Skip to content

recompute support offload tensor #10981

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

Open
wants to merge 1 commit into
base: dsv3-sft
Choose a base branch
from
Open

recompute support offload tensor #10981

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
3 changes: 2 additions & 1 deletion paddlenlp/transformers/deepseek_v2/configuration.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -179,6 +179,7 @@ def __init__(
attention_dropout=0.0,
speculate_model_type=False,
using_flex_token=False,
decoderlayer_act_offload_settings={},
**kwargs,
):
self.vocab_size = vocab_size
Expand Down Expand Up @@ -227,7 +228,7 @@ def __init__(
self.speculate_model_type = speculate_model_type
self.use_fp8 = False
self.using_flex_token = using_flex_token

self.decoderlayer_act_offload_settings = decoderlayer_act_offload_settings
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
Expand Down
28 changes: 28 additions & 0 deletions paddlenlp/transformers/deepseek_v2/modeling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1660,6 +1660,33 @@ def recompute_training_full(
use_cache: bool,
attn_mask_startend_row_indices: Optional[Tensor] = None,
):
decoderlayer_act_offload_settings = self.config.get(
"decoderlayer_act_offload_settings", {"type": "", "value": ""}
)

setting_type = decoderlayer_act_offload_settings["type"]
offload_value = decoderlayer_act_offload_settings["value"]

def get_offload_kwargs(layer_idx, setting_type, offload_value):
offload_kwargs = {}
if "mod" == setting_type:
assert isinstance(offload_value, (list, tuple))
v1, v2 = offload_value
offload_kwargs["offload_indices"] = [0] if layer_idx % v1 == v2 else []
else:
raise ValueError(
f"decoderlayer_act_offload_settings only support type == 'mod' ,but get type {setting_type}"
)
return offload_kwargs

layer_idx = layer_module.layer_idx
# NOTE: the first layer inputs will be used in mtp, so do not offload it
if layer_idx == 0:
offload_kwargs = {}
else:
offload_kwargs = get_offload_kwargs(layer_idx, setting_type, offload_value)
print("recompute offload ", offload_kwargs)

def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
Expand All @@ -1676,6 +1703,7 @@ def custom_forward(*inputs):
use_cache,
attn_mask_startend_row_indices,
use_reentrant=self.config.recompute_use_reentrant,
**offload_kwargs,
)

return hidden_states
Expand Down
39 changes: 37 additions & 2 deletions paddlenlp/transformers/deepseek_v2/modeling_pp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,6 +93,27 @@ def get_attr(layer, name):
return get_attr(layer._layer, name)


def get_offload_kwargs(layer_idx, decoderlayer_act_offload_settings):
setting_type = decoderlayer_act_offload_settings["type"]
offload_value = decoderlayer_act_offload_settings["value"]

# NOTE: the first layer inputs will be used in mtp, so do not offload it
if layer_idx == 0:
offload_kwargs = {}
else:
offload_kwargs = {}
if "mod" == setting_type:
assert isinstance(offload_value, (list, tuple))
v1, v2 = offload_value
offload_kwargs["offload_indices"] = [0] if layer_idx % v1 == v2 else []
elif setting_type is not None and setting_type != "":
raise ValueError(
f"decoderlayer_act_offload_settings only support type == 'mod' ,but get type {setting_type}"
)
print("offload_kwargs ", offload_kwargs)
return offload_kwargs


class DeepseekV2EmbeddingPipe(nn.Layer):
def __init__(self, config: DeepseekV2Config):
super(DeepseekV2EmbeddingPipe, self).__init__()
Expand Down Expand Up @@ -132,12 +153,14 @@ def forward(self, args):
attention_mask = attention_mask[
:, :, : -self.config.num_nextn_predict_layers, : -self.config.num_nextn_predict_layers
]

# attn_mask_startend_row_indices: [b, num_head, seq_len] or [b, num_head, seq_len, C], C is 2 or 4
if attn_mask_startend_row_indices is not None:
if attn_mask_startend_row_indices.ndim == 3:
attn_mask_startend_row_indices = attn_mask_startend_row_indices[
:, :, : -self.config.num_nextn_predict_layers,
:,
:,
: -self.config.num_nextn_predict_layers,
]
elif attn_mask_startend_row_indices.ndim == 4:
attn_mask_startend_row_indices = attn_mask_startend_row_indices[
Expand Down Expand Up @@ -222,6 +245,10 @@ def forward(self, args):
attn_mask_startend_row_indices, position_ids = None, attn_mask_startend_row_indices

if self.enable_recompute and self.config.recompute_granularity == "full" and has_gradient:
decoderlayer_act_offload_settings = self.config.get(
"decoderlayer_act_offload_settings", {"type": "", "value": ""}
)
offload_kwargs = get_offload_kwargs(self.layer_idx, decoderlayer_act_offload_settings)
if attention_mask is not None or attn_mask_startend_row_indices is not None:
hidden_states = recompute(
super().forward,
Expand All @@ -230,6 +257,7 @@ def forward(self, args):
attention_mask=attention_mask,
attn_mask_startend_row_indices=attn_mask_startend_row_indices,
use_reentrant=False,
**offload_kwargs,
)
else:
# for pretrain
Expand All @@ -239,6 +267,7 @@ def forward(self, args):
position_ids=position_ids,
attn_mask_startend_row_indices=attn_mask_startend_row_indices,
use_reentrant=self.config.recompute_use_reentrant,
**offload_kwargs,
)
else:
hidden_states = super().forward(
Expand Down Expand Up @@ -279,6 +308,10 @@ def forward(self, args):
for depth in range(self.config.num_nextn_predict_layers):
inputs_embeds_cur_depth = inputs_embeds_cur_depth_list[depth]
if self.enable_recompute and self.config.recompute_granularity == "full" and has_gradient:
decoderlayer_act_offload_settings = self.config.get(
"decoderlayer_act_offload_settings", {"type": "", "value": ""}
)
offload_kwargs = get_offload_kwargs(self.layer_idx, decoderlayer_act_offload_settings)
if attention_mask is not None or attn_mask_startend_row_indices is not None:
hidden_states = recompute(
super().forward,
Expand All @@ -288,6 +321,7 @@ def forward(self, args):
attention_mask=attention_mask,
attn_mask_startend_row_indices=attn_mask_startend_row_indices,
use_reentrant=False,
**offload_kwargs,
)
else:
# for pretrain
Expand All @@ -298,6 +332,7 @@ def forward(self, args):
position_ids=position_ids,
attn_mask_startend_row_indices=attn_mask_startend_row_indices,
use_reentrant=self.config.recompute_use_reentrant,
**offload_kwargs,
)
else:
hidden_states = super().forward(
Expand Down
Loading