[BlivionIaG]

Purpose

Subtraction of stale serve instructions. Ground truth is docs/rdna2/tuning.md and tools/rdna/serve_rdna.sh RECIPE=<name>. Nothing here changes kernels or serving.

Not a duplicate: open PRs #87, #99, #100, #101, and #50 do not delete these docs. Hunks those PRs already own were left alone (see below). No model-output change, so no model eval. AI assistance was used (Cursor Grok 4.7).

Files

Deleted (nothing in the tree linked them):

• docs/rdna2/MTP-FASTEST-PATH.md — two-shot rdna_ar and VLLM_RDNA_AR_MAX_KB=20480 are gone (gate is 64 KiB; two-shot requests are ignored).
• docs/recipes/qwen38-awq-full-hip-gfx1030.md — VLLM_ROCM_TRUE_FULL is not in the tree. Current dense recipe: RECIPE=full / RECIPE=27b-awq.
• docs/profiling/2026-09-10-truefull-handover.md — required VLLM_FORCE_CUSTOM_ALL_REDUCE=1 (host crash) and VLLM_ROCM_MIXED_LOG (gone).
• tools/bench_streaming_ttft.py — only caller was that handover.
• docs/recipes/ornith-exl3-gfx1030.md — 2026-09-05 snapshot. Current recipe: tools/rdna/recipes/27b-exl3.env.

Trimmed:

• docs/rdna2/w4a8-wiring.md — dropped the vllm-rdna-0.28.0 build/pytest checkout and /tmp/pr9_branch. Pytest is tests/kernels/quantization/test_rdna2_w4a8.py. The in-model paragraph that still says W4A8 default 1 and names m27b_matrix.sh is cleanup(tools): delete finished tools/rdna2_028 campaign drivers #87's hunk and was not edited. The env table already says default OFF, matching tools/rdna/recipes/27b-awq.env (W4A8=0).
• docs/rdna2/v0.31.0-port-notes.md — preamble no longer says nothing ran on GPU (the validation log is later in the same file) or that the PLE sidecar is unwired and always fails at init. Dropped debug knobs that this tree does not read (VLLM_CG_REPLAY_LOG, VLLM_PIECE_DUMP, VLLM_PIECE_IN_DEBUG, VLLM_GDN_OUT_DEBUG, VLLM_LOG_GDN_PTRS). The others in that list are still read and were kept.
• docs/rdna2/flash-next-ple-ipc.md — launcher claim is ALLOC_CONF=auto (expandable_segments off when VLLM_PLE_CPU_OFFLOAD=1). The VMM/IPC failure itself stays.
• docs/rdna2/dsv4-fp16-attention-plan.md — DeepseekV4RDNAAttention and sparse_mla_prefill_rdna2 are in tree and wired. Full serve confirmation is still outstanding.
• docs/rdna2/mixed-batch-stall-probe.md — dropped the cite of missing tools/rdna2_028/prefill_decode_probe.py (and the missing mixed_probe.py in the same sentence). The measured stalls stay.
• tunableop/README.md — scripts/ launchers are tools/rdna/. Line 15 (venv-7.14.0_0.28.0) is context for perf(rdna2): QSA/GDN HIP prefill paths vs Triton/FLA + TunableOp decode/MTP grid (implements #64) #99's rocm10 row and was not edited. perf(rdna2): QSA/GDN HIP prefill paths vs Triton/FLA + TunableOp decode/MTP grid (implements #64) #99's tuning.md rows were not touched.
• tunableop/rocm7.14-rocblas5.5/README.md — same scripts/ → tools/rdna/ path fix. perf(rdna2): QSA/GDN HIP prefill paths vs Triton/FLA + TunableOp decode/MTP grid (implements #64) #99 does not own this file.

Test Plan

Docs and one unreferenced bench script. No GPU.

# deleted names have no remaining references outside bench_results history
rg -n 'MTP-FASTEST-PATH|qwen38-awq-full-hip|ornith-exl3-gfx1030|truefull-handover|bench_streaming_ttft|VLLM_ROCM_TRUE_FULL|VLLM_ROCM_MIXED_LOG' \
  --glob '!bench_results/**'

# cited paths exist
test -f docs/rdna2/tuning.md tools/rdna/serve_rdna.sh \
  tools/rdna/recipes/27b-exl3.env tools/rdna/recipes/27b-awq.env \
  tools/rdna/recipes/full.env tests/kernels/quantization/test_rdna2_w4a8.py \
  vllm/models/deepseek_v4/amd/rdna/attention.py

Test Result

• Those names do not appear outside bench_results/ historical notes.
• VLLM_ROCM_TRUE_FULL and VLLM_ROCM_MIXED_LOG do not appear in vllm/, csrc/, or tools/.
• Cited recipe, launcher, test, and attention files exist. tools/rdna/recipes/27b-awq.env sets W4A8=0. tools/rdna/recipes/full.env sets ALLOC_CONF=auto. prefill_decode_probe.py is absent.

---

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.

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing