tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

reorder the packing check

Browse Source
This commit is contained in:
Wing Lian
2025-05-03 15:38:29 -04:00
parent f301a165c3
commit bed8f354a5
1 changed files with 21 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
src/axolotl/monkeypatch/attention/xformers.py
View File

@@ -41,32 +41,10 @@ def xformers_attention_forward(
attn_bias = xformers.ops.LowerTriangularMask() attn_bias = xformers.ops.LowerTriangularMask()
if attention_mask is not None:
batch_size = query.shape[0]
query, key, value, indices_q, cu_seq_lens, _ = _upad_input(
query, key, value, attention_mask, query_length
)
cu_seqlens_q, cu_seq_lens_k = cu_seq_lens
seq_lengths = []
for i in range(len(cu_seq_lens_q) - 1):
seq_lengths.append(cu_seqlens_q[i + 1] - cu_seq_lens_q[i])
attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
q_seqlen=seq_lengths,
kv_seqlen=seq_lengths,
)
attn_output_unpad = xformers_attention(
query,
key,
value,
attn_bias=attn_bias,
)
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
# If position_ids is provided and check all examples do not contain only 1 sequence, If tensor in increasing # If position_ids is provided and check all examples do not contain only 1 sequence, If tensor in increasing
# then we probably have one sequence, otherwise it is packed. Additionally check we are in pre-fill/training stage. # then we probably have one sequence, otherwise it is packed. Additionally check we are in pre-fill/training stage.
# Use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach # Use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach
elif position_ids is not None and ( if position_ids is not None and (
max_length_q is not None max_length_q is not None
or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all()) or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all())
): ):
@@ -118,7 +96,27 @@ def xformers_attention_forward(
value, value,
attn_bias=attn_bias, attn_bias=attn_bias,
) )
elif attention_mask is not None:
batch_size = query.shape[0]
query, key, value, indices_q, cu_seq_lens, _ = _upad_input(
query, key, value, attention_mask, query_length
)
cu_seqlens_q, cu_seq_lens_k = cu_seq_lens
seq_lengths = []
for i in range(len(cu_seq_lens_q) - 1):
seq_lengths.append(cu_seqlens_q[i + 1] - cu_seq_lens_q[i])
attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
q_seqlen=seq_lengths,
kv_seqlen=seq_lengths,
)
attn_output_unpad = xformers_attention(
query,
key,
value,
attn_bias=attn_bias,
)
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
else: else:
attn_output = xformers_attention( attn_output = xformers_attention(
query, query,