daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Debug print in vision tower forward

Browse source
This commit is contained in:
Daniel Maddern 2026-08-19 22:18:33 +07:00
parent e4d9a6f8ff
commit ffb3c9e729
1 changed files with 7 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
src/h3_blackwell_runtime/qwen3vl_vision.py
View file

@ -696,8 +696,15 @@ class Qwen3VL32BVision(nn.Module):
deepstack_features = []
for layer_num, block in enumerate(self.blocks):
x = block(x, cu_seqlens=cu_seqlens, position_embeddings=position_embeddings)
if layer_num == 0:
print(f"[vision] after block 0: x.shape={tuple(x.shape)}, grid_thw={grid_thw.tolist()}")
if layer_num in self.deepstack_visual_indexes:
# x should be [N_patches, C]; if not, re-flatten.
if x.ndim != 2:
x = x.reshape(-1, x.shape[-1])
deepstack_features.append(self.deepstack_merger_list[self.deepstack_visual_indexes.index(layer_num)](self._merge_tokens(x, grid_thw)))
if x.ndim != 2:
x = x.reshape(-1, x.shape[-1])
return self.merger(self._merge_tokens(x, grid_thw)), deepstack_features