UPSTREAM PR #17728: ggml-cpu: aarm64: remove asserts always evaluating to false in repack

[loci-dev]

Mirrored from ggml-org/llama.cpp#17728

I got an email reporting the issue. I can't find the original comment, but for gemv these asserts don't make sense, as nr will always be 1

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Review Summary - PR #413

Analysis: This PR removes two incorrect assertions from ARM64 GEMV functions (ggml_gemv_q4_K_8x4_q8_K and ggml_gemv_q4_K_8x8_q8_K) in repack.cpp. The assertions assert(nr % 4 == 0) were semantically wrong for vector operations where nr = 1 is expected.

Performance Impact: The modified GEMV functions are not among the top performance-impacted functions. The observed performance changes in the version comparison (parameter handling functions showing 7-12 ns increases in throughput) are unrelated to this PR. The assertion removal eliminates 1-2 instructions per call, resulting in negligible improvement (<0.1 ns).

Inference Impact: No impact on tokens per second. The core inference functions (llama_decode, llama_encode, llama_tokenize) are unchanged. GEMV operations are utility functions in the quantized computation pipeline, not direct inference entry points.

Power Consumption: The 0.21% reduction in libggml-cpu.so power consumption (244 nJ decrease) is driven by the 88 ns throughput improvement in ggml_compute_forward_silu_back_f16, not by this PR's changes.

Conclusion: This is a correctness fix with no measurable performance impact. The change enables proper execution of GEMV operations without affecting inference throughput or power consumption.