Tiny fix single-gpu dumper and add tests for dumper (#16285)

This commit is contained in:
fzyzcjy
2026-01-02 13:30:09 +08:00
committed by GitHub
parent 7254986342
commit 749736ba95
4 changed files with 415 additions and 4 deletions
+7 -4
View File
@@ -250,8 +250,8 @@ class _DumperRpcHandler:
def _create_zmq_rpc_handles(handler, base_port: int) -> Optional[List["_ZmqRpcHandle"]]:
import zmq
rank = dist.get_rank()
world_size = dist.get_world_size()
rank = _get_rank()
world_size = dist.get_world_size() if dist.is_initialized() else 1
port = base_port + rank
local_addr = f"tcp://{_get_local_ip_by_remote()}:{port}"
@@ -274,8 +274,11 @@ def _create_zmq_rpc_handles(handler, base_port: int) -> Optional[List["_ZmqRpcHa
thread.start()
print(f"[Dumper.ZmqRpc] rank={rank} server started at {local_addr}")
all_addresses = [None] * world_size
dist.all_gather_object(all_addresses, local_addr)
if dist.is_initialized():
all_addresses = [None] * world_size
dist.all_gather_object(all_addresses, local_addr)
else:
all_addresses = [local_addr]
print(f"[Dumper.ZmqRpc] rank={rank} all_addresses={all_addresses}")
if rank == 0:
@@ -0,0 +1,136 @@
import os
import tempfile
import unittest
from contextlib import contextmanager
from pathlib import Path
import torch
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
register_cpu_ci(est_time=60, suite="default", nightly=True)
class TestDumpComparator(CustomTestCase):
def test_calc_rel_diff(self):
from sglang.srt.debug_utils.dump_comparator import _calc_rel_diff
x = torch.randn(10, 10)
self.assertAlmostEqual(_calc_rel_diff(x, x).item(), 0.0, places=5)
self.assertAlmostEqual(
_calc_rel_diff(torch.tensor([1.0, 0.0]), torch.tensor([0.0, 1.0])).item(),
1.0,
places=5,
)
def test_argmax_coord(self):
from sglang.srt.debug_utils.dump_comparator import _argmax_coord
x = torch.zeros(2, 3, 4)
x[1, 2, 3] = 10.0
self.assertEqual(_argmax_coord(x), (1, 2, 3))
def test_try_unify_shape(self):
from sglang.srt.debug_utils.dump_comparator import _try_unify_shape
target = torch.Size([3, 4])
self.assertEqual(
_try_unify_shape(torch.randn(1, 1, 3, 4), target).shape, target
)
self.assertEqual(
_try_unify_shape(torch.randn(2, 3, 4), target).shape, (2, 3, 4)
)
def test_compute_smaller_dtype(self):
from sglang.srt.debug_utils.dump_comparator import _compute_smaller_dtype
self.assertEqual(
_compute_smaller_dtype(torch.float32, torch.bfloat16), torch.bfloat16
)
self.assertIsNone(_compute_smaller_dtype(torch.float32, torch.float32))
def test_einops_pattern(self):
from sglang.srt.debug_utils.dump_comparator import (
_get_einops_dim_index,
_split_einops_pattern,
)
self.assertEqual(_split_einops_pattern("a (b c) d"), ["a", "(b c)", "d"])
self.assertEqual(_get_einops_dim_index("a b c", "b"), 1)
def test_load_object(self):
from sglang.srt.debug_utils.dump_comparator import _load_object
with tempfile.TemporaryDirectory() as tmpdir:
path = Path(tmpdir) / "tensor.pt"
torch.save(torch.randn(5, 5), path)
self.assertEqual(_load_object(path).shape, (5, 5))
torch.save({"dict": 1}, path)
self.assertIsNone(_load_object(path))
self.assertIsNone(_load_object("/nonexistent.pt"))
def test_compute_and_print_diff(self):
from sglang.srt.debug_utils.dump_comparator import _compute_and_print_diff
x = torch.ones(10, 10)
self.assertAlmostEqual(
_compute_and_print_diff(x, x, 1e-3)["max_abs_diff"], 0.0, places=5
)
self.assertAlmostEqual(
_compute_and_print_diff(x, x + 0.5, 1e-3)["max_abs_diff"], 0.5, places=4
)
class TestEndToEnd(CustomTestCase):
def test_main(self):
from argparse import Namespace
from sglang.srt.debug_utils.dump_comparator import main
from sglang.srt.debug_utils.dumper import _Dumper
with tempfile.TemporaryDirectory() as d1, tempfile.TemporaryDirectory() as d2:
baseline_tensor = torch.randn(10, 10)
target_tensor = baseline_tensor + torch.randn(10, 10) * 0.01
