Q5_K and Q6_K have no AVX-512 kernel at all, so an AVX-512 host runs its largest tensor on 256-bit lanes #2605

Severity: P1 — found by the pre-release multi-platform dogfood sweep (intel/AVX-512 lane).

Measured against the published apr 0.63.0 from crates.io on the Intel receipt host
(Xeon W-3245, 16C/32T, AVX-512F/BW/CD/DQ/VL/VNNI), not a local build.

Where

crates/aprender-serve/src/quantize/fused_q5k_q6k.rs — fused_q6k_dot has only _avx2 and _simd variants; no AVX-512 sibling exists anywhere in the workspace

Defect

Q5_K and Q6_K have NO AVX-512 kernel at all, so even if F1 were fixed the AVX-512 host would still run its single largest tensor on 256-bit lanes. ISA coverage in the shipped binary is lopsided: 40 avx2 symbols vs 4 avx512 (of which 2 are the VNNI kernels proven dead in F1, plus fused_q4_0_q8_0_dot_avx512_vnni which also never fired). AVX-512 was implemented only for the Q4_K x Q8_K and Q4_0 x Q8_0 dots. This matters concretely on Q4_K_M: output.weight (the lm_head) is Q6_K [1536, 151936] at 182.6 MB — the largest single tensor in the file — and 29 of 339 tensors are Q6_K.

Reproduction

MEASURED on mac-server, published binary:
  nm /home/noah/.cargo/bin/apr | grep -E ' [tT] ' | grep -c '8realizar8quantize.*avx512vnni'  -> 2
  nm … | grep -c '8realizar8quantize.*avx512'                                                  -> 4
  nm … | grep -c '8realizar8quantize.*avx2'                                                    -> 40
  nm … | grep -c '8realizar8quantize.*scalar'                                                  -> 1
  nm … | grep -oE 'fused_q6k_dot[a-z0-9_]*'  ->  fused_q6k_dot, fused_q6k_dot_avx2, fused_q6k_dot_simd   (no avx512)
  grep -rn 'fn ggml_style' --include='*.rs' crates/  -> exactly 2, both _avx2 (q4k_dot_ggml_style.rs:38,63); no AVX-512 sibling
Runtime, same 8-token generation: fused_q6k_dot_avx2 = 8,320,128 calls (identical on Zen 4).
Tensor mix from `apr tensors` on the 1.5B Q4_K_M: 170 Q4_K, 141 F32, 29 Q6_K, largest = output.weight [1536,151936] Q6_K 182.6 MB.

Filed from the 0.64.0 pre-release sweep. Scope AVX512-SPECIFIC.