File size: 8,536 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 | # SPDX-FileCopyrightText: © 2026 Tenstorrent AI ULC
# SPDX-License-Identifier: Apache-2.0
from types import SimpleNamespace
from models.common.modules.lazy_buffer import LazyBuffer
from models.common.modules.sampling import sampling_1d
from models.common.modules.sampling.sampling_1d import Sampling1D
from models.common.sampling.tt_log_probs import LogProbsCalculator
class FakeTensor:
def __init__(self, name):
self.name = name
class FakeLogProbsCalculator:
instances = []
def __init__(self, *_args): # External construction hook; intentionally generic.
self.buffer = FakeTensor(f"log-probs-{len(self.instances)}")
self.instances.append(self)
def release(self):
if self.buffer is not None:
sampling_1d.ttnn.deallocate(self.buffer)
self.buffer = None
def _lazy_buffer(name):
return LazyBuffer(
source=name,
dtype=object(),
layout=object(),
device=object(),
mesh_mapper=object(),
)
def _sampler_config(index_offsets, owned_specs):
return SimpleNamespace(
index_offsets=index_offsets,
invalid_vocab_mask=owned_specs.get("invalid_vocab_mask"),
invalid_vocab_tail_mask=owned_specs.get("invalid_vocab_tail_mask"),
invalid_vocab_tail_width=32,
seeds=owned_specs["seeds"],
user_ids=owned_specs["user_ids"],
mesh_device=object(),
tt_ccl=None,
sub_core_grids=None,
start_core=None,
max_batch_size=32,
is_resolved=lambda: True,
)
def test_sampling_release_is_idempotent_preserves_borrowed_and_allows_reload(monkeypatch):
deallocated = []
allocations = []
# ttnn.from_torch is overloaded; retain backend-specific keyword absorption.
def from_torch(source, **_kwargs):
tensor = FakeTensor(f"{source}-{len(allocations)}")
allocations.append(tensor)
return tensor
monkeypatch.setattr(sampling_1d.ttnn, "Tensor", FakeTensor)
monkeypatch.setattr(sampling_1d.ttnn, "from_torch", from_torch)
monkeypatch.setattr(sampling_1d.ttnn, "deallocate", deallocated.append)
from models.common import utils as common_utils
FakeLogProbsCalculator.instances = []
monkeypatch.setattr(common_utils, "LogProbsCalculator", FakeLogProbsCalculator)
borrowed_index_offsets = FakeTensor("borrowed-index-offsets")
owned_specs = {
"invalid_vocab_mask": _lazy_buffer("invalid-vocab-mask"),
"invalid_vocab_tail_mask": _lazy_buffer("invalid-vocab-tail-mask"),
"seeds": _lazy_buffer("seeds"),
"user_ids": _lazy_buffer("user-ids"),
}
config = _sampler_config(borrowed_index_offsets, owned_specs)
sampler = object.__new__(Sampling1D)
sampler.config = config
sampler._device_buffers_loaded = False
sampler.load_device_buffers()
first_owned = {name: getattr(sampler, f"_{name}") for name in owned_specs}
first_calculator = sampler._log_probs_calculator
assert sampler._index_offsets is borrowed_index_offsets
sampler.release()
deallocations_after_first_release = list(deallocated)
sampler.release()
assert deallocated == deallocations_after_first_release
assert borrowed_index_offsets not in deallocated
assert set(first_owned.values()).issubset(deallocated)
assert first_calculator.buffer is None
assert not sampler._device_buffers_loaded
assert all(spec._value is None for spec in owned_specs.values())
sampler.load_device_buffers()
assert sampler._device_buffers_loaded
assert sampler._index_offsets is borrowed_index_offsets
assert sampler._log_probs_calculator is not first_calculator
for name, old_tensor in first_owned.items():
assert getattr(sampler, f"_{name}") is not old_tensor
def test_partial_load_failure_preserves_primary_and_releases_before_reload(monkeypatch, expect_error):
allocations = []
deallocated = []
allocation_error = RuntimeError("seed allocation failed")
cleanup_error = RuntimeError("mask cleanup failed once")
fail_allocation = True
fail_cleanup = True
def from_torch(source, **_kwargs):
if source == "seeds" and fail_allocation:
