File size: 13,661 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 | # SPDX-FileCopyrightText: © 2026 Tenstorrent AI ULC
# SPDX-License-Identifier: Apache-2.0
import ast
import subprocess
import sys
from pathlib import Path
import pytest
import torch
from models.common.modules.sampling.params import (
PreparedSamplingParams,
place_prepared_sampling_params,
prepare_sampling_params,
slice_prepared_sampling_params,
slice_sampling_params,
)
from models.common.sampling.sampling_params import SamplingParams
def _prepare(params, *, batch_size=4, max_device_top_k=32, allow_force_argmax=True):
return prepare_sampling_params(
params,
batch_size,
max_device_top_k=max_device_top_k,
allow_force_argmax=allow_force_argmax,
)
@pytest.mark.parametrize("top_k", [1, 32])
def test_stochastic_top_k_device_boundaries_are_preserved_exactly(top_k):
prepared = _prepare(SamplingParams(temperature=0.7, top_k=top_k, top_p=0.9))
assert isinstance(prepared, PreparedSamplingParams)
assert prepared.top_k[0] == top_k
assert prepared.sampling_path == "topk"
@pytest.mark.parametrize("top_k", [0, -1, 33, 50, 128256])
def test_unsupported_stochastic_top_k_raises_instead_of_clamping(top_k, expect_error):
with expect_error(ValueError, "route this request to host sampling"):
_prepare(SamplingParams(temperature=0.7, top_k=top_k, top_p=0.9))
@pytest.mark.parametrize("top_k", [-1, 0, 33, 50, 128256])
def test_greedy_normalization_precedes_top_k_validation(top_k):
prepared = _prepare(SamplingParams(temperature=0.0, top_k=top_k, top_p=0.2))
assert prepared.top_k[0] == 1
assert prepared.top_p[0] == 0.0
assert prepared.temperature[0] == 1.0
assert prepared.greedy_mask[0] is True
assert prepared.sampling_path == "argmax"
def test_mixed_greedy_and_stochastic_rows_use_batch_wide_topk_path():
prepared = _prepare(
SamplingParams(
temperature=[0.0, 0.5],
top_k=[128256, 32],
top_p=[0.4, 0.8],
)
)
assert prepared.row_paths[:2] == ("argmax", "topk")
assert prepared.sampling_path == "topk"
assert prepared.top_k[:2] == (1, 32)
def test_one_unsupported_stochastic_row_rejects_the_complete_batch(expect_error):
with expect_error(ValueError, r"top_k\[1\]=33"):
_prepare(
SamplingParams(
temperature=[0.0, 0.5],
top_k=[128256, 33],
top_p=[0.4, 0.8],
)
)
def test_prepared_structure_preserves_all_request_owned_sampling_fields():
params = SamplingParams(
temperature=torch.tensor([0.0, 0.5]),
top_k=torch.tensor([128256, 7]),
top_p=torch.tensor([0.6, 0.75]),
presence_penalty=[0.1, 0.2],
frequency_penalty=[0.3, 0.4],
repetition_penalty=[1.1, 1.2],
seed=[None, 19],
enable_log_probs=[False, True],
num_logprobs=[0, 0],
)
prepared = _prepare(params)
assert prepared.presence_penalty[:2] == pytest.approx((0.1, 0.2))
assert prepared.frequency_penalty[:2] == pytest.approx((0.3, 0.4))
assert prepared.repetition_penalty[:2] == pytest.approx((1.1, 1.2))
assert prepared.seeds[:2] == (None, 19)
assert prepared.logprob_modes[:2] == ("none", "sampled_token")
assert prepared.penalties_enabled is True
assert prepared.log_probs_enabled is True
assert torch.equal(params.top_k, torch.tensor([128256, 7]))
def test_greedy_logprob_request_uses_logprob_capable_topk_path():
prepared = _prepare(
SamplingParams(
temperature=0.0,
top_k=128256,
top_p=1.0,
enable_log_probs=True,
num_logprobs=0,
)
)
assert prepared.logprob_modes[0] == "sampled_token"
assert prepared.row_paths[0] == "topk"
assert prepared.sampling_path == "topk"
def test_vllm_disabled_logprob_sentinel_normalizes_without_rejecting_device_sampling():
prepared = _prepare(
SamplingParams(
temperature=[0.0, 0.8],
top_k=[128256, 7],
top_p=[1.0, 0.9],
enable_log_probs=[False, True],
num_logprobs=[-2, 0],
)
)
assert prepared.num_logprobs[:2] == (0, 0)
assert prepared.logprob_modes[:2] == ("none", "sampled_token")
def test_force_argmax_capability_is_explicit_policy():
params = SamplingParams(temperature=0.0, top_k=128256, top_p=1.0)
assert _prepare(params, allow_force_argmax=True).sampling_path == "argmax"
