Fix RuntimeError by ensuring attn_mask is None when is_causal=True

[IgorAherne]

Prevents an error by modifying the Attention module: avoid passing an explicit attn_mask to F.scaled_dot_product_attention when is_causal=True.

This allows the function's internal causal masking mechanism to be used without conflict, during operations like decoder prefill.

Otherwise getting the following error:

Error processing task 1 ('C:_myDrive\repos\auto-vlog\AutoVlogProj\temp_video_processing\audio_maker_temp\girl and a stick in a forest\chunk_2.wav'): _scaled_dot_product_attention: Explicit attn_mask should not be set when is_causal=True
Traceback (most recent call last):
File "C:_myDrive\repos\auto-vlog\dia\cli.py", line 168, in process_task
output_audio = model.generate(
^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\venv\Lib\site-packages\torch\utils_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\dia\dia\model.py", line 407, in generate
dec_state, dec_output = self._prepare_generation(full_text, audio_prompt, verbose)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\dia\dia\model.py", line 289, in _prepare_generation
self.model.decoder.forward(tokens_BxTxC, dec_state)
File "C:_myDrive\repos\auto-vlog\dia\dia\layers.py", line 605, in forward
x = layer(x, state, self_attn_cache=self_cache, cross_attn_cache=cross_cache, prefill=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\venv\Lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\venv\Lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\dia\dia\layers.py", line 437, in forward
sa_out = self.self_attention(
^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\venv\Lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\venv\Lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:_myDrive\repos\auto-vlog\dia\dia\layers.py", line 251, in forward
attn_output = F.scaled_dot_product_attention(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: _scaled_dot_product_attention: Explicit attn_mask should not be set when is_causal=True
Skipping task due to error.

[IgorAherne]

Hm, looking at my proposed changes, there is more than necessary somehow.

I wanted to only offer this change:

        # If is_causal is True, sdpa handles masking internally, so attn_mask should be None.
        # Otherwise, use the provided attn_mask (e.g., for padding in encoder).
        effective_attn_mask = attn_mask if not is_causal else None

        attn_output = F.scaled_dot_product_attention(
            Xq_BxNxTxH,
            attn_k,
            attn_v,
            attn_mask=attn_mask,
            attn_mask=effective_attn_mask,     #use it here
            scale=1.0,
            enable_gqa=self.num_gqa_groups > 1,
            is_causal=is_causal,

[IgorAherne]

ok, much less intrusive now, see Files Changed tab