feat(KERNEL-QUANT-CIQ-GEMM-ROCM-RDNA3): enable gfx1100 WMMA prefill

[VikashLoomba]

Enable Q4_K and Q6_K WMMA prefill on gfx1100 through the existing quantized matmul dispatch. Both tile bodies remain byte-identical; the change adds quant-specific architecture admission. Attention retains its separate gate, and VT_ROCM_QUANT_WMMA=0 retains the scalar control.

Row

KERNEL-QUANT-CIQ-GEMM-ROCM-RDNA3; resolves ISSUE-LOCAL-01M2F0PQWGSCXG0N4951NF9DPZ.

Before starting

The current source at cef9f821632eb33156bb7d24a0a85eeab7f7484a excludes gfx1100 from the quantized WMMA compile and runtime gates. The existing physical tests could skip that device. The issue/PR and ownership search found no existing gfx1100 admission change. The developer selected one PR for specification and implementation. Specification commit a5b5c92 precedes implementation c3fe98ba6.

Evidence

• The gfx1100 red-first hardware run fails ten dispatch assertions before admission; numerical scalar checks still pass.
• Four focused physical cases pass all 46 assertions, including Q4_K/Q6_K, F32/BF16, partial blocks and asymmetric tails.
• The original upstream MMQ corpus passes all 240 cases against primary outputs and dense references with unchanged tolerances.
• The public completion test checks both launches, all 1,024 logits and eight tokens against a separate scalar process. Removing either launch breaks output checks even with its counter retained.
• The fresh reviewer reports PASS with no findings after architecture, launch, scalar-control and corpus-integrity mutations. The operator independently rebuilt and reran the hardware/public gates.
• A separate final-record reviewer reports PASS with no findings at d6e40c91f634a041c873c7a04516d55c4d05772a. The review verifies receipt mappings and hashes, authored links, lifecycle records, and the explicit performance qualifications.
• The broad HIP gate has 24 executed tests passing and five paired fixture/device skips. The architecture suite passes 16 cases and 109 assertions. The emitted gfx1100 object contains signed-int8 WMMA instructions.

Exact commands, hashes, red/green results, mutation restorations and review evidence are in docs/bench-evidence/rocm-rdna3-quant-wmma/README.md.

Preflight qualification

The implementer and reviewer full staged preflights exit zero with no failed checks and 12 explicit baseline skips. These are not all-green preflights. Seven NumPy-dependent suites subsequently run with an existing isolated NumPy installation; all exit zero, with five model-dependent CLIP subcases still skipped. Four uninvoked architecture checks concern unchanged ARM/CPU/CUDA/Triton paths. The final PR classification check runs separately with its required arguments.

The final operator agent-ready.py run at d6e40c91f634a041c873c7a04516d55c4d05772a reports zero failed checks and compiles all eight affected host translation units. It exits one because its strict preflight skips five argument-dependent checks: ARM ISA, CPU ISA, CUDA gencode, PR classification, and Triton AOT. PR classification passes separately locally and in CI. The other four concern unchanged architecture paths. This is a qualified result, not an all-green readiness claim.

The operator's initial coordinator branch name caused the role checker to treat pending work as landed. Renaming that task branch to the required row/ form resolves the check without changing source or the checker. Initial receipts and scoped resolutions remain preserved. These narrowly scoped baseline/resource qualifications do not waive the gfx1100 correctness, reachability or independent-review gates.

Model evidence and remaining work

The real-model workload uses two raw prompts of 183 and 174 tokens, repeated four times, with 16 greedy output tokens per request. All native input IDs match the primary tokenizer exactly. All 128 completion IDs match across six interleaved native on/off processes and both upstream runtimes. Matching native traces show 1,216 WMMA launches confined to eight prefill windows and none during 120 decode windows; the scalar control shows zero WMMA launches.

Fresh builds of vLLM main 39545e475d3627287ff69c25465dc0bd405f67e1 and llama.cpp master 093a2f86c3e37c54fa3e1f9efb17b304f3433abd execute the retained Qwen3.5-4B Q4_K_M workload; the GGUF plugin main remains d4c1f0d082fc7cd4350da56689109a01c1f29d6c.

Other gfx11 devices and attention WMMA admission remain excluded. Full-model performance floors remain failing under ISSUE-LOCAL-01M2F4WCD6ZK5VH5S8TF83APD6. Dynamic-clock timing observations and sampled memory records carry explicit limits; this PR does not claim accepted clock-attributed speedups, whole-model performance parity, or equal-cache-capacity memory efficiency. RDNA4 hardware was not rerun; its architecture admission and tile bodies are preserved.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:gpt-6-astra [Codex]