Skip to content

fix(rt-detr): use float32 on MPS in build_2d_sinusoidal_position_embedding #46513

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

Closed
Shylin26 wants to merge 1 commit into from
Closed

fix(rt-detr): use float32 on MPS in build_2d_sinusoidal_position_embedding #46513

Show file tree
Hide file tree
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
11 changes: 6 additions & 5 deletions src/transformers/models/rt_detr/modeling_rt_detr.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -847,7 +847,7 @@ def build_2d_sinusoidal_position_embedding(
temperature: Base for the frequency decay.
cls_token: If `True`, prepend a zero row for a CLS token.
device: Target device; defaults to CPU.
dtype: Output dtype; frequency arithmetic uses float64 internally.
dtype: Output dtype; frequency arithmetic uses float64 internally (float32 on MPS).

Returns:
Tensor of shape ``(height * width [+1], embed_dim)``.
Expand All @@ -856,11 +856,12 @@ def build_2d_sinusoidal_position_embedding(
raise ValueError(f"`embed_dim` must be divisible by 4, got {embed_dim}")

pos_dim = embed_dim // 4
omega = torch.arange(pos_dim, dtype=torch.float64, device=device) / pos_dim
compute_dtype = torch.float64 if str(device) != "mps" else torch.float32
omega = torch.arange(pos_dim, dtype=compute_dtype, device=device) / pos_dim
omega = 1.0 / temperature**omega # (D/4,)

grid_h = torch.arange(height, dtype=torch.float64, device=device)
grid_w = torch.arange(width, dtype=torch.float64, device=device)
grid_h = torch.arange(height, dtype=compute_dtype, device=device)
grid_w = torch.arange(width, dtype=compute_dtype, device=device)
grid_h, grid_w = torch.meshgrid(grid_h, grid_w, indexing="ij") # (H, W) each

emb_h = grid_h.flatten().outer(omega) # (H*W, D/4)
Expand All @@ -869,7 +870,7 @@ def build_2d_sinusoidal_position_embedding(
pos_embed = torch.cat([emb_h.sin(), emb_h.cos(), emb_w.sin(), emb_w.cos()], dim=1)

if cls_token:
pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=torch.float64, device=device), pos_embed], dim=0)
pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=compute_dtype, device=device), pos_embed], dim=0)

return pos_embed.to(dtype)

Expand Down
11 changes: 6 additions & 5 deletions src/transformers/models/rt_detr_v2/modeling_rt_detr_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -976,7 +976,7 @@ def build_2d_sinusoidal_position_embedding(
temperature: Base for the frequency decay.
cls_token: If `True`, prepend a zero row for a CLS token.
device: Target device; defaults to CPU.
dtype: Output dtype; frequency arithmetic uses float64 internally.
dtype: Output dtype; frequency arithmetic uses float64 internally (float32 on MPS).

Returns:
Tensor of shape ``(height * width [+1], embed_dim)``.
Expand All @@ -985,11 +985,12 @@ def build_2d_sinusoidal_position_embedding(
raise ValueError(f"`embed_dim` must be divisible by 4, got {embed_dim}")

pos_dim = embed_dim // 4
omega = torch.arange(pos_dim, dtype=torch.float64, device=device) / pos_dim
compute_dtype = torch.float64 if str(device) != "mps" else torch.float32
omega = torch.arange(pos_dim, dtype=compute_dtype, device=device) / pos_dim
omega = 1.0 / temperature**omega # (D/4,)

grid_h = torch.arange(height, dtype=torch.float64, device=device)
grid_w = torch.arange(width, dtype=torch.float64, device=device)
grid_h = torch.arange(height, dtype=compute_dtype, device=device)
grid_w = torch.arange(width, dtype=compute_dtype, device=device)
grid_h, grid_w = torch.meshgrid(grid_h, grid_w, indexing="ij") # (H, W) each

emb_h = grid_h.flatten().outer(omega) # (H*W, D/4)
Expand All @@ -998,7 +999,7 @@ def build_2d_sinusoidal_position_embedding(
pos_embed = torch.cat([emb_h.sin(), emb_h.cos(), emb_w.sin(), emb_w.cos()], dim=1)

if cls_token:
pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=torch.float64, device=device), pos_embed], dim=0)
pos_embed = torch.cat([torch.zeros(1, embed_dim, dtype=compute_dtype, device=device), pos_embed], dim=0)

return pos_embed.to(dtype)

Expand Down