From ffb3c9e729bd6cdad83be998336c3fb73ab08fa8 Mon Sep 17 00:00:00 2001 From: Daniel Maddern Date: Wed, 19 Aug 2026 22:18:33 +0700 Subject: [PATCH] Debug print in vision tower forward --- src/h3_blackwell_runtime/qwen3vl_vision.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/src/h3_blackwell_runtime/qwen3vl_vision.py b/src/h3_blackwell_runtime/qwen3vl_vision.py index a011059..ef2fc80 100644 --- a/src/h3_blackwell_runtime/qwen3vl_vision.py +++ b/src/h3_blackwell_runtime/qwen3vl_vision.py @@ -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