Merge upstream vLLM code into gfx11

[amd-callumm]

Purpose

Merge all commits from vllm-project/vllm:main since the last common ancestor with ROCm/vllm:gfx11.

The commit count and change volume are huge, and included a fair number of conflict resolutions. Thus, substantial testing is needed to ensure critical functionality and optimizations are not lost.

Customer safeguard: in the event that some functionality breaks or performance regresses as a result of this merge, the gfx11_20260528 tag can be used to obtain stable pre-merge code. This package is based on the May 28 nightly build which showed good numbers and test results during nightly regressions.

Test Plan

• Run representative subsets of all of the following test suites on a Strix Halo machine pre- and post-merge:

  • tests/basic_correctness/test_basic_correctness.py
  • tests/kernels/core/ (fundamental kernel operations)
  • tests/kernels/moe/test_exllama_moe.py
  • tests/kernels/moe/test_hybrid_w4a16_moe.py
  • tests/kernels/quantization/test_awq_gemv_moe.py
  • tests/kernels/quantization/test_hip_w4a16.py
  • tests/kernels/quantization/test_hybrid_w4a16_triton.py
  • tests/kernels/quantization/test_rocm_compressed_tensors_w4a16.py
  • tests/kernels/quantization/test_rocm_skinny_gemms.py
  • tests/kernels/quantization/test_dynamic_int8_lm_head.py
  • tests/kernels/test_wvsplitk_fused_silu.py
  • tests/kernels/attention/test_rocm_attention_selector.py
  • tests/kernels/attention/test_triton_unified_attention.py
  • tests/kernels/attention/test_cache.py (KV cache operations + reshaping)
  • tests/kernels/attention/test_flash_attn.py
  • tests/kernels/attention/test_paged_attn.py
  • tests/kernels/moe/test_batched_moe.py
  • tests/kernels/moe/test_fused_topk.py
  • tests/quantization/test_hip_w4a16_kernel.py
  • tests/quantization/test_fp8.py
  • tests/quantization/test_compressed_tensors.py
  • tests/samplers/test_beam_search.py
  • tests/samplers/test_ignore_eos.py
  • tests/models/language/generation/test_common.py
  • tests/entrypoints/llm/test_generate.py
  • tests/rocm/aiter/test_grouped_quant.py (grouped FP8 quantization)
  • tests/rocm/aiter/test_mla_fp8_support_check.py (MLA FP8 support)
  • tests/rocm/aiter/test_fused_qk_norm_mrope_kvcache.py

• Run attention benchmarking tests to check for any regressions

• Run nightly regressions suite to check for severe performance regressions versus May 28 2026 nightly results

• For any regressions found, plan next steps

Test Result

For all correctness/functionality tests, exactly the same test cases are passing both pre- and post-merge. Out of ~1700 tests, 8 tests failed, all for pre-existing and low-risk reasons such as model gating.

Performance checks are pending.