Enhance diff and tensor-info logging in dumper grafter (#24512)

This commit is contained in:
fzyzcjy
2026-05-06 16:58:08 +08:00
committed by GitHub
parent 75943cfbcf
commit 8527db0a91
2 changed files with 68 additions and 2 deletions
@@ -17,6 +17,7 @@ import torch.distributed as dist
from sglang.srt.debug_utils.dumper import (
DumperConfig,
_collective_with_timeout,
_compare_tensors_quick,
_deepcopy_or_clone,
_detect_recompute_status,
_Dumper,
@@ -397,6 +398,37 @@ class TestLog:
assert ", t=" in out, out
class TestCompareTensorsQuick:
def test_identical(self):
a = torch.tensor([1.0, 2.0, 3.0])
s = _compare_tensors_quick(a, a.clone())
assert "rel_diff=0" in s, s
assert "max_abs=0" in s, s
def test_diverged(self):
a = torch.tensor([1.0, 2.0, 3.0])
b = torch.tensor([1.0, 2.0, 4.0]) # last element differs by 1
s = _compare_tensors_quick(a, b)
assert "max_abs=1" in s, s
assert "rel_diff=" in s, s
def test_shape_mismatch(self):
s = _compare_tensors_quick(torch.zeros(3), torch.zeros(4))
assert "shape mismatch" in s, s
def test_dtype_unified(self):
s = _compare_tensors_quick(
torch.zeros(3, dtype=torch.float32),
torch.zeros(3, dtype=torch.float64),
)
assert "rel_diff=" in s, s
assert "max_abs=" in s, s
def test_empty(self):
s = _compare_tensors_quick(torch.zeros(0), torch.zeros(0))
assert s == "empty"
class TestCollectiveTimeout:
def test_watchdog_fires_on_timeout(self):
block_event = threading.Event()