Support method decorator for tagging and add minimalistic comparator in dumper (#19559)

This commit is contained in:
fzyzcjy
2026-02-28 18:04:54 +08:00
committed by GitHub
parent 9bf3638a25
commit 706ab9296a
6 changed files with 551 additions and 90 deletions
@@ -332,7 +332,6 @@ class TestEntrypointGroupingRaw:
enable=True,
dir=str(side_dir),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
dumper.__dict__["_static_meta"] = {"world_rank": 0, "world_size": 1}
@@ -1110,7 +1109,6 @@ class TestEntrypointAlignment:
enable=True,
dir=str(d),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
@@ -1167,7 +1165,6 @@ class TestEntrypointAlignment:
enable=True,
dir=str(sglang_dir),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
@@ -1226,7 +1223,6 @@ class TestEntrypointAlignment:
enable=True,
dir=str(megatron_dir),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
@@ -1468,9 +1464,7 @@ def _assert_single_comparison_passed(records: list[AnyRecord]) -> ComparisonReco
def _make_dumper(directory: Path) -> _Dumper:
return _Dumper(
config=DumperConfig(enable=True, dir=str(directory), enable_http_server=False)
)
return _Dumper(config=DumperConfig(enable=True, dir=str(directory)))
def _create_dumps(
@@ -1528,7 +1522,6 @@ def _create_non_tensor_rank_dump(
enable=True,
dir=str(directory),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
dumper.__dict__["_static_meta"] = {"world_rank": rank, "world_size": 1}
@@ -1611,7 +1604,6 @@ def _create_rank_dump(
enable=True,
dir=str(directory),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
@@ -1951,7 +1943,6 @@ class TestEntrypointThdCpZigzag:
enable=True,
dir=str(sglang_dir),
exp_name=_FIXED_EXP_NAME,
enable_http_server=False,
)
)
@@ -403,79 +403,5 @@ class TestAlignerPlanInComparisonRecord:
assert "unsharder" in text
def _make_aligner_plan() -> AlignerPlan:
unsharder = UnsharderPlan(
axis=ParallelAxis.TP,
params=ConcatParams(dim_name="h"),
groups=[[0, 1]],
)
return AlignerPlan(
per_step_plans=Pair(
x=[
AlignerPerStepPlan(
step=0, input_object_indices=[0, 1], sub_plans=[unsharder]
)
],
y=[
AlignerPerStepPlan(
step=0, input_object_indices=[0, 1], sub_plans=[unsharder]
)
],
),
)
class TestAlignerPlanInComparisonRecord:
def test_comparison_record_with_aligner_plan(self) -> None:
plan: AlignerPlan = _make_aligner_plan()
record: ComparisonRecord = _make_comparison_record(
diff=_make_diff_info(passed=True),
)
record_with_plan = record.model_copy(update={"aligner_plan": plan})
assert record_with_plan.aligner_plan is not None
assert record_with_plan.aligner_plan.per_step_plans.x[0].step == 0
def test_aligner_plan_json_roundtrip(self) -> None:
plan: AlignerPlan = _make_aligner_plan()
record: ComparisonRecord = _make_comparison_record(
diff=_make_diff_info(passed=True),
)
record_with_plan = record.model_copy(update={"aligner_plan": plan})
json_str: str = record_with_plan.model_dump_json()
parsed = json.loads(json_str)
assert "aligner_plan" in parsed
assert (
parsed["aligner_plan"]["per_step_plans"]["x"][0]["sub_plans"][0]["type"]
== "unsharder"
)
roundtripped: ComparisonRecord = parse_record_json(json_str)
assert roundtripped.aligner_plan is not None
assert (
roundtripped.aligner_plan.per_step_plans.x[0].sub_plans[0].type
== "unsharder"
)
def test_comparison_record_without_aligner_plan(self) -> None:
record: ComparisonRecord = _make_comparison_record(
diff=_make_diff_info(passed=True),
)
json_str: str = record.model_dump_json()
