fix: use HF-style attn mask path for qwen3-vl RoPE in BSHD training (#2582)

Signed-off-by: joosungy <joosungy@nvidia.com>
Co-authored-by: Li Jinliang <jinliangl@nvidia.com>

Author: eagle705

src/megatron/bridge/models/qwen_vl/modelling_qwen3_vl/model.py

@@ -487,6 +487,9 @@ def forward(
 
         if position_ids is None:
             # BSHD
+            # Megatron uses 4D bool masks ([B|1,1,S,S], True=masked); HF uses 2D keep masks ([B,S], 1=keep)
+            # For simplicity, we set hf_attention_mask to None.
+            hf_attention_mask = None
             position_ids, _ = get_rope_index(
                 self.config.spatial_merge_size,
                 self.image_token_id,
@@ -495,7 +498,7 @@ def forward(
                 input_ids,
                 image_grid_thw=image_grid_thw,
                 video_grid_thw=video_grid_thw,
-                attention_mask=attention_mask,
+                attention_mask=hf_attention_mask,
             )  #  [3*b*s]
             if packed_seq_params is not None:
                 # convert position_ids to THD format