Skip to content

Commit

Permalink
Change repeat to expand in GQA (#628)
Browse files Browse the repository at this point in the history
  • Loading branch information
sashaDoubov committed Sep 26, 2023
1 parent 547b313 commit 61dfbd6
Showing 1 changed file with 34 additions and 9 deletions.
43 changes: 34 additions & 9 deletions llmfoundry/models/layers/attention.py
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,23 @@ def _reset_is_causal(num_query_tokens: int, num_key_tokens: int,
return original_is_causal


def repeat_kv_for_gqa(hidden: torch.Tensor, n_rep: int) -> torch.Tensor:
"""Perform repeat of kv heads along a particular dimension.
hidden.shape expected to be: (batch size, seq len, kv_n_heads, head_dim)
n_rep: amount of repetitions of kv_n_heads
Unlike torch.repeat_interleave, this function avoids allocating new memory.
"""
if n_rep == 1:
return hidden

b, s, kv_n_heads, d = hidden.shape

hidden = hidden[:, :, :, None, :].expand(b, s, kv_n_heads, n_rep, d)

return hidden.reshape(b, s, kv_n_heads * n_rep, d)


def scaled_multihead_dot_product_attention(
query: torch.Tensor,
key: torch.Tensor,
Expand Down Expand Up @@ -84,8 +101,11 @@ def scaled_multihead_dot_product_attention(

# grouped query case
if kv_n_heads > 1 and kv_n_heads < n_heads:
k = k.repeat_interleave(n_heads // kv_n_heads, dim=1)
v = v.repeat_interleave(n_heads // kv_n_heads, dim=1)
# necessary to do a transpose to swap (b h s d) -> (b s h d) for repeat_kv_for_gqa function
k = repeat_kv_for_gqa(k.transpose(1, 2),
n_heads // kv_n_heads).transpose(1, 2)
v = repeat_kv_for_gqa(v.transpose(1, 2),
n_heads // kv_n_heads).transpose(1, 2)

if softmax_scale is None:
softmax_scale = 1 / math.sqrt(d)
Expand Down Expand Up @@ -243,10 +263,16 @@ def flash_attn_fn(
elif kv_n_heads < n_heads:
# Each query belong to a group of kv heads of group size n_heads // kv_n_heads
# We repeat each kv head by the group size number to use the underlying MHA kernels
# done along the head dimension = 1
key_unpad = key_unpad.repeat_interleave(n_heads // kv_n_heads, dim=1)
value_unpad = value_unpad.repeat_interleave(n_heads // kv_n_heads,
dim=1)

# since repeat_kv_for_gqa expects input dims of (b, s, kv_n_heads, d)
# we use .view to modify {key, value}_unpad appropriately

key_unpad = repeat_kv_for_gqa(
key_unpad.view(batch_size, seqlen, kv_n_heads, -1),
n_heads // kv_n_heads).view(batch_size * seqlen, n_heads, -1)
value_unpad = repeat_kv_for_gqa(
value_unpad.view(batch_size, seqlen, kv_n_heads, -1),
n_heads // kv_n_heads).view(batch_size * seqlen, n_heads, -1)

dropout_p = dropout_p if training else 0.0

Expand Down Expand Up @@ -383,9 +409,8 @@ def triton_flash_attn_fn(
elif kv_n_heads < n_heads:
# Each query belong to a group of kv heads of group size n_heads // kv_n_heads
# We repeat each kv head by the group size number to use the underlying MHA kernels
# done along dim = 2, unlike the implementation for flash and torch attn
key = key.repeat_interleave(n_heads // kv_n_heads, dim=2)
value = value.repeat_interleave(n_heads // kv_n_heads, dim=2)
key = repeat_kv_for_gqa(key, n_heads // kv_n_heads)
value = repeat_kv_for_gqa(value, n_heads // kv_n_heads)

reset_is_causal = _reset_is_causal(query.size(1), key.size(1), is_causal)
attn_output = flash_attn_func( # type: ignore
Expand Down

0 comments on commit 61dfbd6

Please sign in to comment.