torch.bfloat16
manager.validate_vllm_cache_spec(block_size=32, dtype=torch.bfloat16, num_blocks=8)
with expect_error(ValueError, "incompatible"):
manager.validate_vllm_cache_spec(block_size=32, dtype=torch.float32)
with expect_error(ValueError, "block_size"):
manager.validate_vllm_cache_spec(block_size=16, dtype=torch.bfloat16)
with expect_error(ValueError, "exceeds configured maximum"):
manager.validate_vllm_cache_spec(block_size=32, dtype=torch.bfloat16, num_blocks=9)
def test_nonuniform_model_dtypes_remain_per_layer():
manager = PagedKVCacheManager(
FakeModel(dtypes=(ttnn.bfloat8_b, ttnn.bfloat16)),
cache_config(dtype=ttnn.bfloat8_b),
)
assert manager.per_layer_dtypes == (ttnn.bfloat8_b, ttnn.bfloat16)
manager.validate_vllm_cache_spec(block_size=32, dtype=torch.bfloat16)
def test_allocate_derives_shapes_and_dtypes_binds_exact_borrowed_handle(fake_allocator, expect_error):
allocated, _ = fake_allocator
model = FakeModel(dtypes=(ttnn.bfloat8_b, ttnn.bfloat16))
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
cache = manager.allocate()
assert cache is model.bound_cache
assert manager.cache_shapes == ((4, 4, 32, 16), (4, 4, 32, 16))
assert [entry[0].dtype for entry in allocated] == [
ttnn.bfloat8_b,
ttnn.bfloat8_b,
ttnn.bfloat16,
ttnn.bfloat16,
]
assert all(tuple(entry[1].shape) == (4, 4, 32, 16) for entry in allocated)
assert len({id(entry[1]) for entry in allocated}) == 1
assert all(entry[2]["memory_config"] == ttnn.DRAM_MEMORY_CONFIG for entry in allocated)
assert all(entry[2]["cache_file_name"] is None for entry in allocated)
manager.validate_borrowed_handle(cache)
with expect_error(ValueError, "exact manager-owned"):
manager.validate_borrowed_handle([pair[:] for pair in cache])
assert manager.bound_context.config is manager.config
assert manager.bound_context.tensors[0][0] is cache[0][0]
def test_allocate_reuses_legacy_cache_files_when_dtype_is_unambiguous(fake_allocator, tmp_path):
allocated, _ = fake_allocator
model = FakeModel()
model.model_args = SimpleNamespace(model_cache_path=tmp_path)
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
manager.allocate()
shape = (4, 4, 32, 16)
expected = [
tmp_path / f"empty_kcache_paged_attention{shape}",
tmp_path / f"empty_vcache_paged_attention{shape}",
]
assert [entry[2]["cache_file_name"] for entry in allocated] == expected * 2
assert len({id(entry[1]) for entry in allocated}) == 1
def test_allocate_avoids_cache_file_collision_for_nonuniform_device_dtypes(fake_allocator, tmp_path):
allocated, _ = fake_allocator
model = FakeModel(dtypes=(ttnn.bfloat8_b, ttnn.bfloat16))
model.model_args = SimpleNamespace(model_cache_path=tmp_path)
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
manager.allocate()
assert all(entry[2]["cache_file_name"] is None for entry in allocated)
def test_allocation_requires_resolved_capacity_and_happens_once(fake_allocator, expect_error):
model = FakeModel()
manager = PagedKVCacheManager(model, cache_config())
with expect_error(RuntimeError, "must be resolved"):
manager.allocate()
manager.configure(replace(manager.config, num_blocks=4))
cache = manager.allocate()
with expect_error(RuntimeError, "already been allocated"):
manager.allocate()
assert cache is model.bound_cache
def test_partial_allocation_failure_deallocates_created_tensors(monkeypatch, expect_error):
first = FakeTensor((4, 4, 32, 16), ttnn.bfloat8_b)
calls = 0
deallocated = []
def fail_second_allocation(
host_tensor,
*,
device,
mesh_mapper,
layout,
memory_config,
dtype,
cache_file_name,
):
nonlocal calls
calls += 1
if calls == 2:
raise RuntimeError("allocation failed")
return first
monkeypatch.setattr(ttnn, "as_tensor", fail_second_allocation)
monkeypatch.setattr(ttnn, "ReplicateTensorToMesh", lambda mesh: None)
monkeypatch.setattr(ttnn, "deallocate", lambda tensor: deallocated.append(tensor))
model = FakeModel()
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
with expect_error(RuntimeError, "allocation failed"):
manager.allocate()
assert deallocated == [first]
assert model.set_calls == []
def test_bind_failure_unbinds_then_deallocates_all_tensors(fake_allocator, expect_error):
allocated, deallocated = fake_allocator
model = FakeModel(bind_fails=True)
