File size: 11,052 Bytes
a402b9b | 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 | import sys
import pytest
import torch
from sglang.srt.debug_utils.comparator.dp_utils import (
_extract_dp_info,
_group_has_data,
filter_to_non_empty_dp_rank,
)
from sglang.srt.debug_utils.dump_loader import ValueWithMeta
from sglang.test.ci.ci_register import register_cpu_ci
register_cpu_ci(est_time=15, suite="default", nightly=True)
def _make_sglang_meta(
*, tp_rank: int = 0, tp_size: int = 1, dp_rank: int = 0, dp_size: int = 1
) -> dict:
return {
"sglang_parallel_info": {
"tp_rank": tp_rank,
"tp_size": tp_size,
"dp_rank": dp_rank,
"dp_size": dp_size,
}
}
def _make_megatron_meta(
*, tp_rank: int = 0, tp_size: int = 1, dp_rank: int = 0, dp_size: int = 1
) -> dict:
return {
"megatron_parallel_info": {
"tp_rank": tp_rank,
"tp_size": tp_size,
"dp_rank": dp_rank,
"dp_size": dp_size,
}
}
def _make_item(value: object, meta: dict) -> ValueWithMeta:
return ValueWithMeta(value=value, meta=meta)
# ---------------------------------------------------------------------------
# _extract_dp_info
# ---------------------------------------------------------------------------
class TestExtractDpInfo:
def test_sglang_dp(self) -> None:
meta: dict = _make_sglang_meta(dp_rank=1, dp_size=4)
assert _extract_dp_info(meta) == (1, 4)
def test_megatron_dp(self) -> None:
meta: dict = _make_megatron_meta(dp_rank=2, dp_size=8)
assert _extract_dp_info(meta) == (2, 8)
def test_no_parallel_info(self) -> None:
assert _extract_dp_info({}) is None
def test_no_dp_fields(self) -> None:
meta: dict = {"sglang_parallel_info": {"tp_rank": 0, "tp_size": 2}}
assert _extract_dp_info(meta) is None
# ---------------------------------------------------------------------------
# _group_has_data
# ---------------------------------------------------------------------------
class TestGroupHasData:
def test_non_empty_tensor(self) -> None:
item: ValueWithMeta = _make_item(value=torch.tensor([1, 2, 3]), meta={})
assert _group_has_data([item]) is True
def test_empty_tensor(self) -> None:
item: ValueWithMeta = _make_item(value=torch.tensor([]), meta={})
assert _group_has_data([item]) is False
def test_non_tensor_value(self) -> None:
item: ValueWithMeta = _make_item(value="hello", meta={})
assert _group_has_data([item]) is False
def test_empty_group(self) -> None:
assert _group_has_data([]) is False
# ---------------------------------------------------------------------------
# filter_to_non_empty_dp_rank
# ---------------------------------------------------------------------------
class TestFilterToNonEmptyDpRank:
def test_dp_size_1_returns_unchanged(self) -> None:
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0]),
meta=_make_sglang_meta(dp_size=1),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert result is items
def test_no_parallel_info_returns_unchanged(self) -> None:
items: list[ValueWithMeta] = [
_make_item(value=torch.tensor([1.0]), meta={}),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert result is items
def test_empty_list_returns_empty(self) -> None:
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank([])
assert result == []
def test_dp2_all_non_tensor_returns_unchanged(self) -> None:
"""DP=2 with non-tensor values: skip filtering, return unchanged."""
