Correct native HRX explicit RoPE frequency factors

Row: BACKEND-ROCM

Native HRX's Qwen postprocess matcher treats ggml_rope_ext's explicit freq_factors tensor as an inverse-frequency table. The CPU reference divides the base theta by each factor, while the HRX kernel multiplies the position by each supplied value.

Application pin: AMD-Ecosystem/llama.cpp 6319038132ed12f968ea68f37753f705da830ea8. Runtime pin: ROCm/hrx-system 6bcd5a4ff111fa5bf160ab9f4592ca8e7cc810b1, with the separately retained #3081 Loom compilation repair. The #3083 scheduling repair makes the BF16 Qwen postprocess region reachable.

Source proof: ggml/src/ggml-cpu/ops.cpp:5852 computes theta/ff. ggml/src/ggml-hrx/dispatch_registration/qwen/dispatch-qwen-attention-postprocess.cpp:410 binds the explicit tensor directly. ggml/src/ggml-hrx/kernel-corpus/kernels/qwen_moe/qwen3_moe/attention_postprocess_f32_f16.loom:44 computes position * inverse_frequencies. The upstream test fixture creates inverse frequencies and supplies them as ggml factors at tests/test-hrx-ops.cpp:737 and :2790. Production Qwen3 uses nullptr at src/models/qwen3.cpp:91 and :99, so it takes the separate implicit mode.

The extended upstream numerical fixture uses BF16 weights, 16 tokens, input1024, Q16, KV8, cache4096, and the unchanged upstream threshold 2 + 0.05 * abs(CPU). GPU relay aborts at output index2865: HRX0.61159, CPU-1.4684, difference2.07999, allowed2.07342. This index uses token1, head6, channel49. The same supplied float0.0008659643353894353 gives CPU angle1.0000017 and HRX angle0.000865964. The original quantized two-token fixture passes its unchanged tolerance.

Evidence: /home/vikash/vllm.cpp-hrx-evaluation/build-hrx-evaluation/native-prefill-repaired-numerics retains command, complete artifact hashes, exit-6, and stderr. The source calculation is retained in /home/vikash/vllm.cpp-hrx-prefill-repair-3083/build-hrx-prefill-repair/evidence/explicit-rope-diagnosis.json.

Owner: BACKEND-ROCM's #3080 evaluation operator. The #3083 scheduling spec lists this under Owed. The scoped scheduling repair does not change kernel arithmetic or tolerances. Resolve the explicit mode under a committed scope, keep its reproducer, and independently review any semantic repair. This issue does not explain the separate corrupted model output in the implicit mode.

Start with ggml/src/ggml-cpu/ops.cpp:5852, dispatch-qwen-attention-postprocess.cpp:410, and attention_postprocess_f32_f16.loom:44 to compare explicit frequency handling. Run the extended fixture described in tests/test-hrx-ops.cpp and review the retained diagnostic artifacts. Done means the explicit mode agrees with the CPU reference while the implicit mode, original quantized fixture, and stated tolerances remain unchanged.