[Bugfix] Fix missing seq_start_loc in xformers prefill metadata

[Isotr0py]

Phi-3-small with blocksparse attention is not working with xformers attention prefill metadata because of missing seq_start_loc:

Logs

[rank0]: Traceback (most recent call last):
[rank0]:   File "/kaggle/working/vllm/vllm/worker/model_runner_base.py", line 116, in _wrapper
[rank0]:     return func(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/worker/model_runner.py", line 1718, in execute_model
[rank0]:     hidden_or_intermediate_states = model_executable(
[rank0]:                                     ^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/model_executor/models/phi3_small.py", line 445, in forward
[rank0]:     output_hidden_states = self.model(
[rank0]:                            ^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/model_executor/models/phi3_small.py", line 358, in forward
[rank0]:     hidden_states = layer(
[rank0]:                     ^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/model_executor/models/phi3_small.py", line 290, in forward
[rank0]:     hidden_states = self.self_attn(
[rank0]:                     ^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/model_executor/models/phi3_small.py", line 249, in forward
[rank0]:     attn_output = self.attn(q, k, v, kv_cache, attn_metadata=attn_metadata)
[rank0]:                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/attention/layer.py", line 177, in forward
[rank0]:     return torch.ops.vllm.unified_attention(
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/_ops.py", line 1116, in __call__
[rank0]:     return self._op(*args, **(kwargs or {}))
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/attention/layer.py", line 271, in unified_attention
[rank0]:     return self.impl.forward(self, query, key, value, kv_cache, attn_metadata)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/attention/backends/blocksparse_attn.py", line 424, in forward
[rank0]:     output = self.bs_attn(
[rank0]:              ^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1747, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/attention/ops/blocksparse_attention/interface.py", line 223, in forward
[rank0]:     return self.spda(
[rank0]:            ^^^^^^^^^^
[rank0]:   File "/kaggle/working/vllm/vllm/attention/ops/blocksparse_attention/interface.py", line 185, in spda
[rank0]:     cu_seqlens = cu_seqlens_k.cpu()
[rank0]:                  ^^^^^^^^^^^^^^^^
[rank0]: AttributeError: 'NoneType' object has no attribute 'cpu'

• This PR fixes the missing seq_start_loc in xformers prefill metadata.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

Thanks for fixing! cc @WoosukKwon