[Bugfix] Removed the NPU-specific code path in _run_local_attention

[gcanlin]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Following #273, this PR fixes #596 and #598.

• Removed the NPU-specific code path that tried to call self.attention() directly (which wasn't callable). Now it simply calls self.attention.forward(query, key, value, attn_metadata) for all platforms, delegating the platform-specific logic to the attention backend implementation. The AscendAttentionBackendImpl.forward() method will handle the NPU dispatch automatically.
• The second issue was that when output_joint is created by slicing attn_output, it becomes a non-contiguous view of the original tensor. PyTorch's dist.all_gather requires contiguous tensors for the collective operation. The NPU backend is stricter about this requirement than the GPU backend, which is maybe why it worked on GPU but failed on NPU.

Test Plan

python text_to_image.py \
  --model Qwen/Qwen-Image \
  --prompt "a cup of coffee on the table" \
  --seed 42 \
  --cfg_scale 4.0 \
  --num_images_per_prompt 1 \
  --num_inference_steps 50 \
  --cache_backend  cache_dit \
  --ulysses_degree 2

Test Result

[Stage-0] INFO 01-03 17:10:12 [diffusion_engine.py:86] Generation completed successfully.
[Stage-0] INFO 01-03 17:10:12 [diffusion_engine.py:109] Post-processing completed in 0.1150 seconds
INFO 01-03 17:10:12 [log_utils.py:549] {'type': 'request_level_metrics',
INFO 01-03 17:10:12 [log_utils.py:549]  'request_id': '0_eb52da8d-61b8-40e8-b7a1-6682171e6411',
INFO 01-03 17:10:12 [log_utils.py:549]  'e2e_time_ms': 26954.47301864624,
INFO 01-03 17:10:12 [log_utils.py:549]  'e2e_tpt': 0.0,
INFO 01-03 17:10:12 [log_utils.py:549]  'e2e_total_tokens': 0,
INFO 01-03 17:10:12 [log_utils.py:549]  'transfers_total_time_ms': 0.0,
INFO 01-03 17:10:12 [log_utils.py:549]  'transfers_total_bytes': 0,
INFO 01-03 17:10:12 [log_utils.py:549]  'stages': {0: {'stage_gen_time_ms': 26936.877489089966,
INFO 01-03 17:10:12 [log_utils.py:549]                 'num_tokens_out': 0,
INFO 01-03 17:10:12 [log_utils.py:549]                 'num_tokens_in': 0}}}
Processed prompts: 100%|███████████████████████████████████| 1/1 [00:26<00:00, 26.95s/img, est. speed stage-0 img/s: 0.00, avg e2e_lat: 0.0ms]
INFO 01-03 17:10:12 [omni.py:687] [Summary] {'e2e_requests': 1,1 [00:26<00:00, 26.95s/img, est. speed stage-0 img/s: 0.00, avg e2e_lat: 0.0ms]
INFO 01-03 17:10:12 [omni.py:687]  'e2e_total_time_ms': 26955.371618270874,
INFO 01-03 17:10:12 [omni.py:687]  'e2e_sum_time_ms': 26954.47301864624,
INFO 01-03 17:10:12 [omni.py:687]  'e2e_total_tokens': 0,
INFO 01-03 17:10:12 [omni.py:687]  'e2e_avg_time_per_request_ms': 26954.47301864624,
INFO 01-03 17:10:12 [omni.py:687]  'e2e_avg_tokens_per_s': 0.0,
INFO 01-03 17:10:12 [omni.py:687]  'wall_time_ms': 26955.371618270874,
INFO 01-03 17:10:12 [omni.py:687]  'final_stage_id': {'0_eb52da8d-61b8-40e8-b7a1-6682171e6411': 0},
INFO 01-03 17:10:12 [omni.py:687]  'stages': [{'stage_id': 0,
INFO 01-03 17:10:12 [omni.py:687]              'requests': 1,
INFO 01-03 17:10:12 [omni.py:687]              'tokens': 0,
INFO 01-03 17:10:12 [omni.py:687]              'total_time_ms': 26954.70142364502,
INFO 01-03 17:10:12 [omni.py:687]              'avg_time_per_request_ms': 26954.70142364502,
INFO 01-03 17:10:12 [omni.py:687]              'avg_tokens_per_s': 0.0}],
INFO 01-03 17:10:12 [omni.py:687]  'transfers': []}
Adding requests:   0%|                                                                                                  | 0/1 [00:26<?, ?it/s]
[Stage-0] ERROR 01-03 17:10:12 [omni_stage.py:636] Received shutdown signal
[Stage-0] INFO 01-03 17:10:12 [gpu_worker.py:265] Worker 0: Received shutdown message
[Stage-0] INFO 01-03 17:10:12 [gpu_worker.py:287] event loop terminated.
[Stage-0] INFO 01-03 17:10:12 [gpu_worker.py:265] Worker 1: Received shutdown message
[Stage-0] INFO 01-03 17:10:12 [gpu_worker.py:287] event loop terminated.
[Stage-0] INFO 01-03 17:10:12 [npu_worker.py:126] Worker 1: Shutdown complete.
[Stage-0] INFO 01-03 17:10:12 [npu_worker.py:126] Worker 0: Shutdown complete.
INFO 01-03 17:10:17 [text_to_image.py:169] Outputs: [OmniRequestOutput(request_id='', finished=True, stage_id=0, final_output_type='image', request_output=[OmniRequestOutput(request_id='0_eb52da8d-61b8-40e8-b7a1-6682171e6411', finished=True, stage_id=None, final_output_type='image', request_output=None, images=[1 PIL Images], prompt='a cup of coffee on the table', latents=None, metrics={})], images=[], prompt=None, latents=None, metrics={})]
Saved generated image to qwen_image_output.png

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[gcanlin]

cc @mxuax @Gaohan123

[hsliuustc0106]

lgtm