mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

Remove sdpa context mgrs

Browse Source
This commit is contained in:
Ross Wightman 2023-09-25 23:30:56 -07:00
parent 2734bb76ce
commit 379780bb6c
1 changed files with 16 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
timm/models/vision_transformer_packed.py
View File

@ -124,7 +124,7 @@ def pack_images(
):
max_seq_len = max_grid_size[0] * max_grid_size[1]
# patchify if needed, generate position indices, apply patch drop, record seq lengths
# patchify, generate position indices, apply patch drop, record seq lengths
img_tokens = []
img_pos_indices = []
img_seq_lens = []
@ -144,6 +144,7 @@ def pack_images(
indexing='ij'),
dim=-1,
)
# FIXME patch drop here
img_tokens.append(patches.flatten(0, 1))
img_pos_indices.append(pos_indices.flatten(0, 1))
img_seq_lens.append(seq_len)
@ -221,7 +222,6 @@ class Attention(nn.Module):
attn_mask = attn_mask.expand((-1, self.num_heads, -1, -1))
if self.fused_attn:
with torch.backends.cuda.sdp_kernel(enable_mem_efficient=False):
x = F.scaled_dot_product_attention(
q, k, v,
attn_mask=attn_mask,
@ -374,7 +374,6 @@ class ParallelScalingBlock(nn.Module):
k = self.k_norm(k.view(B, N, self.num_heads, self.head_dim)).transpose(1, 2)
v = v.view(B, N, self.num_heads, self.head_dim).transpose(1, 2)
if self.fused_attn:
with torch.backends.cuda.sdp_kernel(enable_mem_efficient=False):
x_attn = F.scaled_dot_product_attention(
q, k, v,
attn_mask=attn_mask,
@ -507,7 +506,6 @@ class AttentionPoolLatent(nn.Module):
q = self.q_norm(q)
k = self.k_norm(k)
if False:
with torch.backends.cuda.sdp_kernel(enable_mem_efficient=False):
x = F.scaled_dot_product_attention(
q, k, v,
attn_mask=attn_mask,