lIlBrother
commited on
Commit
β’
c251804
1
Parent(s):
f8d0c20
Update: model
Browse files- config.json +1 -1
- {global_step23940 β global_step165430}/mp_rank_00_model_states.pt +2 -2
- {global_step23940 β global_step165430}/zero_pp_rank_0_mp_rank_00_optim_states.pt +1 -1
- {global_step23940 β global_step165430}/zero_pp_rank_1_mp_rank_00_optim_states.pt +1 -1
- {global_step23940 β global_step165430}/zero_pp_rank_2_mp_rank_00_optim_states.pt +1 -1
- {global_step23940 β global_step165430}/zero_pp_rank_3_mp_rank_00_optim_states.pt +1 -1
- latest +1 -1
- zero_to_fp32.py +54 -37
config.json
CHANGED
@@ -24,7 +24,7 @@
|
|
24 |
"position_buckets": 256,
|
25 |
"relative_attention": true,
|
26 |
"share_att_key": true,
|
27 |
-
"transformers_version": "4.
|
28 |
"type_vocab_size": 0,
|
29 |
"vocab_size": 64100
|
30 |
}
|
|
|
24 |
"position_buckets": 256,
|
25 |
"relative_attention": true,
|
26 |
"share_att_key": true,
|
27 |
+
"transformers_version": "4.38.2",
|
28 |
"type_vocab_size": 0,
|
29 |
"vocab_size": 64100
|
30 |
}
|
{global_step23940 β global_step165430}/mp_rank_00_model_states.pt
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5b04e95e0d0b4bb47fea41cd9de0af4ddfe4a9e3350ec64d4c1f8b77de9f9541
|
3 |
+
size 1077570796
|
{global_step23940 β global_step165430}/zero_pp_rank_0_mp_rank_00_optim_states.pt
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 808085192
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4fe7887eab40a10ededf92b4403561338b9b7d9e0269623913f2dcbada17cc00
|
3 |
size 808085192
|
{global_step23940 β global_step165430}/zero_pp_rank_1_mp_rank_00_optim_states.pt
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 808095752
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9046b582e4329869f31dc4562371217f8ee62dd267816ac7ba57990bb808be1e
|
3 |
size 808095752
|
{global_step23940 β global_step165430}/zero_pp_rank_2_mp_rank_00_optim_states.pt
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 808085064
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:091d4c4fd84894549e6e66fadb7699e8c700d6e532ebca6feba502eda8d5d047
|
3 |
size 808085064
|
{global_step23940 β global_step165430}/zero_pp_rank_3_mp_rank_00_optim_states.pt
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
size 808095496
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1470545a59094d545cbc35cc96811ce3765d7a602fc6f93739adf6295b514415
|
3 |
size 808095496
|
latest
CHANGED
@@ -1 +1 @@
|
|
1 |
-
|
|
|
1 |
+
global_step165430
|
zero_to_fp32.py
CHANGED
@@ -24,9 +24,18 @@ from dataclasses import dataclass
|
|
24 |
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
25 |
# DeepSpeed data structures it has to be available in the current python environment.
|
26 |
from deepspeed.utils import logger
|
27 |
-
from deepspeed.checkpoint.constants import (
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
|
32 |
@dataclass
|
@@ -42,7 +51,7 @@ class zero_model_state:
|
|
42 |
debug = 0
|
43 |
|
44 |
# load to cpu
|
45 |
-
device = torch.device(
|
46 |
|
47 |
|
48 |
def atoi(text):
|
@@ -50,12 +59,12 @@ def atoi(text):
|
|
50 |
|
51 |
|
52 |
def natural_keys(text):
|
53 |
-
|
54 |
alist.sort(key=natural_keys) sorts in human order
|
55 |
http://nedbatchelder.com/blog/200712/human_sorting.html
|
56 |
(See Toothy's implementation in the comments)
|
57 |
-
|
58 |
-
return [atoi(c) for c in re.split(r
|
59 |
|
60 |
|
61 |
def get_model_state_file(checkpoint_dir, zero_stage):
|
@@ -127,12 +136,14 @@ def parse_model_states(files):
|
|
127 |
|
128 |
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
129 |
|
130 |
-
z_model_state = zero_model_state(
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
|
|
|
|
136 |
zero_model_states.append(z_model_state)
|
137 |
|
138 |
return zero_model_states
|
@@ -208,7 +219,7 @@ def _get_fp32_state_dict_from_zero_checkpoint(ds_checkpoint_dir):
|
|
208 |
model_files = get_model_state_files(ds_checkpoint_dir)
|
209 |
|
210 |
zero_model_states = parse_model_states(model_files)
|
211 |
-
print(f
|
212 |
|
213 |
if zero_stage <= 2:
|
214 |
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
@@ -225,13 +236,13 @@ def _zero2_merge_frozen_params(state_dict, zero_model_states):
|
|
225 |
|
226 |
if debug:
|
227 |
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
228 |
-
print(f
|
229 |
|
230 |
wanted_params = len(frozen_param_shapes)
|
231 |
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
232 |
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
233 |
-
print(f
|
234 |
-
print(f
|
235 |
|
236 |
total_params = 0
|
237 |
total_numel = 0
|
@@ -273,7 +284,8 @@ def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero
|
|
273 |
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
274 |
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
275 |
