Missing updates for Llama4 on main

[Luca-Calabria]

Added Llama4 missing fixes from #881 #862 #884 on main branch

[Copilot]

Pull request overview

This PR ports Llama4-specific fixes from PRs #881, #862, and #884 to the main branch, focusing on attention scaling and chunked attention layer handling improvements.

Changes:

• Updated _get_attn_scale_for_hpu implementation to remove closure dependency and match the actual attention scale calculation
• Refactored chunked attention layer detection to be a standalone function and changed the signature of apply_model_specific_patches to accept model_runner instead of model
• Consolidated model-specific patches by removing duplicate maybe_set_chunked_attention_layers method from the class

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The comment 'add explicit warning' suggests that an exception handler should log a warning, but the current implementation silently ignores exceptions. Consider adding a proper warning message using a logger to help with debugging when chunked attention setup fails.

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
17b17c068453e6dc6af79240bb94857ae175cc51

[github-actions]

✅ CI Passed

All checks passed successfully against the following vllm commit:
17b17c068453e6dc6af79240bb94857ae175cc51