Skip to content

[upgrade] upgrade gpt2 #6291

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
merged 5 commits into from
May 8, 2025
Merged

[upgrade] upgrade gpt2 #6291

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
fdcc369
fix
flybird11111 Apr 29, 2025
f8caea7
fix
flybird11111 May 1, 2025
840b9f3
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 1, 2025
519c2d0
fix
flybird11111 May 1, 2025
a612e6d
fix
flybird11111 May 7, 2025
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
17 changes: 11 additions & 6 deletions colossalai/shardformer/modeling/gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,15 @@ def _get_attention_mask(
sp_mode = shard_config.sequence_parallelism_mode
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]

if self.config.add_cross_attention and encoder_hidden_states is not None:
assert not sp_mode == "ring_attn", "Ring Attention only supports decoder-only."
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
if shard_config.enable_flash_attention:
encoder_attention_mask = ColoAttention.prepare_attn_kwargs(
(encoder_batch_size, 1, seq_len, encoder_sequence_length),
dtype=hidden_states.dtype,
dtype2=encoder_hidden_states.dtype,
device=encoder_hidden_states.device,
q_padding_mask=attention_mask,
kv_padding_mask=encoder_attention_mask,
)
Expand All @@ -77,7 +78,6 @@ def _get_attention_mask(
if shard_config.enable_flash_attention:
if attention_mask is not None:
attention_mask = attention_mask.view(batch_size, -1)

attention_mask = ColoAttention.prepare_attn_kwargs(
(batch_size, 1, seq_len, seq_len + past_key_values_length),
hidden_states.dtype,
Expand Down Expand Up @@ -835,9 +835,12 @@ def forward(
attention_mask = encoder_attention_mask
else:
query, key, value = self.c_attn(hidden_states).split(self.split_size, dim=2)
query = self._split_heads(query, self.num_heads, self.head_dim)
key = self._split_heads(key, self.num_heads, self.head_dim)
value = self._split_heads(value, self.num_heads, self.head_dim)

shape_q = (*query.shape[:-1], -1, self.head_dim)
shape_kv = (*key.shape[:-1], -1, self.head_dim)
query = query.view(shape_q).transpose(1, 2)
key = key.view(shape_kv).transpose(1, 2)
value = value.view(shape_kv).transpose(1, 2)

if layer_past is not None:
past_key, past_value = layer_past
Expand Down Expand Up @@ -871,7 +874,9 @@ def forward(
)
else:
attn_output = ColoAttention.attention(query, key, value, **attention_mask, dropout_p=dropout_p, scale=scale)
attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)

attn_output = attn_output.permute(0, 2, 1, 3).contiguous()
attn_output = attn_output.reshape(*attn_output.shape[:-2], -1).contiguous()
attn_output = self.c_proj(attn_output)
attn_output = self.resid_dropout(attn_output)
outputs = (attn_output, present, None)
Expand Down
8 changes: 1 addition & 7 deletions colossalai/shardformer/policies/gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,14 +38,8 @@ def preprocess(self):
def module_policy(self):
from transformers.models.gpt2.modeling_gpt2 import GPT2MLP, GPT2Attention, GPT2Block, GPT2Model

ATTN_IMPLEMENTATION = {
"eager": GPT2Attention,
}

policy = {}

attn_cls = ATTN_IMPLEMENTATION[self.origin_attn_implement]

embedding_cls = None
if self.shard_config.enable_tensor_parallelism:
embedding_cls = col_nn.VocabParallelEmbedding1D
Expand Down Expand Up @@ -280,7 +274,7 @@ def module_policy(self):
"forward": get_gpt2_flash_attention_forward(shard_config=self.shard_config),
},
policy=policy,
target_key=attn_cls,
target_key=GPT2Attention,
)

if not self.shard_config.pipeline_stage_manager and self.shard_config.enable_sequence_parallelism:
Expand Down
6 changes: 3 additions & 3 deletions tests/test_shardformer/test_model/test_shard_gpt2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,7 +180,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
"enable_sequence_parallelism": True,
"sequence_parallelism_mode": "split_gather",
"enable_flash_attention": True,
"use_lazy_init": True,
"use_lazy_init": False,
"precision": "fp16",
"initial_scale": 1,
},
Expand Down Expand Up @@ -238,7 +238,7 @@ def run_gpt2_test(test_config):
"tp_size": 2,
"pp_size": 2,
"num_microbatches": 4,
"enable_all_optimization": False,
"enable_all_optimization": True,
"use_lazy_init": False,
"precision": "fp32",
"initial_scale": 1,
Expand All @@ -247,7 +247,7 @@ def run_gpt2_test(test_config):
"tp_size": 2,
"pp_size": 2,
"num_microbatches": 4,
"enable_all_optimization": False,
"enable_all_optimization": True,
"use_lazy_init": False,
"precision": "fp16",
"zero_stage": 1,
Expand Down