DLcomm 0.3.1__tar.gz → 0.3.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {dlcomm-0.3.1 → dlcomm-0.3.2}/PKG-INFO +1 -1
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/comm/collectives.py +10 -10
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/dl_comm_main.py +6 -6
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/PKG-INFO +1 -1
- {dlcomm-0.3.1 → dlcomm-0.3.2}/pyproject.toml +1 -1
- {dlcomm-0.3.1 → dlcomm-0.3.2}/README.md +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/analysis/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/analysis/bandwidth.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/analysis/ccl_parser.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/analysis/correctness.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/comm/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/comm/comm_setup.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/config_spec.json +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/mpi_utils.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/setup_algorithms.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/system_info.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/config/validation.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/timer/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/timer/timer.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/utils/__init__.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dl_comm/utils/utility.py +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/SOURCES.txt +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/dependency_links.txt +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/entry_points.txt +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/requires.txt +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/dlcomm.egg-info/top_level.txt +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/setup.cfg +0 -0
- {dlcomm-0.3.1 → dlcomm-0.3.2}/tests/test_mxm_unitrace.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: dlcomm
|
|
3
|
-
Version: 0.3.
|
|
3
|
+
Version: 0.3.2
|
|
4
4
|
Summary: Distributed GPU Communication Benchmarking Framework for Deep Learning
|
|
5
5
|
Author-email: Musa Cim <musaoguzhancim@gmail.com>, Kaushik Velusamy <kaushikvelusamy@gmail.com>
|
|
6
6
|
Project-URL: Homepage, https://github.com/argonne-lcf/DLcomm_benchmark
|
|
@@ -49,7 +49,7 @@ def register_collective(name: str, needs_op: bool = False):
|
|
|
49
49
|
return decorator
|
|
50
50
|
|
|
51
51
|
@register_collective("allreduce", needs_op=True)
|
|
52
|
-
def _allreduce(
|
|
52
|
+
def _allreduce(tensor, op, group=None, dist=None, log=None, framework="pytorch"):
|
|
53
53
|
if framework == 'pytorch':
|
|
54
54
|
dist.all_reduce(tensor, op=op, group=group)
|
|
55
55
|
elif framework == 'jax':
|
|
@@ -57,7 +57,7 @@ def _allreduce(framework, tensor, op, group=None, dist=None, log=None):
|
|
|
57
57
|
|
|
58
58
|
|
|
59
59
|
@register_collective("reduce", needs_op=True)
|
|
60
|
-
def _reduce(
|
|
60
|
+
def _reduce(tensor, op, group=None, dist=None,log=None, framework="pytorch"):
|
|
61
61
|
if framework == 'pytorch':
|
|
62
62
|
if group is None:
|
|
63
63
|
smallest_rank = 0
|
|
@@ -70,7 +70,7 @@ def _reduce(framework, tensor, op, group=None, dist=None,log=None ):
|
|
|
70
70
|
|
|
71
71
|
|
|
72
72
|
@register_collective("broadcast", needs_op=False)
|
|
73
|
-
def _broadcast(
|
|
73
|
+
def _broadcast(tensor, op, group=None, dist=None, log=None, framework="pytorch"):