avail_numel = sum(
|
276 |
-
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups]
|
|
|
277 |
|
278 |
if debug:
|
279 |
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
@@ -292,7 +304,7 @@ def _zero2_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero
|
|
292 |
avail_numel = full_single_fp32_vector.numel()
|
293 |
for name, shape in shapes.items():
|
294 |
|
295 |
-
unpartitioned_numel = shape.numel() if _has_callable(shape,
|
296 |
total_numel += unpartitioned_numel
|
297 |
total_params += 1
|
298 |
|
@@ -361,14 +373,14 @@ def _zero3_merge_frozen_params(state_dict, world_size, zero_model_states):
|
|
361 |
if debug:
|
362 |
for i in range(world_size):
|
363 |
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
364 |
-
print(f
|
365 |
|
366 |
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
367 |
wanted_params = len(frozen_param_shapes)
|
368 |
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
369 |
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
370 |
-
print(f
|
371 |
-
print(f
|
372 |
|
373 |
total_params = 0
|
374 |
total_numel = 0
|
@@ -430,9 +442,11 @@ def _zero3_merge_trainable_params(state_dict, world_size, fp32_flat_groups, zero
|
|
430 |
)
|
431 |
|
432 |
# XXX: memory usage doubles here
|
433 |
-
state_dict[name] =
|
434 |
-
tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)),
|
435 |
-
|
|
|
|
|
436 |
offset += partitioned_numel
|
437 |
|
438 |
offset *= world_size
|
@@ -499,9 +513,9 @@ def get_fp32_state_dict_from_zero_checkpoint(checkpoint_dir, tag=None):
|
|
499 |
|
500 |
"""
|
501 |
if tag is None:
|
502 |
-
latest_path = os.path.join(checkpoint_dir,
|
503 |
if os.path.isfile(latest_path):
|
504 |
-
with open(latest_path,
|
505 |
tag = fd.read().strip()
|
506 |
else:
|
507 |
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
@@ -572,19 +586,22 @@ def load_state_dict_from_zero_checkpoint(model, checkpoint_dir, tag=None):
|
|
572 |
if __name__ == "__main__":
|
573 |
|
574 |
parser = argparse.ArgumentParser()
|
575 |
-
parser.add_argument(
|
576 |
-
|
577 |
-
|
578 |
parser.add_argument(
|
579 |
"output_file",
|
580 |
type=str,
|
581 |
-
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)"
|
582 |
-
|
583 |
-
|
584 |
-
|
585 |
-
|
586 |
-
|
587 |
-
|
|
|
|
|
|
|
588 |
args = parser.parse_args()
|
589 |
|
590 |
debug = args.debug
|
|
|
24 |
# while this script doesn't use deepspeed to recover data, since the checkpoints are pickled with
|
25 |
# DeepSpeed data structures it has to be available in the current python environment.
|
26 |
from deepspeed.utils import logger
|
27 |
+
from deepspeed.checkpoint.constants import (
|
28 |
+
DS_VERSION,
|
29 |
+
OPTIMIZER_STATE_DICT,
|
30 |
+
SINGLE_PARTITION_OF_FP32_GROUPS,
|
31 |
+
FP32_FLAT_GROUPS,
|
32 |
+
ZERO_STAGE,
|
33 |
+
PARTITION_COUNT,
|
34 |
+
PARAM_SHAPES,
|
35 |
+
BUFFER_NAMES,
|
36 |
+
FROZEN_PARAM_SHAPES,
|
37 |
+
FROZEN_PARAM_FRAGMENTS,
|
38 |
+
)
|
39 |
|
40 |
|
41 |
@dataclass
|
|
|
51 |
debug = 0
|
52 |
|
53 |
# load to cpu
|
54 |
+
device = torch.device("cpu")
|
55 |
|
56 |
|
57 |
def atoi(text):
|
|
|
59 |
|
60 |
|
61 |
def natural_keys(text):
|
62 |
+
"""
|
63 |
alist.sort(key=natural_keys) sorts in human order
|
64 |
http://nedbatchelder.com/blog/200712/human_sorting.html
|
65 |
(See Toothy's implementation in the comments)
|
66 |
+
"""
|
67 |
+
return [atoi(c) for c in re.split(r"(\d+)", text)]
|
68 |
|
69 |
|
70 |
def get_model_state_file(checkpoint_dir, zero_stage):
|
|
|
136 |
|
137 |
frozen_param_fragments = state_dict.get(FROZEN_PARAM_FRAGMENTS, None)
|
138 |
|
139 |
+
z_model_state = zero_model_state(
|
140 |
+
buffers=buffers,
|
141 |
+
param_shapes=param_shapes,
|
142 |
+
shared_params=shared_params,
|
143 |
+
ds_version=ds_version,
|
144 |
+
frozen_param_shapes=frozen_param_shapes,
|
145 |
+
frozen_param_fragments=frozen_param_fragments,
|
146 |
+
)
|
147 |
zero_model_states.append(z_model_state)
|
148 |
|
149 |
return zero_model_states
|
|
|
219 |
model_files = get_model_state_files(ds_checkpoint_dir)
|
220 |
|
221 |
zero_model_states = parse_model_states(model_files)
|
222 |
+
print(f"Parsing checkpoint created by deepspeed=={zero_model_states[0].ds_version}")
|
223 |
|
224 |
if zero_stage <= 2:
|
225 |
return _get_fp32_state_dict_from_zero2_checkpoint(world_size, fp32_flat_groups, zero_model_states)
|
|
|
236 |
|
237 |
if debug:
|
238 |
num_elem = sum(s.numel() for s in frozen_param_shapes.values())
|
239 |
+
print(f"rank 0: {FROZEN_PARAM_SHAPES}.numel = {num_elem}")
|
240 |
|
241 |
wanted_params = len(frozen_param_shapes)
|
242 |
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
243 |
avail_numel = sum([p.numel() for p in frozen_param_fragments.values()])
|
244 |
+
print(f"Frozen params: Have {avail_numel} numels to process.")