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
with expect_error(RuntimeError, "bind failed"):
manager.allocate()
assert model.set_calls[-1] is None
assert deallocated == [entry[0] for entry in reversed(allocated)]
assert model.bound_cache is None
def test_release_unbinds_before_deallocating_and_is_idempotent(monkeypatch, expect_error):
operations = []
model = FakeModel()
original_set = model.set_kv_cache
def set_kv_cache(cache):
operations.append(("bind", cache))
original_set(cache)
model.set_kv_cache = set_kv_cache
monkeypatch.setattr(
ttnn,
"as_tensor",
lambda host_tensor, **kwargs: FakeTensor(host_tensor.shape, kwargs["dtype"]),
)
monkeypatch.setattr(ttnn, "ReplicateTensorToMesh", lambda mesh: None)
monkeypatch.setattr(ttnn, "deallocate", lambda tensor: operations.append(("deallocate", tensor)))
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
cache = manager.allocate()
operations.clear()
manager.release()
assert operations[0] == ("bind", None)
assert [operation[1] for operation in operations[1:]] == [tensor for pair in cache for tensor in pair]
assert manager.bound_context is None
manager.release()
assert len(operations) == 5
with expect_error(RuntimeError, "terminal"):
manager.allocate()
def test_borrowed_handle_mutation_cannot_redirect_owned_tensor_release(fake_allocator, expect_error):
allocated, deallocated = fake_allocator
model = FakeModel()
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
cache = manager.allocate()
owned_tensors = [entry[0] for entry in allocated]
replacement = FakeTensor(cache[0][0].shape, cache[0][0].dtype)
cache[0][0] = replacement
with expect_error(ValueError, "exact manager-owned K/V tensors"):
manager.validate_borrowed_handle(cache)
manager.release()
assert deallocated == owned_tensors
assert replacement not in deallocated
def test_release_failure_retains_only_failed_tensor_and_retries(monkeypatch, expect_error):
tensors = []
def as_tensor(host_tensor, **kwargs):
tensor = FakeTensor(host_tensor.shape, kwargs["dtype"])
tensors.append(tensor)
return tensor
monkeypatch.setattr(ttnn, "as_tensor", as_tensor)
monkeypatch.setattr(ttnn, "ReplicateTensorToMesh", lambda mesh: None)
model = FakeModel()
manager = PagedKVCacheManager(model, cache_config(num_blocks=4))
manager.allocate()
failed_tensor = tensors[0]
attempts = []
def fail_once(tensor):
attempts.append(tensor)
if tensor is failed_tensor and attempts.count(tensor) == 1:
raise RuntimeError("deallocate failed")
monkeypatch.setattr(ttnn, "deallocate", fail_once)
with expect_error(RuntimeError, "Failed to deallocate 1"):
manager.release()
assert manager.bound_context is None
assert model.bound_cache is None
assert attempts == tensors
manager.release()
assert attempts.count(failed_tensor) == 2
assert all(attempts.count(tensor) == 1 for tensor in tensors[1:])
def test_partial_allocation_cleanup_failure_preserves_tensor_for_release_retry(monkeypatch, expect_error):
tensor = FakeTensor((4, 4, 32, 16), ttnn.bfloat8_b)
allocation_calls = 0
deallocation_calls = 0
def fail_second_allocation(
host_tensor,
*,
device,
mesh_mapper,
layout,
memory_config,
dtype,
cache_file_name,
):
nonlocal allocation_calls
allocation_calls += 1
if allocation_calls == 2:
raise RuntimeError("allocation failed")
return tensor
def fail_first_deallocation(value):
nonlocal deallocation_calls
deallocation_calls += 1
if deallocation_calls == 1:
raise RuntimeError("cleanup failed")
monkeypatch.setattr(ttnn, "as_tensor", fail_second_allocation)
monkeypatch.setattr(ttnn, "ReplicateTensorToMesh", lambda mesh: None)
monkeypatch.setattr(ttnn, "deallocate", fail_first_deallocation)
manager = PagedKVCacheManager(FakeModel(), cache_config(num_blocks=4))
with expect_error(RuntimeError, "allocation failed") as exc_info:
manager.allocate()
assert [str(error) for error in exc_info.value.cleanup_failures] == ["cleanup failed"]
manager.release()
assert deallocation_calls == 2
with expect_error(RuntimeError, "terminal"):
manager.allocate()
def test_release_before_allocation_is_terminal_and_idempotent(expect_error):
manager = PagedKVCacheManager(FakeModel(), cache_config())
manager.release()
manager.release()
with expect_error(RuntimeError, "terminal"):
manager.allocate()
|