8000 Fix collective schedule logging and runtime tests (#160260) · pytorch/pytorch@fc80f68 · GitHub
[go: up one dir, main page]
Skip to content

Commit fc80f68

Browse files
skarjalapytorchmergebot
skarjala
authored and
pytorchmergebot
committed
Fix collective schedule logging and runtime tests (#160260)
Summary: - Fix collective schedule logging so that only logs when collectives present - Fix runtime estimate test to check if each op has a number value Pull Request resolved: #160260 Approved by: https://github.com/Skylion007
1 parent cf0a0dc commit fc80f68

File tree

2 files changed

+12
-29
lines changed
8000

2 files changed

+12
-29
lines changed

‎test/dynamo/test_structured_trace.py‎

Lines changed: 9 additions & 28 deletions
Original file line numberDiff line numberDiff line change
@@ -260,7 +260,6 @@ def test_schedule(self):
260260
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
261261
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
262262
{"artifact": {"name": "triton_kernel_info", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
263-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
264263
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
265264
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
266265
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
@@ -293,7 +292,6 @@ def test_cudagraphs(self):
293292
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
294293
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
295294
{"artifact": {"name": "triton_kernel_info", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
296-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
297295
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
298296
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
299297
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
@@ -333,7 +331,6 @@ def fn(x, y):
333331
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
334332
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
335333
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
336-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
337334
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
338335
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
339336
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
@@ -354,7 +351,6 @@ def fn(x, y):
354351
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
355352
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
356353
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
357-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
358354
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
359355
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 1, "attempt": 0, "has_payload": "HASH"}
360356
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 1, "attempt": 0}
@@ -385,7 +381,6 @@ def test_example_fn(self):
385381
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
386382
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
387383
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
388-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
389384
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
390385
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
391386
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
@@ -443,7 +438,6 @@ def test_example_training_fn(self):
443438
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
444439
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
445440
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
446-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
447441
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
448442
{"dynamo_cpp_guards_str": {}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
449443
{"compilation_metrics": "METRICS", "frame_id": 2, "frame_compile_id": 0, "attempt": 1}
@@ -453,7 +447,6 @@ def test_example_training_fn(self):
453447
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
454448
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
455449
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
456-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
457450
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 2, "frame_compile_id": 0, "attempt": 1, "has_payload": "HASH"}
458451
{"bwd_compilation_metrics": "METRICS", "frame_id": 2, "frame_compile_id": 0, "attempt": 1}
459452
{"dynamo_start": {"stack": "STACK"}, "frame_id": 4, "frame_compile_id": 0, "attempt": 0}
@@ -678,7 +671,6 @@ def forward(self, x):
678671
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
679672
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
680673
{"inductor_output_code": {"filename": "FILENAME"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
681-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
682674
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
683675
{"describe_storage": {"id": 16, "describer_id": "ID", "size": 4194304}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0}
684676
{"describe_tensor": {"id": 29, "ndim": 2, "dtype": "torch.float32", "device": "device(type='cuda', index=0)", "size": [1024, 1024], "is_leaf": true, "requires_grad": true, "is_parameter": true, "stride": [1024, 1], "storage": 16, "view_func": "VIEW_FUNC", "describer_id": "ID"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0}
@@ -698,7 +690,6 @@ def forward(self, x):
698690
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
699691
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
700692
{"inductor_output_code": {"filename": "FILENAME"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
701-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
702693
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
703694
{"dynamo_cpp_guards_str": {}, "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
704695
{"compilation_metrics": "METRICS", "rank": 0, "frame_id": 4, "frame_compile_id": 0, "attempt": 0}
@@ -738,7 +729,6 @@ def fn(x):
738729
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
739730
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
740731
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
741-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
742732
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
743733
{"dynamo_cpp_guards_str": {}, "frame_id": 1, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
744734
{"compilation_metrics": "METRICS", "frame_id": 1, "frame_compile_id": 0, "attempt": 0}
@@ -898,7 +888,6 @@ def fn(a):
898888
{"artifact": {"name": "before_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
899889
{"artifact": {"name": "after_post_grad_graph", "encoding": "string"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
900890
{"inductor_output_code": {"filename": "FILENAME"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
901-
{"artifact": {"name": "inductor_collective_schedule", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
902891
{"artifact": {"name": "fx_graph_cache_miss", "encoding": "json"}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
903892
{"dynamo_cpp_guards_str": {}, "frame_id": 0, "frame_compile_id": 0, "attempt": 0, "has_payload": "HASH"}
904893
{"compilation_metrics": "METRICS", "frame_id": 0, "frame_compile_id": 0, "attempt": 0}
@@ -1159,9 +1148,9 @@ def test_collective_schedule_empty(self):
11591148

11601149
log_collective_schedule([])
11611150

1162-
self.assertIn('"inductor_collective_schedule"', self.buffer.getvalue())
1163-
self.assertEqual(json.loads(payload_buffer.getvalue()), [])
1164-
self.assertParses()
1151+
# With no collectives, artifact should not be logged and payload should be empty
1152+
self.assertNotIn('"inductor_collective_schedule"', self.buffer.getvalue())
1153+
self.assertEqual(payload_buffer.getvalue().strip(), "")
11651154

11661155
@requires_tlparse
11671156
@requires_distributed()
@@ -1271,14 +1260,10 @@ def forward(self, x):
12711260

12721261
self.assertTrue(len(compute_ops) > 0 or len(collective_ops) > 0)
12731262

1274-
# All ops should have runtime > 0 except wait_tensor can be 0
1263+
# Just check each op has an estimated runtime value (any value, including 0)
12751264
for op in ops:
1276-
if "wait_tensor" not in op["name"]:
1277-
self.assertGreater(
1278-
op["estimated_runtime_ns"],
1279-
0,
1280-
f"Op {op['name']} should have runtime > 0",
1281-
)
1265+
self.assertIn("estimated_runtime_ns", op)
1266+
self.assertIsNotNone(op["estimated_runtime_ns"])
12821267

12831268
self.assertParses()
12841269
finally:
@@ -1339,14 +1324,10 @@ def forward(self, x):
13391324
self.assertIn("compute", op_types)
13401325
self.assertIn("collective", op_types)
13411326

1342-
# All ops should have runtime > 0 except wait_tensor can be 0
1327+
# Just check each op has an estimated runtime value (any value, including 0)
13431328
for op in ops:
1344-
if "wait_tensor" not in op["name"]:
1345-
self.assertGreater(
1346-
op["estimated_runtime_ns"],
1347-
0,
1348-
f"Op {op['name']} should have runtime > 0",
1349-
)
1329+
self.assertIn("estimated_runtime_ns", op)
1330+
self.assertIsNotNone(op["estimated_runtime_ns"])
13501331

13511332
self.assertParses()
13521333
finally:

‎torch/_inductor/debug.py‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -719,7 +719,9 @@ def log_collective_schedule(nodes: Sequence[BaseSchedulerNode]) -> None:
719719
if isinstance(op := getattr(node, "node", None), ir._CollectiveKernel)
720720
]
721721

722-
_dump_collective_schedule(schedule)
722+
# Only log when there is at least one collective op
723+
if schedule:
724+
_dump_collective_schedule(schedule)
723725

724726

725727
def log_runtime_estimates(node_runtimes: Sequence[tuple[Any, float]]) -> None:

0 commit comments

Comments
 (0)
0