|
245 |
+
print(f"Frozen params: Need {wanted_numel} numels in {wanted_params} params")
|
246 |
|
247 |
total_params = 0
|
248 |
total_numel = 0
|
|
|
284 |
full_single_fp32_vector = torch.cat(merged_partitions, 0)
|
285 |
merged_single_partition_of_fp32_groups.append(full_single_fp32_vector)
|
286 |
avail_numel = sum(
|
287 |
+
[full_single_fp32_vector.numel() for full_single_fp32_vector in merged_single_partition_of_fp32_groups]
|
288 |
+
)
|
289 |
|
290 |
if debug:
|
291 |
wanted_params = sum([len(shapes) for shapes in param_shapes])
|
|
|
304 |
avail_numel = full_single_fp32_vector.numel()
|
305 |
for name, shape in shapes.items():
|
306 |
|
307 |
+
unpartitioned_numel = shape.numel() if _has_callable(shape, "numel") else math.prod(shape)
|
308 |
total_numel += unpartitioned_numel
|
309 |
total_params += 1
|
310 |
|
|
|
373 |
if debug:
|
374 |
for i in range(world_size):
|
375 |
num_elem = sum(s.numel() for s in zero_model_states[i].frozen_param_fragments.values())
|
376 |
+
print(f"rank {i}: {FROZEN_PARAM_SHAPES}.numel = {num_elem}")
|
377 |
|
378 |
frozen_param_shapes = zero_model_states[0].frozen_param_shapes
|
379 |
wanted_params = len(frozen_param_shapes)
|
380 |
wanted_numel = sum(s.numel() for s in frozen_param_shapes.values())
|
381 |
avail_numel = sum([p.numel() for p in zero_model_states[0].frozen_param_fragments.values()]) * world_size
|
382 |
+
print(f"Frozen params: Have {avail_numel} numels to process.")
|
383 |
+
print(f"Frozen params: Need {wanted_numel} numels in {wanted_params} params")
|
384 |
|
385 |
total_params = 0
|
386 |
total_numel = 0
|
|
|
442 |
)
|
443 |
|
444 |
# XXX: memory usage doubles here
|
445 |
+
state_dict[name] = (
|
446 |
+
torch.cat(tuple(fp32_flat_groups[i].narrow(0, offset, partitioned_numel) for i in range(world_size)), 0)
|
447 |
+
.narrow(0, 0, unpartitioned_numel)
|
448 |
+
.view(shape)
|
449 |
+
)
|
450 |
offset += partitioned_numel
|
451 |
|
452 |
offset *= world_size
|
|
|
513 |
|
514 |
"""
|
515 |
if tag is None:
|
516 |
+
latest_path = os.path.join(checkpoint_dir, "latest")
|
517 |
if os.path.isfile(latest_path):
|
518 |
+
with open(latest_path, "r") as fd:
|
519 |
tag = fd.read().strip()
|
520 |
else:
|
521 |
raise ValueError(f"Unable to find 'latest' file at {latest_path}")
|
|
|
586 |
if __name__ == "__main__":
|
587 |
|
588 |
parser = argparse.ArgumentParser()
|
589 |
+
parser.add_argument(
|
590 |
+
"checkpoint_dir", type=str, help="path to the desired checkpoint folder, e.g., path/checkpoint-12"
|
591 |
+
)
|
592 |
parser.add_argument(
|
593 |
"output_file",
|
594 |
type=str,
|
595 |
+
help="path to the pytorch fp32 state_dict output file (e.g. path/checkpoint-12/pytorch_model.bin)",
|
596 |
+
)
|
597 |
+
parser.add_argument(
|
598 |
+
"-t",
|
599 |
+
"--tag",
|
600 |
+
type=str,
|
601 |
+
default=None,
|
602 |
+
help="checkpoint tag used as a unique identifier for checkpoint. e.g., global_step1",
|
603 |
+
)
|
604 |
+
parser.add_argument("-d", "--debug", action="store_true", help="enable debug")
|
605 |
args = parser.parse_args()
|
606 |
|
607 |
debug = args.debug
|