assert _prepare(params, allow_force_argmax=False).sampling_path == "topk"
def test_scalar_seed_stays_request_scoped_and_inactive_rows_use_safe_defaults():
prepared = _prepare(SamplingParams(temperature=0.8, top_k=4, top_p=0.9, seed=123))
assert prepared.seeds == (123, None, None, None)
assert prepared.top_k[1:] == (1, 1, 1)
assert prepared.top_p[1:] == (0.0, 0.0, 0.0)
assert prepared.temperature[1:] == (1.0, 1.0, 1.0)
assert prepared.row_paths[1:] == ("inactive", "inactive", "inactive")
def test_vllm_negative_one_seed_sentinel_is_native_unseeded_state():
prepared = _prepare(
SamplingParams(
temperature=[0.8, 0.8],
top_k=[4, 4],
top_p=[0.9, 0.9],
seed=[-1, 123],
)
)
assert prepared.seeds[:2] == (None, 123)
def test_slice_sampling_params_preserves_field_alignment_without_mutating_input():
params = SamplingParams(
temperature=[0.1, 0.2, 0.3],
top_k=[1, 2, 3],
top_p=[0.4, 0.5, 0.6],
presence_penalty=[0.7, 0.8, 0.9],
seed=[11, None, 33],
enable_log_probs=[False, True, False],
)
sliced = slice_sampling_params(params, [2, 0])
assert sliced.temperature == [0.3, 0.1]
assert sliced.top_k == [3, 1]
assert sliced.top_p == [0.6, 0.4]
assert sliced.presence_penalty == [0.9, 0.7]
assert sliced.seed == [33, 11]
assert sliced.enable_log_probs == [False, False]
assert params.temperature == [0.1, 0.2, 0.3]
def test_slice_sampling_params_broadcasts_scalars_to_every_selected_row():
# A demo-style request: one scalar per field, meant for every active decode slot.
params = SamplingParams(temperature=0.0, top_k=32, top_p=0.08, seed=7)
sliced = slice_sampling_params(params, [4, 0, 9])
assert sliced.temperature == [0.0, 0.0, 0.0]
assert sliced.top_k == [32, 32, 32]
assert sliced.top_p == [0.08, 0.08, 0.08]
assert sliced.presence_penalty == [0.0, 0.0, 0.0]
assert sliced.enable_log_probs == [False, False, False]
# Seeds are request-owned and never broadcast to sibling rows.
assert sliced.seed == 7
assert params.temperature == 0.0
def test_scalar_sampling_request_places_onto_every_active_decode_slot():
# Regression for #55953: qwen3-32b eval-32-perf-report compiles decode for 32 active slots with
# SamplingParams(temperature=0.0, top_k=32, top_p=0.08); the sliced request must describe 32 rows.
slots = tuple(range(32))
request = slice_sampling_params(SamplingParams(temperature=0.0, top_k=32, top_p=0.08), slots)
prepared = prepare_sampling_params(request, 32, max_device_top_k=32, allow_force_argmax=True)
placed = place_prepared_sampling_params(prepared, slots)
assert prepared.active_rows == 32
assert placed.active_mask == tuple([True] * 32)
assert placed.top_k == tuple([1] * 32) # temperature 0 normalizes to greedy rows
assert placed.greedy_mask == tuple([True] * 32)
assert placed.temperature == tuple([1.0] * 32) # greedy rows carry the neutral temperature
def test_prepared_slice_preserves_prompt_output_and_slot_remap_alignment():
prompt_tokens = torch.tensor([[10, 11], [20, 21], [30, 31]])
output_tokens = [[100], [200, 201], [300]]
slot_remap = torch.tensor([2, 0, 1], dtype=torch.int32)
prepared = prepare_sampling_params(
SamplingParams(
temperature=[0.0, 0.5, 0.8],
top_k=[128256, 5, 7],
top_p=[1.0, 0.9, 0.8],
seed=[11, 22, 33],
),
4,
max_device_top_k=32,
allow_force_argmax=True,
prompt_tokens=prompt_tokens,
output_tokens=output_tokens,
slot_remap=slot_remap,
)
lane = slice_prepared_sampling_params(prepared, [2, 0])
assert lane.top_k == (7, 1)
assert lane.seeds == (33, 11)
assert lane.row_paths == ("topk", "argmax")
assert lane.sampling_path == "topk"
assert torch.equal(lane.prompt_tokens, torch.tensor([[30, 31], [10, 11]]))
assert lane.output_tokens == [[300], [100]]
assert torch.equal(lane.slot_remap, torch.tensor([1, 2], dtype=torch.int32))
assert torch.equal(prepared.prompt_tokens, prompt_tokens)
assert prepared.output_tokens is output_tokens
assert prepared.slot_remap is slot_remap
def test_prefill_request_rows_are_placed_into_lane_local_slots_with_history():
prepared = prepare_sampling_params(
SamplingParams(
temperature=[0.0, 0.5],
