ROCm : Sage Attn Out Of Resources

Nobody has claimed this yet.

Hi,

On Ubuntu, RX 7900 XTX (GFX1100). Sage Attn is installed by default.

```
sageattention 1.0.6
torch 2.7.0+rocm6.2.4
torchaudio 2.7.0+rocm6.2.4
triton 3.3.0
```

$dem_graio.py

```
Currently enabled native sdp backends: ['flash', 'math', 'mem_efficient', 'cudnn']
Xformers is not installed!
Flash Attn is not installed!
Sage Attn is installed!
```

These are the error messages:

```
Unloaded SiglipVisionModel as complete.
Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB
0%| | 0/25 [00:00

File "/home/user/sd/FramePack/diffusers_helper/models/hunyuan_video_packed.py", line 111, in attn_varlen_func
x = sageattn(q, k, v, tensor_layout='NHD')
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/lib/python3.12/site-packages/sageattention/core.py", line 110, in sageattn
o = attn_false(q_int8, k_int8, v, q_scale, k_scale, tensor_layout=tensor_layout, output_dtype=dtype)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/lib/python3.12/site-packages/sageattention/attn_qk_int8_per_block.py", line 113, in forward
_attn_fwd[grid](
File "/home/user/miniconda3/lib/python3.12/site-packages/triton/runtime/jit.py", line 347, in
return lambda *args, **kwargs: self.run(grid=grid, warmup=False, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/user/miniconda3/lib/python3.12/site-packages/triton/runtime/jit.py", line 591, in run
kernel.run(grid_0, grid_1, grid_2, stream, kernel.function, kernel.packed_metadata,
^^^^^^^^^^
File "/home/user/miniconda3/lib/python3.12/site-packages/triton/compiler/compiler.py", line 413, in __getattribute__
self._init_handles()
File "/home/user/miniconda3/lib/python3.12/site-packages/triton/compiler/compiler.py", line 401, in _init_handles
raise OutOfResources(self.metadata.shared, max_shared, "shared memory")
triton.runtime.errors.OutOfResources: out of resource: shared memory, Required: 90112, Hardware limit: 65536. Reducing block sizes or `num_stages` may help.
```

There is no error when using `flash_attn 2.7.4.post1`

Start with diffusers_helper/models/hunyuan_video_packed.py at attn_varlen_func and reproduce the reported SageAttention path on the RX 7900 XTX with the listed PyTorch and ROCm versions. Trace the call into the installed SageAttention and Triton packages; done means the attention path runs without the shared-memory OutOfResources failure.

Written by the indexing model from the issue text.