Skip to content

Fix VLM position_ids packing in Megatron strategy with sequence packing enabled #452

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
sanmuf wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix VLM position_ids packing in Megatron strategy with sequence packing enabled #452

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
25 changes: 20 additions & 5 deletions roll/distributed/strategy/megatron_strategy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -341,10 +341,13 @@ def _pack_sequences(self, input_tensor, attention_mask, pad_packed_seq_to=None,
seq_tokens = input_tensor_unpadded[b]

# Pad sequence if needed
if padded_seq_len > seq_len:
seq_tokens = torch.nn.functional.pad(
seq_tokens, (0, padded_seq_len - seq_len), value=pad_val
)
pad_seq_len = padded_seq_len - seq_len
if pad_seq_len > 0:
if seq_tokens.dim() == 1:
pad_config = (0, pad_seq_len)
else:
pad_config = tuple([0, 0] * (seq_tokens.dim() - 1) + [0, pad_seq_len])
seq_tokens = torch.nn.functional.pad(seq_tokens, pad_config, value=pad_val)
all_input_tensor_padded.append(seq_tokens)

if cp_size > 1:
Expand Down Expand Up @@ -411,20 +414,32 @@ def inner_forward_step(self, loss_func, data_iterator: Iterator[DataProto], mode
attention_mask = data.batch["attention_mask"]
labels = data.batch["labels"] if "labels" in data.batch else None # labels is only used for sft
packed_seq_params = None
position_ids = None
raw_position_ids = None
if "position_ids" in data.batch.keys() and data.batch["position_ids"].dim() == 3: # qwen2vl/qwen3vl mrope
raw_position_ids = data.batch["position_ids"]
if raw_position_ids.size(1) == 4:
raw_position_ids = raw_position_ids[:, 1:, :].contiguous() # (bsz, 4, seqlen) -> (bsz, 3, seqlen)

if self.use_sequence_packing:
input_ids, packed_seq_params, cu_seqlens, cu_seqlens_padded = self._pack_sequences(
input_ids, attention_mask,
)
if labels is not None:
labels, _, _, _ = self._pack_sequences(labels, attention_mask, pad_val=IGNORE_INDEX)
if raw_position_ids is not None:
position_ids_for_pack = raw_position_ids.transpose(1, 2).contiguous() # (bsz, C, seqlen) -> (bsz, seqlen, C)
packed_position_ids, _, _, _ = self._pack_sequences(position_ids_for_pack, attention_mask, pad_val=0)
position_ids = packed_position_ids.squeeze(0).transpose(0, 1).unsqueeze(1).contiguous()
attention_mask = None
else:
input_ids = self._get_feature_on_this_cp_rank(input_ids, "input_ids")
attention_mask = self._get_feature_on_this_cp_rank(attention_mask, "attention_mask")
if labels is not None:
labels = self._get_feature_on_this_cp_rank(labels, "labels")
position_ids = None
if raw_position_ids is not None:
attention_mask = None
position_ids = raw_position_ids.transpose(0, 1) # (bsz, C, seqlen) -> (C, bsz, seqlen)
# attention_mask: SelfAttention defalt to te DotProductAttention with
# AttnMaskType.causal in which attention_mask would not be used, pass
# it mainly for moe aux loss without pad token and it is 2D
Expand Down