Skip to content
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

Flash Attention Refactor #2

Merged
merged 1 commit into from
Jul 15, 2024
Merged

Flash Attention Refactor #2

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
127 changes: 11 additions & 116 deletions src/transformers/models/mamba2/modeling_mamba2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -51,8 +51,7 @@
logger = logging.get_logger(__name__)

if is_flash_attn_2_available():
from flash_attn import flash_attn_func, flash_attn_varlen_func
from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
from ...modeling_flash_attention_utils import _flash_attention_forward

if is_mamba_ssm_greater_or_equal_2_0_4():
from mamba_ssm.ops.triton.selective_state_update import selective_state_update
Expand Down Expand Up @@ -660,19 +659,6 @@ def _attn(self, query, key, value, attention_mask=None):
return attn_output, attn_weights


# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
max_seqlen_in_batch = seqlens_in_batch.max().item()
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
return (
indices,
cu_seqlens,
max_seqlen_in_batch,
)


# Adapted from transformers.models.gpt_neox.modeling_gpt_neox.GPTNeoXFlashAttention2 with GPTNeoX->Mamba2
class Mamba2FlashAttention2(Mamba2Attention):
"""
Expand Down Expand Up @@ -742,8 +728,16 @@ def forward(
value = value.to(target_dtype)

# Compute attention
attn_weights = self._flash_attention_forward(
query, key, value, attention_mask, query_length, dropout=0.0, softmax_scale=None
attn_weights = _flash_attention_forward(
query,
key,
value,
attention_mask,
query_length,
dropout=0.0,
softmax_scale=None,
is_causal=self.is_causal,
use_top_left_mask=self._flash_attn_uses_top_left_mask,
)

# Reshape outputs
Expand All @@ -755,105 +749,6 @@ def forward(

return attn_output, attn_weights

# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward
def _flash_attention_forward(
self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
):
"""
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
first unpad the input, then computes the attention scores and pad the final attention scores.

Args:
query_states (`torch.Tensor`):
Input query states to be passed to Flash Attention API
key_states (`torch.Tensor`):
Input key states to be passed to Flash Attention API
value_states (`torch.Tensor`):
Input value states to be passed to Flash Attention API
attention_mask (`torch.Tensor`):
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
position of padding tokens and 1 for the position of non-padding tokens.
dropout (`float`):
Attention dropout
softmax_scale (`float`, *optional*):
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
"""
if not self._flash_attn_uses_top_left_mask:
causal = self.is_causal
else:
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
causal = self.is_causal and query_length != 1

# Contains at least one padding token in the sequence
if attention_mask is not None:
batch_size = query_states.shape[0]
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
query_states, key_states, value_states, attention_mask, query_length
)

cu_seqlens_q, cu_seqlens_k = cu_seq_lens
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens

attn_output_unpad = flash_attn_varlen_func(
query_states,
key_states,
value_states,
cu_seqlens_q=cu_seqlens_q,
cu_seqlens_k=cu_seqlens_k,
max_seqlen_q=max_seqlen_in_batch_q,
max_seqlen_k=max_seqlen_in_batch_k,
dropout_p=dropout,
softmax_scale=softmax_scale,
causal=causal,
)

attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
else:
attn_output = flash_attn_func(
query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
)

return attn_output

# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._upad_input with num_heads->num_attention_heads
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape

key_layer = index_first_axis(
key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
)
value_layer = index_first_axis(
value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
)
if query_length == kv_seq_len:
query_layer = index_first_axis(
query_layer.reshape(batch_size * kv_seq_len, self.num_attention_heads, head_dim), indices_k
)
cu_seqlens_q = cu_seqlens_k
max_seqlen_in_batch_q = max_seqlen_in_batch_k
indices_q = indices_k
elif query_length == 1:
max_seqlen_in_batch_q = 1
cu_seqlens_q = torch.arange(
batch_size + 1, dtype=torch.int32, device=query_layer.device
) # There is a memcpy here, that is very bad.
indices_q = cu_seqlens_q[:-1]
query_layer = query_layer.squeeze(1)
else:
# The -q_len: slice assumes left padding.
attention_mask = attention_mask[:, -query_length:]
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)

return (
query_layer,
key_layer,
value_layer,
indices_q,
(cu_seqlens_q, cu_seqlens_k),
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
)


# Adapted from transformers.models.gpt_neox.modeling_gpt_neox.GPTNeoXSdpaAttention with GPTNeoX->Mamba2
class Mamba2SdpaAttention(Mamba2Attention):
Expand Down