quoccuongle/Grounded-SAM-2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

better support for non-CUDA devices (CPU, MPS) (#192)

Browse Source
This commit is contained in:
Ronghang Hu
2024-08-12 10:46:50 -07:00
committed by GitHub
parent 778e112740
commit 1034ee2a1a
8 changed files with 213 additions and 377 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
sam2/modeling/position_encoding.py
View File

@@ -211,6 +211,11 @@ def apply_rotary_enc(
# repeat freqs along seq_len dim to match k seq_len
if repeat_freqs_k:
r = xk_.shape[-2] // xq_.shape[-2]
freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
if freqs_cis.is_cuda:
freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
else:
# torch.repeat on complex numbers may not be supported on non-CUDA devices
# (freqs_cis has 4 dims and we repeat on dim 2) so we use expand + flatten
freqs_cis = freqs_cis.unsqueeze(2).expand(-1, -1, r, -1, -1).flatten(2, 3)
xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)

4
sam2/modeling/sam2_base.py
View File

@@ -567,10 +567,10 @@ class SAM2Base(torch.nn.Module):
continue # skip padding frames
# "maskmem_features" might have been offloaded to CPU in demo use cases,
# so we load it back to GPU (it's a no-op if it's already on GPU).
feats = prev["maskmem_features"].cuda(non_blocking=True)
feats = prev["maskmem_features"].to(device, non_blocking=True)
to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
# Spatial positional encoding (it might have been offloaded to CPU in eval)
maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
maskmem_enc = prev["maskmem_pos_enc"][-1].to(device)
maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
# Temporal positional encoding
maskmem_enc = (