dump_dirs = []
for d, tensor in [(d1, baseline_tensor), (d2, target_tensor)]:
with _with_env("SGLANG_DUMPER_DIR", d), _with_env(
"SGLANG_DUMPER_SERVER_PORT", "-1"
):
dumper = _Dumper()
dumper.on_forward_pass_start()
dumper.dump("tensor_a", tensor)
dumper.on_forward_pass_start()
dumper.dump("tensor_b", tensor * 2)
dump_dirs.append(Path(d) / f"sglang_dump_{dumper._partial_name}")
args = Namespace(
baseline_path=str(dump_dirs[0]),
target_path=str(dump_dirs[1]),
start_id=1,
end_id=2,
baseline_start_id=1,
diff_threshold=1e-3,
filter=None,
)
main(args)
@contextmanager
def _with_env(name: str, value: str):
old = os.environ.get(name)
os.environ[name] = value
try:
yield
finally:
if old is None:
os.environ.pop(name, None)
else:
os.environ[name] = old
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,67 @@
import tempfile
import unittest
from pathlib import Path
import polars as pl
import torch
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
register_cpu_ci(est_time=30, suite="default", nightly=True)
class TestDumpLoader(CustomTestCase):
def test_read_meta(self):
from sglang.srt.debug_utils.dump_loader import read_meta
with tempfile.TemporaryDirectory() as tmpdir:
for fn in [
"forward_pass_id=1___rank=0___dump_index=1___name=a.pt",
"forward_pass_id=2___rank=0___dump_index=2___name=b.pt",
]:
torch.save(torch.randn(5), Path(tmpdir) / fn)
df = read_meta(tmpdir)
self.assertEqual(len(df), 2)
self.assertTrue(
all(c in df.columns for c in ["forward_pass_id", "rank", "name"])
)
def test_find_row(self):
from sglang.srt.debug_utils.dump_loader import find_row
df = pl.DataFrame({"id": [1, 2], "name": ["a", "b"], "file": ["f1", "f2"]})
self.assertEqual(find_row(df, {"id": 2})["file"], "f2")
self.assertIsNone(find_row(df, {"id": 999}))
df_dup = pl.DataFrame({"id": [1, 1], "file": ["f1", "f2"]})
self.assertIsNone(find_row(df_dup, {"id": 1}))
def test_cast_to_polars_dtype(self):
from sglang.srt.debug_utils.dump_loader import _cast_to_polars_dtype
self.assertEqual(_cast_to_polars_dtype("42", pl.Int64), 42)
self.assertEqual(_cast_to_polars_dtype("3.14", pl.Float64), 3.14)
def test_add_duplicate_index(self):
from sglang.srt.debug_utils.dump_loader import _add_duplicate_index
df = pl.DataFrame(
{
"name": ["a", "a", "b"],
"dump_index": [1, 2, 3],
"filename": ["f1", "f2", "f3"],
}
)
result = _add_duplicate_index(df)
self.assertEqual(
result.filter(pl.col("name") == "a")
.sort("dump_index")["duplicate_index"]
.to_list(),
[0, 1],
)
if __name__ == "__main__":
unittest.main()
+205
View File
@@ -0,0 +1,205 @@
import os
import tempfile
import time
import unittest
from pathlib import Path
import requests
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=30, suite="nightly-2-gpu", nightly=True)
class TestDumperPureFunctions(CustomTestCase):
def test_get_truncated_value(self):
from sglang.srt.debug_utils.dumper import get_truncated_value
self.assertIsNone(get_truncated_value(None))
self.assertEqual(get_truncated_value(42), 42)
self.assertEqual(
len(get_truncated_value((torch.randn(10), torch.randn(20)))), 2
)
self.assertEqual(get_truncated_value(torch.randn(10, 10)).shape, (10, 10))
self.assertEqual(get_truncated_value(torch.randn(100, 100)).shape, (5, 5))
def test_obj_to_dict(self):
from sglang.srt.debug_utils.dumper import _obj_to_dict
self.assertEqual(_obj_to_dict({"a": 1}), {"a": 1})
class Obj:
x, y = 10, 20
def method(self):
pass
result = _obj_to_dict(Obj())
self.assertEqual(result["x"], 10)
self.assertNotIn("method", result)
def test_get_tensor_info(self):
from sglang.srt.debug_utils.dumper import get_tensor_info
info = get_tensor_info(torch.randn(10, 10))
for key in ["shape=", "dtype=", "min=", "max=", "mean="]:
self.assertIn(key, info)
self.assertIn("value=42", get_tensor_info(42))
self.assertIn("min=None", get_tensor_info(torch.tensor([])))
