fix(qwen3_moe): pass kwargs to self_attn

This is needed to avoid `.item()` calls in `_flash_attention_forward`.

Author: llllvvuu

src/transformers/models/qwen3_moe/modeling_qwen3_moe.py

@@ -355,6 +355,7 @@ def forward(
             use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
+            **kwargs,
         )
         hidden_states = residual + hidden_states
 

src/transformers/models/qwen3_moe/modular_qwen3_moe.py

@@ -189,6 +189,7 @@ def forward(
             use_cache=use_cache,
             cache_position=cache_position,
             position_embeddings=position_embeddings,
+            **kwargs,
         )
         hidden_states = residual + hidden_states