From b085ef02e83fc5772dacd6d06e970b88841fb113 Mon Sep 17 00:00:00 2001 From: Daniel Maddern Date: Wed, 19 Aug 2026 22:36:32 +0700 Subject: [PATCH] Merge tokens expands C by merge^2 (merger takes pre-expanded) --- src/h3_blackwell_runtime/qwen3vl_vision.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/h3_blackwell_runtime/qwen3vl_vision.py b/src/h3_blackwell_runtime/qwen3vl_vision.py index 0887a85..635e180 100644 --- a/src/h3_blackwell_runtime/qwen3vl_vision.py +++ b/src/h3_blackwell_runtime/qwen3vl_vision.py @@ -681,7 +681,7 @@ class Qwen3VL32BVision(nn.Module): C = int(x.shape[-1]) merge = 2 target = (t, h // merge, merge, w // merge, merge, C) - return x.view(target).permute(0, 1, 3, 2, 4, 5).reshape(-1, C // (merge * merge)) + return x.view(target).permute(0, 1, 3, 2, 4, 5).reshape(-1, C * merge * merge) def forward(self, flatten_patches: torch.Tensor, grid_thw: torch.Tensor) -> tuple[torch.Tensor, list[torch.Tensor]]: """Run the visual tower -> (merged, deepstack)."""