class TestDumperDistributed(CustomTestCase):
def test_basic(self):
with tempfile.TemporaryDirectory(prefix="test_dumper_") as tmpdir:
_run_distributed_test(_test_basic_func, tmpdir=tmpdir)
def test_http_enable(self):
_run_distributed_test(_test_http_func)
def test_filter(self):
with tempfile.TemporaryDirectory(prefix="test_dumper_") as tmpdir:
_run_distributed_test(_test_filter_func, tmpdir=tmpdir)
def test_write_disabled(self):
with tempfile.TemporaryDirectory(prefix="test_dumper_") as tmpdir:
_run_distributed_test(_test_write_disabled_func, tmpdir=tmpdir)
def _test_basic_func(rank, tmpdir):
os.environ["SGLANG_DUMPER_DIR"] = tmpdir
from sglang.srt.debug_utils.dumper import dumper
tensor = torch.randn(10, 10, device=f"cuda:{rank}")
dumper.on_forward_pass_start()
dumper.dump("tensor_a", tensor, arg=100)
dumper.on_forward_pass_start()
dumper.set_ctx(ctx_arg=200)
dumper.dump("tensor_b", tensor)
dumper.set_ctx(ctx_arg=None)
dumper.on_forward_pass_start()
dumper.override_enable(False)
dumper.dump("tensor_skip", tensor)
dumper.override_enable(True)
dumper.on_forward_pass_start()
dumper.dump_dict("obj", {"a": torch.randn(3, device=f"cuda:{rank}"), "b": 42})
dist.barrier()
filenames = _get_filenames(tmpdir)
_assert_files(
filenames,
exist=["tensor_a", "tensor_b", "arg=100", "ctx_arg=200", "obj_a", "obj_b"],
not_exist=["tensor_skip"],
)
def _test_http_func(rank):
os.environ["SGLANG_DUMPER_ENABLE"] = "0"
from sglang.srt.debug_utils.dumper import dumper
assert not dumper._enable
dumper.on_forward_pass_start()
for enable in [True, False]:
dist.barrier()
if rank == 0:
time.sleep(0.1)
requests.post(
"http://localhost:40000/dumper", json={"enable": enable}
).raise_for_status()
dist.barrier()
assert dumper._enable == enable
def _test_filter_func(rank, tmpdir):
os.environ["SGLANG_DUMPER_DIR"] = tmpdir
os.environ["SGLANG_DUMPER_FILTER"] = "keep"
from sglang.srt.debug_utils.dumper import dumper
dumper.on_forward_pass_start()
dumper.dump("keep_this", torch.randn(5, device=f"cuda:{rank}"))
dumper.dump("skip_this", torch.randn(5, device=f"cuda:{rank}"))
dist.barrier()
filenames = _get_filenames(tmpdir)
_assert_files(filenames, exist=["keep_this"], not_exist=["skip_this"])
def _test_write_disabled_func(rank, tmpdir):
os.environ["SGLANG_DUMPER_DIR"] = tmpdir
os.environ["SGLANG_DUMPER_WRITE_FILE"] = "0"
from sglang.srt.debug_utils.dumper import dumper
dumper.on_forward_pass_start()
dumper.dump("no_write", torch.randn(5, device=f"cuda:{rank}"))
dist.barrier()
assert len(_get_filenames(tmpdir)) == 0
def _get_filenames(tmpdir):
return {f.name for f in Path(tmpdir).glob("sglang_dump_*/*.pt")}
def _assert_files(filenames, *, exist=(), not_exist=()):
for p in exist:
assert any(p in f for f in filenames), f"{p} not found in {filenames}"
for p in not_exist:
assert not any(
p in f for f in filenames
), f"{p} should not exist in {filenames}"
def _run_distributed_test(func, world_size=2, **kwargs):
ctx = mp.get_context("spawn")
result_queue = ctx.Queue()
processes = []
for rank in range(world_size):
p = ctx.Process(
target=_run_worker, args=(rank, world_size, func, result_queue, kwargs)
)
p.start()
processes.append(p)
for p in processes:
p.join()
errors = [result_queue.get() for _ in range(world_size)]
errors = [e for e in errors if e]
if errors:
raise AssertionError("\n".join(errors))
def _run_worker(rank, world_size, func, result_queue, kwargs):
os.environ.update(
MASTER_ADDR="localhost",
MASTER_PORT="29500",
RANK=str(rank),
WORLD_SIZE=str(world_size),
)
torch.cuda.set_device(rank)
dist.init_process_group(backend="nccl", rank=rank, world_size=world_size)
try:
func(rank, **kwargs)
result_queue.put(None)
except Exception as e:
import traceback
result_queue.put(f"Rank {rank}: {e}\n{traceback.format_exc()}")
finally:
dist.destroy_process_group()
if __name__ == "__main__":
unittest.main()