Skip to content

using container when loading inference checkpoints #2875

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
HeyangQin merged 10 commits into from
Feb 28, 2023
Merged

using container when loading inference checkpoints #2875

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
be50a64
move load_params() from container to policy
HeyangQin Feb 22, 2023
3567dfb
fix formattng
HeyangQin Feb 22, 2023
34446a7
Revert "move load_params() from container to policy"
HeyangQin Feb 22, 2023
0a9dd27
using container when loading checkpoints
HeyangQin Feb 22, 2023
a8b5755
Merge branch 'master' into HeyangQin/fix_issue_2770
HeyangQin Feb 23, 2023
2fe914e
fix format
HeyangQin Feb 23, 2023
48681f6
Merge branch 'master' into HeyangQin/fix_issue_2770
HeyangQin Feb 23, 2023
15c3d82
Merge branch 'master' into HeyangQin/fix_issue_2770
tjruwase Feb 24, 2023
7f2ce3e
Merge branch 'master' into HeyangQin/fix_issue_2770
HeyangQin Feb 27, 2023
4e9035d
Merge branch 'master' into HeyangQin/fix_issue_2770
mrwyattii Feb 28, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions deepspeed/module_inject/containers/bloom.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,8 +55,8 @@ def load_params(self, module, sd, weight_quantizer, mp_replace, prefix):
transformer_param_names[i],
prefix + param_names[i],
qkv=True,
megatron_v2=self.is_megatron_v2,
split_qkv=self.split_qkv)
megatron_v2=self.policy.is_megatron_v2,
split_qkv=self.policy.split_qkv)
for i in range(2, 4):
maybe_copy(module.attention,
sd,
Expand Down
2 changes: 1 addition & 1 deletion deepspeed/module_inject/containers/gptj.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,7 @@ def load_params(self, module, sd, weight_quantizer, mp_replace, prefix):
[prefix + param_names[0],
prefix + param_names[1],
prefix + param_names[2]],
split_qkv=self.split_qkv)
split_qkv=self.policy.split_qkv)
for i in range(3, 4):
maybe_copy(module.attention,
sd,
Expand Down
2 changes: 1 addition & 1 deletion deepspeed/module_inject/containers/gptneo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,7 @@ def load_params(self, module, sd, weight_quantizer, mp_replace, prefix):
[prefix + param_names[0],
prefix + param_names[1],
prefix + param_names[2]],
split_qkv=self.split_qkv)
split_qkv=self.policy.split_qkv)
for i in range(3, 5):
maybe_copy(module.attention,
sd,
Expand Down
6 changes: 3 additions & 3 deletions deepspeed/module_inject/containers/gptneox.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,9 +53,9 @@ def load_params(self, module, sd, weight_quantizer, mp_replace, prefix):
transformer_param_names[i],
prefix + param_names[i],
qkv=True,
megatron_v2=self.is_megatron_v2,
split_qkv=self.split_qkv,
heads=self.client_module.attention.num_attention_heads)
megatron_v2=self.policy.is_megatron_v2,
split_qkv=self.policy.split_qkv,
heads=self.policy.client_module.attention.num_attention_heads)
for i in range(2, 4):
maybe_copy(module.attention,
sd,
Expand Down
2 changes: 1 addition & 1 deletion deepspeed/module_inject/containers/opt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,7 @@ def load_params(self, module, sd, weight_quantizer, mp_replace, prefix):
prefix + param_names[i + 1],
prefix + param_names[i + 2]
],
split_qkv=self.split_qkv)
split_qkv=self.policy.split_qkv)
for i in range(6, 8):
maybe_copy(module.attention,
sd,
Expand Down
10 changes: 3 additions & 7 deletions deepspeed/module_inject/load_checkpoint.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@ def load_model_with_checkpoint(r_module,
ckpt_mp_size,
weight_quantizer=None,
rank=0,
replace_policy=None):
container=None):
error_msgs = []

def transpose(data):
Expand Down Expand Up @@ -199,11 +199,7 @@ def load_parameters(module, prefix):
for n, child in module.named_children():
load_parameters(child, prefix + n + '.')
else:
replace_policy.load_params(module,
sd[0],
weight_quantizer,
mp_replace,
prefix)
container.load_params(module, sd[0], weight_quantizer, mp_replace, prefix)

try:
import transformers
Expand Down Expand Up @@ -274,7 +270,7 @@ def load_module_recursive(module, prefix='', level=0):
else:
load_module_recursive(
child,
prefix if (level == 0 and ckpt_type == 'pp') and replace_policy.use_load_prefix else \
prefix if (level == 0 and ckpt_type == 'pp') and container.policy.use_load_prefix else \
prefix + name + '.',
level + 1)

Expand Down
6 changes: 3 additions & 3 deletions deepspeed/module_inject/replace_module.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -566,7 +566,7 @@ def replace_fn(child, _policy, layer_id=0):
ckpt_type,
ckpt_mp_size,
quantizer,
replace_policy=container_g.policy)
container=container_g)
pbar.update(1)
else:
import gc
Expand Down Expand Up @@ -597,7 +597,7 @@ def replace_fn(child, _policy, layer_id=0):
ckpt_mp_size,
quantizer,
int(rank % tp_split_size),
replace_policy=container_g.policy)
container=container_g)
sds = [None for _ in sds]
gc.collect()

Expand All @@ -619,7 +619,7 @@ def replace_fn(child, _policy, layer_id=0):
ckpt_mp_size,
quantizer,
int(rank % tp_split_size),
replace_policy=container_g.policy)
container=container_g)
sds = [None for _ in sds]
gc.collect()
print(f"checkpoint loading time at rank {rank}: {time.time()-start_time} sec")
Expand Down