top_k=[128256, 7],
top_p=[1.0, 0.8],
seed=[11, 22],
repetition_penalty=[1.1, 1.2],
),
4,
max_device_top_k=32,
allow_force_argmax=True,
prompt_tokens=torch.tensor([[10, 11], [20, 21]]),
output_tokens=torch.tensor([[100, -1], [200, 201]]),
)
placed = place_prepared_sampling_params(prepared, [3, 1])
assert placed.active_mask == (False, True, False, True)
assert placed.top_k == (1, 7, 1, 1)
assert placed.seeds == (None, 22, None, 11)
assert placed.row_paths == ("inactive", "topk", "inactive", "argmax")
assert torch.equal(
placed.prompt_tokens,
torch.tensor([[-1, -1], [20, 21], [-1, -1], [10, 11]]),
)
assert torch.equal(
placed.output_tokens,
torch.tensor([[-1, -1], [200, 201], [-1, -1], [100, -1]]),
)
@pytest.mark.parametrize(
("relative_path", "class_name"),
[
("models/common/models/llama3_8b/generator.py", "Llama3Generator"),
("models/common/models/llama33_70b/generator.py", "Llama33_70BGenerator"),
("models/common/models/qwen3_32b/generator.py", "Qwen3_32BGenerator"),
],
)
def test_target_generator_capabilities_advertise_exact_device_top_k(relative_path, class_name):
repository_root = Path(__file__).parents[5]
source = (repository_root / relative_path).read_text(encoding="utf-8")
tree = ast.parse(source, filename=relative_path)
class_node = next(node for node in tree.body if isinstance(node, ast.ClassDef) and node.name == class_name)
assignment = next(
node
for node in class_node.body
if isinstance(node, ast.Assign)
and any(isinstance(target, ast.Name) and target.id == "model_capabilities" for target in node.targets)
)
capabilities = ast.literal_eval(assignment.value)
assert capabilities["supports_sample_on_device"] is True
assert capabilities["max_device_top_k"] == 32
@pytest.mark.parametrize(
"relative_path",
[
"models/tt_transformers/tt/generator_vllm.py",
"models/demos/llama3_70b_galaxy/tt/generator_vllm.py",
],
)
def test_legacy_device_sampling_capabilities_declare_the_exact_limit(relative_path):
tree = ast.parse(Path(relative_path).read_text(encoding="utf-8"), filename=relative_path)
advertised = []
for node in tree.body:
if not isinstance(node, ast.ClassDef):
continue
for statement in node.body:
if not isinstance(statement, ast.Assign):
continue
if not any(
isinstance(target, ast.Name) and target.id == "model_capabilities" for target in statement.targets
):
continue
capabilities = ast.literal_eval(statement.value)
if capabilities.get("supports_sample_on_device"):
advertised.append((node.name, capabilities))
assert advertised
assert all(capabilities.get("max_device_top_k") == 32 for _, capabilities in advertised)
def test_tttv2_runtime_import_boundary_excludes_legacy_sampling_state():
roots = [Path("models/common/llm_runtime")]
files = [path for root in roots for path in root.rglob("*.py")]
files.extend(
Path(path)
for path in (
"models/common/models/llama3_8b/executor.py",
"models/common/models/llama33_70b/executor.py",
"models/common/models/qwen3_32b/executor.py",
)
)
forbidden_modules = {
"models.common.sampling",
"models.common.sampling.generator",
"models.common.sampling.tt_penalties",
}
violations = []
for path in files:
tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
for node in ast.walk(tree):
if isinstance(node, ast.ImportFrom) and node.module in forbidden_modules:
violations.append((str(path), node.lineno, node.module))
assert violations == []
def test_loading_tttv2_runtime_does_not_load_legacy_sampling_modules():
script = """
import importlib
import sys
importlib.import_module('models.common.llm_runtime.decode')
forbidden = {
'models.common.sampling.generator',
'models.common.sampling.tt_penalties',
'models.common.sampling.tt_sampling',
}
loaded = sorted(forbidden.intersection(sys.modules))
if loaded:
raise SystemExit('loaded legacy sampling modules: ' + ', '.join(loaded))
"""
result = subprocess.run(
[sys.executable, "-c", script],
cwd=Path.cwd(),
text=True,
capture_output=True,
check=False,
)
assert result.returncode == 0, result.stdout + result.stderr
|