|
|
74
74
|
if framework == 'pytorch':
|
|
75
75
|
if group is None:
|
|
76
76
|
smallest_rank = 0
|
|
@@ -84,7 +84,7 @@ def _broadcast(framework, tensor, op, group=None, dist=None, log=None ):
|
|
|
84
84
|
|
|
85
85
|
|
|
86
86
|
@register_collective("allgather", needs_op=False)
|
|
87
|
-
def _allgather(
|
|
87
|
+
def _allgather(tensor, op=None, group=None, dist=None, log=None, framework="pytorch"):
|
|
88
88
|
if framework == 'pytorch':
|
|
89
89
|
world_size = dist.get_world_size(group)
|
|
90
90
|
tensor_list = [torch.empty_like(tensor) for _ in range(world_size)]
|
|
@@ -94,7 +94,7 @@ def _allgather(framework, tensor , op=None, group=None, dist=None, log=None ):
|
|
|
94
94
|
pass
|
|
95
95
|
|
|
96
96
|
@register_collective("gather", needs_op=False)
|
|
97
|
-
def _gather(
|
|
97
|
+
def _gather(tensor, op=None, group=None, dist=None, log=None, framework="pytorch"):
|
|
98
98
|
if framework == 'pytorch':
|
|
99
99
|
if group is None:
|
|
100
100
|
smallest_rank = 0
|
|
@@ -116,7 +116,7 @@ def _gather(framework, tensor, op=None, group=None, dist=None, log=None):
|
|
|
116
116
|
|
|
117
117
|
|
|
118
118
|
@register_collective("scatter", needs_op=False)
|
|
119
|
-
def _scatter(
|
|
119
|
+
def _scatter(tensor, op=None, group=None, dist=None,log=None, framework="pytorch"):
|
|
120
120
|
if framework == 'pytorch':
|
|
121
121
|
if group is None:
|
|
122
122
|
smallest_rank = 0
|
|
@@ -137,7 +137,7 @@ def _scatter(framework, tensor, op=None, group=None, dist=None,log=None):
|
|
|
137
137
|
|
|
138
138
|
|
|
139
139
|
@register_collective("reducescatter", needs_op=True)
|
|
140
|
-
def _reduce_scatter(
|
|
140
|
+
def _reduce_scatter(tensor, op, group=None, dist=None,log=None, framework="pytorch"):
|
|
141
141
|
if framework == 'pytorch':
|
|
142
142
|
world_size = dist.get_world_size(group)
|
|
143
143
|
|
|
@@ -160,7 +160,7 @@ def _reduce_scatter(framework, tensor, op, group=None, dist=None,log=None):
|
|
|
160
160
|
|
|
161
161
|
|
|
162
162
|
@register_collective("alltoall", needs_op=False)
|
|
163
|
-
def _all_to_all(
|
|
163
|
+
def _all_to_all(tensor, op=None, group=None, dist=None,log=None, framework="pytorch"):
|
|
164
164
|
if framework == 'pytorch':
|
|
165
165
|
world_size = dist.get_world_size(group)
|
|
166
166
|
|
|
@@ -175,7 +175,7 @@ def _all_to_all(framework, tensor, op=None, group=None, dist=None,log=None):
|
|
|
175
175
|
|
|
176
176
|
|
|
177
177
|
@register_collective("alltoallsingle", needs_op=False)
|
|
178
|
-
def _all_to_all_single(
|
|
178
|
+
def _all_to_all_single(tensor, op=None, group=None, dist=None, log=None, framework="pytorch"):
|
|
179
179
|
if framework == 'pytorch':
|
|
180
180
|
output_tensor = torch.empty_like(tensor)
|
|
181
181
|
dist.all_to_all_single(output_tensor, tensor, group=group)
|
|
@@ -185,7 +185,7 @@ def _all_to_all_single(framework, tensor, op=None, group=None, dist=None, log=No
|
|
|
185
185
|
|
|
186
186
|
|
|
187
187
|
@register_collective("barrier", needs_op=False)
|
|
188
|
-
def _barrier(
|
|
188
|
+
def _barrier(tensor, op=None, group=None, dist=None,log=None, framework="pytorch"):
|
|
189
189
|
if framework == 'pytorch':
|
|
190
190
|
dist.barrier(group=group)
|
|
191
191
|
elif framework == 'jax':
|
|
@@ -412,15 +412,15 @@ def main(cfg: DictConfig):
|
|
|
412
412
|
|
|
413
413
|
if comm_mode == "flatview":
|
|
414
414
|
if flat_group is not None and participating:
|
|
415
|
-
result = run_collective(