roundtripped: ComparisonRecord = parse_record_json(json_str)
assert roundtripped.aligner_plan is None
def test_aligner_plan_text_format(self) -> None:
plan: AlignerPlan = _make_aligner_plan()
record: ComparisonRecord = _make_comparison_record(
diff=_make_diff_info(passed=True),
)
record_with_plan = record.model_copy(update={"aligner_plan": plan})
text: str = record_with_plan.to_text()
assert "Aligner Plan:" in text
assert "unsharder" in text
if __name__ == "__main__":
sys.exit(pytest.main([__file__]))
@@ -0,0 +1,159 @@
from argparse import Namespace
from pathlib import Path
import pytest
import torch
from sglang.srt.debug_utils.dump_comparator import (
_argmax_coord,
_calc_rel_diff,
_compute_smaller_dtype,
_try_unify_shape,
main,
)
from sglang.srt.debug_utils.dumper import DumperConfig, _Dumper
from sglang.test.ci.ci_register import register_cpu_ci
register_cpu_ci(est_time=30, suite="default", nightly=True)
# ----------------------------- Unit tests -----------------------------
class TestCalcRelDiff:
def test_identical_vectors(self) -> None:
x: torch.Tensor = torch.randn(10, 10)
assert _calc_rel_diff(x, x).item() == pytest.approx(0.0, abs=1e-5)
def test_zero_vectors(self) -> None:
z: torch.Tensor = torch.zeros(5)
result = _calc_rel_diff(z, z)
assert not torch.isnan(result) or True # should not crash
class TestArgmaxCoord:
def test_known_position(self) -> None:
x: torch.Tensor = torch.zeros(2, 3, 4)
x[1, 2, 3] = 10.0
assert _argmax_coord(x) == (1, 2, 3)
class TestTryUnifyShape:
def test_squeeze_leading_ones(self) -> None:
target_shape: torch.Size = torch.Size([3, 4])
result: torch.Tensor = _try_unify_shape(torch.randn(1, 1, 3, 4), target_shape)
assert result.shape == target_shape
def test_no_op_when_no_leading_ones(self) -> None:
target_shape: torch.Size = torch.Size([3, 4])
result: torch.Tensor = _try_unify_shape(torch.randn(2, 3, 4), target_shape)
assert result.shape == (2, 3, 4)
class TestComputeSmallerDtype:
def test_known_pair(self) -> None:
assert _compute_smaller_dtype(torch.float32, torch.bfloat16) == torch.bfloat16
assert _compute_smaller_dtype(torch.bfloat16, torch.float32) == torch.bfloat16
def test_none_for_same_dtype(self) -> None:
assert _compute_smaller_dtype(torch.float32, torch.float32) is None
# ----------------------------- Integration tests -----------------------------
def _make_dumper(directory: Path) -> _Dumper:
return _Dumper(
config=DumperConfig(
enable=True,
dir=str(directory),
)
)
def _create_dumps(
tmp_path: Path,
tensor_names: list[str],
*,
baseline_names: list[str] | None = None,
) -> tuple[Path, Path]:
if baseline_names is None:
baseline_names = tensor_names
d_baseline: Path = tmp_path / "baseline"
d_target: Path = tmp_path / "target"
d_baseline.mkdir()
d_target.mkdir()
torch.manual_seed(42)
baseline_tensor: torch.Tensor = torch.randn(10, 10)
target_tensor: torch.Tensor = baseline_tensor + torch.randn(10, 10) * 0.01
exp_paths: list[Path] = []
for d, names, tensor in [
(d_baseline, baseline_names, baseline_tensor),
(d_target, tensor_names, target_tensor),
]:
dumper: _Dumper = _make_dumper(d)
for name in names:
dumper.dump(name, tensor)
dumper.step()
exp_paths.append(d / dumper._config.exp_name)
return exp_paths[0], exp_paths[1]
def _make_args(
baseline_path: Path,
target_path: Path,
*,
filter_pattern: str | None = None,
) -> Namespace:
return Namespace(
baseline_path=str(baseline_path),
target_path=str(target_path),
