[BlivionIaG]

Purpose

Subtraction-only cleanup of the W4A16 prefill host path on rdna_extras. This is a redo of the substance of #48. #48 stays on rdna_extra/v0.28.0 and is not retargeted. Default config and split-K selection are the same when the removed env vars are unset.

AI assistance was used (Cursor Grok 4.7). A human should review every changed line and run the gfx1030 build and tests below before this leaves draft.

Why this is not a duplicate

Open PRs were searched for q_gemm_rdna2_prefill, FORCE_CONFIG, prefill_direct, and hc_prefill. The only PR that deletes these hooks is #48, and its base is rdna_extra/v0.28.0. Nothing else open on rdna_extras edits this host path. This PR also drops the unregistered gptq_gemm_rdna2_prefill_direct alias, the undocumented HC prefill tile/split overrides, the false Triton sentences in q_gemm_rdna2.cu, and the force-var row in docs/rdna2/tuning.md. #48 deleted VLLM_RDNA2_PREFILL_DEBUG; this PR keeps it.

What was deleted

• csrc/rocm/q_gemm_rdna2_prefill.cu: both banners above select_config. Inner branch comments stay. select_config branches are unchanged.
• VLLM_RDNA2_PREFILL_FORCE_CONFIG early return inside select_config.
• maybe_force_split_k and VLLM_RDNA2_PREFILL_FORCE_SPLIT_K. launch_dispatch calls compute_split_k<Config*> directly.
• VLLM_RDNA2_PREFILL_DEBUG is unchanged.
• Unregistered gptq_gemm_rdna2_prefill_direct (not in ops.h or torch_bindings.cpp), its wrapper and fake in vllm/_custom_ops.py, and the fixture line in benchmarks/kernels/w4a16_compile_dispatch/test_probe_cpu.py. That line was not counted: the probe regex ends the op name on a word boundary, and the assertion still expects three gptq_gemm_rdna2_prefill calls.
• csrc/rocm/hc_prefill_rdna2.cu: VLLM_RDNA_HC_PREFILL_TILE / _SPLIT and the env_int helper that only served them. Tile selection is M <= 64 ? 0 : (M <= 384 ? 1 : 2). Split count is (target_blocks + mn / 2) / mn, then clamped to [1, 16].
• csrc/rocm/q_gemm_rdna2.cu: the three sentences that said prefill falls through to Triton. Prefill is gptq_gemm_rdna2_prefill.
• docs/rdna2/tuning.md: the one table row that documented both VLLM_RDNA2_PREFILL_FORCE_CONFIG and _FORCE_SPLIT_K. They were not two rows. That row is not in perf(rdna2): QSA/GDN HIP prefill paths vs Triton/FLA + TunableOp decode/MTP grid (implements #64) #99's hunk (which adds VLLM_RDNA_ALL_HIP and VLLM_RDNA_MROPE_BLOCK after the QSA row) and not in feat(rdna2): opt-in W8A8 int8 sdot4 dense and MoE kernels for gfx1030 #93's hunk (which inserts the W8A8 row after VLLM_RDNA2_W4A8_SDOT4).

ops.h, torch_bindings.cpp, and CMakeLists.txt are not edited, so this does not touch the #57 / #93 / #99 hunks in those files.

vllm/_custom_ops.py: #57 edits moe_skinny_int4_decode (around the old line 3072). This deletion is the unregistered prefill alias (old lines 922-950). The hunks do not overlap. The alias removal is included anyway.

tools/rdna2_028/probe_force_split.py still sets VLLM_RDNA2_PREFILL_FORCE_SPLIT_K. #87 deletes the tools/rdna2_028 probes, so this PR does not edit that file. bench_results/2026-09-29_w4a16-regression-audit/SUMMARY.md names the var in a past writeup and is left as history.

Unset env path (unchanged)

select_config used to cache atoi(VLLM_RDNA2_PREFILL_FORCE_CONFIG) (default -1) and return it when >= 0. Unset, that return never fired. The remaining branches, in order, are the same:

• size_m > 256: size_n >= 4096 returns ConfigA, otherwise ConfigC
• size_m < 4 && size_n > 4096 returns ConfigA
• 4 <= size_m <= 8 && size_n >= 2048 && size_k >= 1024 returns ConfigC
• size_m == 12 && 2560 <= size_n <= 8192 && size_k >= 512 returns ConfigC
• size_m == 32 && 5120 <= size_n <= 6144 && size_k >= 1536 returns ConfigC
• otherwise ConfigV1

launch_dispatch was maybe_force_split_k(compute_split_k<Config>(...)). That helper returned the computed split unless VLLM_RDNA2_PREFILL_FORCE_SPLIT_K was a positive integer (unset default -1). Each arm now passes compute_split_k straight through. compute_split_k itself, including the VLLM_RDNA2_PREFILL_DEBUG print, is untouched.

HC tile was forced >= 0 ? forced : (M-based). env_int returned -1 when unset, so the M-based value was already the result. HC split was forced > 0 ? forced : (target_blocks + mn / 2) / mn. Unset, forced was -1, so the formula was already the result. The [1, 16] clamp is unchanged.

A host replica of those predicates (envs unset) matched on 720 (M, N, K) shapes, 37 computed split values, 4096 HC M values, and 48 HC split pairs. Setting the vars used to change the result; that is what this removal does.

Test Plan

No ROCm toolchain in this environment (hipcc is absent, no /opt/rocm). On gfx1030, before merge:

# Build with PYTORCH_ROCM_ARCH=gfx1030, then:
.venv/bin/python -m pytest \
  tests/kernels/quantization/test_rdna2_w4a16.py \
  tests/kernels/quantization/test_rdna2_w4a16_awq.py \
  tests/kernels/quantization/test_rdna2_w4a16_selection.py -v

Host checks run here: rg for the removed names, python3 -m py_compile on the two Python files, and the unset-env replica above.

Test Result

• rg over csrc/, vllm/, benchmarks/, docs/, tests/, tools/: the only remaining reader is tools/rdna2_028/probe_force_split.py (left for cleanup(tools): delete finished tools/rdna2_028 campaign drivers #87). VLLM_RDNA2_PREFILL_DEBUG is still read in compute_split_k.
• python3 -m py_compile on vllm/_custom_ops.py and benchmarks/kernels/w4a16_compile_dispatch/test_probe_cpu.py: pass. pytest for that file was not run: this environment has no .venv and no regex.
• Brace/paren/bracket balance on the three .cu files: even. No line in those files exceeds 88 columns.
• gfx1030 build and tests/kernels/quantization/test_rdna2_w4a16*.py: not run (no hipcc, no GPU). No model eval: the unset dispatch path is the same code, and no serving run was done.

---

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. (dropped the force-var row in docs/rdna2/tuning.md)

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