raise allocation_error
tensor = FakeTensor(f"{source}-{len(allocations)}")
allocations.append(tensor)
return tensor
def deallocate(value):
nonlocal fail_cleanup
deallocated.append(value)
if value.name.startswith("invalid-vocab-mask") and fail_cleanup:
fail_cleanup = False
raise cleanup_error
monkeypatch.setattr(sampling_1d.ttnn, "Tensor", FakeTensor)
monkeypatch.setattr(sampling_1d.ttnn, "from_torch", from_torch)
monkeypatch.setattr(sampling_1d.ttnn, "deallocate", deallocate)
from models.common import utils as common_utils
FakeLogProbsCalculator.instances = []
monkeypatch.setattr(common_utils, "LogProbsCalculator", FakeLogProbsCalculator)
borrowed_index_offsets = FakeTensor("borrowed-index-offsets")
owned_specs = {
"invalid_vocab_mask": _lazy_buffer("invalid-vocab-mask"),
"seeds": _lazy_buffer("seeds"),
"user_ids": _lazy_buffer("user-ids"),
}
sampler = object.__new__(Sampling1D)
sampler.config = _sampler_config(borrowed_index_offsets, owned_specs)
sampler._device_buffers_loaded = False
with expect_error(RuntimeError, "seed allocation failed") as caught:
sampler.load_device_buffers()
assert caught.value is allocation_error
assert cleanup_error in caught.value.cleanup_failures
assert owned_specs["invalid_vocab_mask"]._value is not None
assert not sampler._device_buffers_loaded
assert borrowed_index_offsets not in deallocated
sampler.release()
assert owned_specs["invalid_vocab_mask"]._value is None
fail_allocation = False
sampler.load_device_buffers()
assert sampler._device_buffers_loaded
sampler.release()
def test_sampling_release_is_best_effort_and_retries_only_failed_buffer(monkeypatch, expect_error):
attempts = []
cleanup_error = RuntimeError("seed cleanup failed once")
monkeypatch.setattr(sampling_1d.ttnn, "Tensor", FakeTensor)
monkeypatch.setattr(
sampling_1d.ttnn,
"from_torch",
lambda source, **_kwargs: FakeTensor(source),
)
from models.common import utils as common_utils
FakeLogProbsCalculator.instances = []
monkeypatch.setattr(common_utils, "LogProbsCalculator", FakeLogProbsCalculator)
borrowed_index_offsets = FakeTensor("borrowed-index-offsets")
owned_specs = {
"seeds": _lazy_buffer("seeds"),
"user_ids": _lazy_buffer("user-ids"),
}
sampler = object.__new__(Sampling1D)
sampler.config = _sampler_config(borrowed_index_offsets, owned_specs)
sampler._device_buffers_loaded = False
sampler.load_device_buffers()
failed = sampler._seeds
def deallocate(value):
attempts.append(value)
if value is failed and attempts.count(value) == 1:
raise cleanup_error
monkeypatch.setattr(sampling_1d.ttnn, "deallocate", deallocate)
with expect_error(RuntimeError, "seed cleanup failed once") as caught:
sampler.release()
assert caught.value is cleanup_error
assert sampler._seeds is failed
assert owned_specs["seeds"]._value is failed
assert owned_specs["user_ids"]._value is None
sampler.release()
assert attempts.count(failed) == 2
assert attempts.count(borrowed_index_offsets) == 0
assert owned_specs["seeds"]._value is None
def test_log_probs_calculator_release_deallocates_unique_owned_tensors_once(monkeypatch):
deallocated = []
monkeypatch.setattr(sampling_1d.ttnn, "deallocate", deallocated.append)
shared = FakeTensor("shared")
tensors = {
"global_max": shared,
"global_exp_sum": FakeTensor("global-exp-sum"),
"mask": FakeTensor("mask"),
"output_tensor": shared,
"topk_logprobs_output": FakeTensor("topk-logprobs"),
"topk_indices_output": FakeTensor("topk-indices"),
}
calculator = object.__new__(LogProbsCalculator)
for name, tensor in tensors.items():
setattr(calculator, name, tensor)
calculator.release()
calculator.release()
assert len(deallocated) == len({id(tensor) for tensor in tensors.values()})
assert {id(tensor) for tensor in deallocated} == {id(tensor) for tensor in tensors.values()}
assert all(getattr(calculator, name) is None for name in tensors)
|