start_step=0,
end_step=1000000,
diff_threshold=1e-3,
filter=filter_pattern,
)
class TestMainBasic:
def test_matching_tensors(
self, tmp_path: Path, capsys: pytest.CaptureFixture
) -> None:
baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
args: Namespace = _make_args(baseline_path, target_path)
main(args)
captured: str = capsys.readouterr().out
assert "✅" in captured
def test_with_filter(self, tmp_path: Path, capsys: pytest.CaptureFixture) -> None:
baseline_path, target_path = _create_dumps(tmp_path, ["tensor_a", "tensor_b"])
args: Namespace = _make_args(
baseline_path, target_path, filter_pattern="tensor_a"
)
main(args)
captured: str = capsys.readouterr().out
assert "tensor_a" in captured
assert "Check:" in captured
def test_no_match_skips(
self, tmp_path: Path, capsys: pytest.CaptureFixture
) -> None:
baseline_path, target_path = _create_dumps(
tmp_path,
["only_in_target"],
baseline_names=["only_in_baseline"],
)
args: Namespace = _make_args(baseline_path, target_path)
main(args)
captured: str = capsys.readouterr().out
assert "Skip" in captured
+58 -5
View File
@@ -451,7 +451,6 @@ class TestDumperDistributed:
config=DumperConfig(
enable=True,
collective_timeout=3,
enable_http_server=False,
),
)
@@ -663,12 +662,11 @@ class TestDumpDictFormat:
def _make_test_dumper(tmp_path, **overrides) -> _Dumper:
"""Create a _Dumper for CPU testing without HTTP server or distributed."""
"""Create a _Dumper for CPU testing without distributed."""
defaults = dict(
enable=True,
dir=str(tmp_path),
exp_name="test",
enable_http_server=False,
)
defaults.update(overrides)
config = DumperConfig(**defaults)
@@ -2293,7 +2291,6 @@ class TestDumperDims:
config=DumperConfig(
enable=True,
dir=str(tmp_path),
enable_http_server=False,
enable_grad=True,
)
)
@@ -2323,7 +2320,6 @@ class TestDumperDims:
config=DumperConfig(
enable=True,
dir=str(tmp_path),
enable_http_server=False,
enable_grad=True,
)
)
@@ -2340,5 +2336,62 @@ class TestDumperDims:
assert grad_data["meta"]["dims"] == "b h(tp)"
class TestCtxDecorator:
def test_ctx_dynamic_lambda(self, tmp_path: Path) -> None:
d = _make_test_dumper(tmp_path)
class FakeLayer:
def __init__(self, layer_id: int) -> None:
self.layer_id = layer_id
@d.ctx(lambda self: dict(layer_id=self.layer_id))
def forward(self, x: torch.Tensor) -> torch.Tensor:
d.dump("hidden", x)
return x
layer = FakeLayer(layer_id=42)
layer.forward(torch.randn(3))
filenames = _get_filenames(tmp_path)
_assert_files(filenames, exist=["layer_id=42"])
def test_ctx_static_kwargs(self, tmp_path: Path) -> None:
d = _make_test_dumper(tmp_path)
@d.ctx(phase="decode")
def decode_step(x: torch.Tensor) -> torch.Tensor:
d.dump("step_out", x)
return x
decode_step(torch.randn(3))
filenames = _get_filenames(tmp_path)
_assert_files(filenames, exist=["phase=decode"])
def test_ctx_clears_on_exception(self, tmp_path: Path) -> None:
d = _make_test_dumper(tmp_path)
@d.ctx(phase="train")
def buggy_fn() -> None:
raise RuntimeError("boom")
with pytest.raises(RuntimeError, match="boom"):
buggy_fn()
assert d._state.global_ctx == {}
def test_ctx_rejects_mixed_args(self) -> None:
d = _make_test_dumper("/tmp")
with pytest.raises(ValueError, match="cannot mix"):
d.ctx(lambda self: dict(a=1), phase="x")
def test_ctx_rejects_empty_args(self) -> None:
d = _make_test_dumper("/tmp")
with pytest.raises(ValueError, match="must provide"):
d.ctx()
if __name__ == "__main__":
sys.exit(pytest.main([__file__]))