|
|
415
|
+
result = run_collective(x, op_obj, group=flat_group, dist=dist, framework=framework)
|
|
416
416
|
|
|
417
417
|
elif comm_mode == "within_node":
|
|
418
418
|
if my_within_group is not None and participating:
|
|
419
|
-
result = run_collective(
|
|
419
|
+
result = run_collective(x, op_obj, group=my_within_group, dist=dist, log=log, framework=framework)
|
|
420
420
|
|
|
421
421
|
elif comm_mode == "across_node":
|
|
422
422
|
if my_across_group is not None and participating:
|
|
423
|
-
result = run_collective(
|
|
423
|
+
result = run_collective(x, op_obj, group=my_across_group, dist=dist, log=log, framework=framework)
|
|
424
424
|
|
|
425
425
|
MPI.COMM_WORLD.Barrier()
|
|
426
426
|
if mpi_rank == 0:
|
|
@@ -460,7 +460,7 @@ def main(cfg: DictConfig):
|
|
|
460
460
|
if flat_group is not None:
|
|
461
461
|
time_barrier(group=flat_group, device=device)
|
|
462
462
|
with timer("(Flatview)"):
|
|
463
|
-
result = run_collective(
|
|
463
|
+
result = run_collective(x, op_obj, group=flat_group, dist=dist, framework=framework)
|
|
464
464
|
time_barrier(group=flat_group, device=device)
|
|
465
465
|
if enable_correctness:
|
|
466
466
|
check_collective_correctness(context, x, coll_name, op=op_obj, group=flat_group, result_data=result, group_type="Flatview", group_id="All")
|
|
@@ -469,7 +469,7 @@ def main(cfg: DictConfig):
|
|
|
469
469
|
if my_within_group is not None and participating:
|
|
470
470
|
time_barrier(group=my_within_group, device=device)
|
|
471
471
|
with timer(f"(Within-Group-{within_group_id})"):
|
|
472
|
-
result = run_collective(
|
|
472
|
+
result = run_collective(x, op_obj, group=my_within_group, dist=dist, log=log, framework=framework)
|
|
473
473
|
time_barrier(group=my_within_group, device=device)
|
|
474
474
|
if enable_correctness:
|
|
475
475
|
check_collective_correctness(context, x, coll_name, op=op_obj, group=my_within_group, result_data=result, group_type="Within", group_id=within_group_id)
|
|
@@ -478,7 +478,7 @@ def main(cfg: DictConfig):
|
|
|
478
478
|
if my_across_group is not None:
|
|
479
479
|
time_barrier(group=my_across_group , device=device)
|
|
480
480
|
with timer(f"(Across-Group-{across_group_id})"):
|
|
481
|
-
result = run_collective(
|
|
481
|
+
result = run_collective(x, op_obj, group=my_across_group, dist=dist, log=log, framework=framework)
|
|
482
482
|
time_barrier(group=my_across_group, device=device)
|
|
483
483
|
if enable_correctness:
|
|
484
484
|
check_collective_correctness(context, x, coll_name, op=op_obj, group=my_across_group, result_data=result, group_type="Across", group_id=across_group_id)
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: dlcomm
|
|
3
|
-
Version: 0.3.
|
|
3
|
+
Version: 0.3.2
|
|
4
4
|
Summary: Distributed GPU Communication Benchmarking Framework for Deep Learning
|
|
5
5
|
Author-email: Musa Cim <musaoguzhancim@gmail.com>, Kaushik Velusamy <kaushikvelusamy@gmail.com>
|
|
6
6
|
Project-URL: Homepage, https://github.com/argonne-lcf/DLcomm_benchmark
|
|
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
|
|
|
4
4
|
|
|
5
5
|
[project]
|
|
6
6
|
name = "dlcomm"
|
|
7
|
-
version = "0.3.
|
|
7
|
+
version = "0.3.2"
|
|
8
8
|
description = "Distributed GPU Communication Benchmarking Framework for Deep Learning"
|
|
9
9
|
authors = [
|
|
10
10
|
{ name = "Musa Cim", email = "musaoguzhancim@gmail.com" },
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|
|
File without changes
|