items: list[ValueWithMeta] = [
_make_item(
value=["req_A"],
meta=_make_sglang_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=["req_A"],
meta=_make_sglang_meta(dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert result is items
def test_dp2_one_empty_one_nonempty_sglang(self) -> None:
"""DP=2, rank 0 has data, rank 1 has empty tensor."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0, 2.0]),
meta=_make_sglang_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([]),
meta=_make_sglang_meta(dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert len(result) == 1
assert torch.equal(result[0].value, torch.tensor([1.0, 2.0]))
def test_dp2_one_empty_one_nonempty_megatron(self) -> None:
"""DP=2 megatron, rank 1 has data, rank 0 has empty tensor."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([]),
meta=_make_megatron_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([3.0, 4.0]),
meta=_make_megatron_meta(dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert len(result) == 1
assert torch.equal(result[0].value, torch.tensor([3.0, 4.0]))
def test_dp2_both_nonempty_raises(self) -> None:
"""DP=2, both ranks have data: assertion error."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0]),
meta=_make_sglang_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([2.0]),
meta=_make_sglang_meta(dp_rank=1, dp_size=2),
),
]
with pytest.raises(
AssertionError, match="Expected exactly 1 non-empty dp_rank"
):
filter_to_non_empty_dp_rank(items)
def test_dp2_with_tp2_filters_correctly(self) -> None:
"""DP=2 x TP=2: 4 items total, 2 non-empty from dp_rank=0."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0]),
meta=_make_sglang_meta(tp_rank=0, tp_size=2, dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([2.0]),
meta=_make_sglang_meta(tp_rank=1, tp_size=2, dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([]),
meta=_make_sglang_meta(tp_rank=0, tp_size=2, dp_rank=1, dp_size=2),
),
_make_item(
value=torch.tensor([]),
meta=_make_sglang_meta(tp_rank=1, tp_size=2, dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(items)
assert len(result) == 2
assert torch.equal(result[0].value, torch.tensor([1.0]))
assert torch.equal(result[1].value, torch.tensor([2.0]))
# ---------------------------------------------------------------------------
# dp_group_alias tests
# ---------------------------------------------------------------------------
class TestExtractDpInfoWithAlias:
def test_alias_found(self) -> None:
meta: dict = {
"sglang_parallel_info": {
"dp_rank": 0,
"dp_size": 2,
"moe_dp_rank": 1,
"moe_dp_size": 4,
}
}
assert _extract_dp_info(meta, dp_group_alias="moe_dp") == (1, 4)
def test_alias_not_found_returns_none(self) -> None:
meta: dict = _make_sglang_meta(dp_rank=0, dp_size=2)
assert _extract_dp_info(meta, dp_group_alias="moe_dp") is None
def test_alias_none_uses_default(self) -> None:
meta: dict = _make_sglang_meta(dp_rank=1, dp_size=4)
assert _extract_dp_info(meta, dp_group_alias=None) == (1, 4)
class TestFilterToNonEmptyDpRankWithAlias:
def test_alias_none_unchanged_behavior(self) -> None:
"""dp_group_alias=None → same behavior as before (regression)."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0, 2.0]),
meta=_make_sglang_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([]),
meta=_make_sglang_meta(dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(
items, dp_group_alias=None
)
assert len(result) == 1
assert torch.equal(result[0].value, torch.tensor([1.0, 2.0]))
def test_alias_group_absent_noop(self) -> None:
"""Alias group not in metadata → noop, return items unchanged."""
items: list[ValueWithMeta] = [
_make_item(
value=torch.tensor([1.0]),
meta=_make_sglang_meta(dp_rank=0, dp_size=2),
),
_make_item(
value=torch.tensor([2.0]),
meta=_make_sglang_meta(dp_rank=1, dp_size=2),
),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(
items, dp_group_alias="moe_dp"
)
assert result is items
def test_alias_size_1_noop(self) -> None:
"""Alias group present but size=1 → noop."""
meta: dict = {
"sglang_parallel_info": {
"dp_rank": 0,
"dp_size": 2,
"moe_dp_rank": 0,
"moe_dp_size": 1,
}
}
items: list[ValueWithMeta] = [
_make_item(value=torch.tensor([1.0]), meta=meta),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(
items, dp_group_alias="moe_dp"
)
assert result is items
def test_alias_filters_correctly(self) -> None:
"""Alias group size=2, one empty rank → correctly filters."""
meta_rank0: dict = {
"sglang_parallel_info": {
"dp_rank": 0,
"dp_size": 2,
"moe_dp_rank": 0,
"moe_dp_size": 2,
}
}
meta_rank1: dict = {
"sglang_parallel_info": {
"dp_rank": 0,
"dp_size": 2,
"moe_dp_rank": 1,
"moe_dp_size": 2,
}
}
items: list[ValueWithMeta] = [
_make_item(value=torch.tensor([1.0, 2.0]), meta=meta_rank0),
_make_item(value=torch.tensor([]), meta=meta_rank1),
]
result: list[ValueWithMeta] = filter_to_non_empty_dp_rank(
items, dp_group_alias="moe_dp"
)
assert len(result) == 1
assert torch.equal(result[0].value, torch.tensor([1.0, 2.0]))
if __name__ == "__main__":
sys.exit(pytest.main([__file__]))
|