#2801 [Feature]: SYCL: cheaper IQ weight decodes in the grouped expert kernels (opt-in, same bits)

Descripción

## What do you want to do?

Decode faster on Intel cards whose expert kernels are instruction-bound rather than bandwidth-bound.

On an Arc Pro B70 (SYCL port, Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS, two cards with a layer split) an ISA dump of the
lane kernels that read the routed experts (`native_gu_kernel` / `native_down_kernel` in
`sycl/src/kernels/cuda/iq_kernels.dp.cpp`) shows 280-480 instructions per 32-weight part for the part's 8 dp4a: the
Q2_0 down dot spends ~300 of its 476 on the dpct per-byte emulation of `byte_level_permute`, the IQ4_NL dot ~100 of
279 on the compare/select codebook lookup, and every format pays popcount/multiply ksigns unpacking and 64-bit
`kbx * sizeof(block)` address products each iteration. The kernels read the experts at 19-44% of the card's DRAM
bandwidth, so decode is slower than the memory system allows.

I would like an opt-in flag (`STRATA_IQ_DECODE_FAST=1`) that switches the seven IQ expert formats to kernels
producing exactly the same bits with fewer instructions: table-driven ksigns masks, the signed grid byte in one xor
and one add, Q2_0 fields spread by shifts and masks, the IQ4_NL codebook through small position tables, a loop that
walks block pointers, two rows per lane in 16-lane sub-groups, and a multi-entry form for IQ2_XXS.

Measured on two B70 (32 GB, kernel 7.1, Compute Runtime 26.22.38646.4, oneAPI 2026.1), main against a branch with
these kernels, `--layer-split 30 --max-context 65536 --kv int8 --spec 4` with the BF16 draft layer, greedy, a fixed
512-token coding reply: decode 79.5 -> 85.6 tok/s median (+7.7%), 76.0 -> 81.4 at a 40,796-token context, output
tokens and draft acceptance counts identical, prompt speed unchanged. A grouped-call microbenchmark on one card put
the Q2_0 down product at 63.7 -> 26.4 us per call and the IQ2_XXS/Q2_0 layers at 114.8 -> 59.8.

Workaround today: none; the kernels are what they are. The default path should stay byte-identical, so the change
must be opt-in and come with a parity test (`native_grouped_parity --iq`) that compares the new kernels against the
current ones bit for bit on random blobs for every IQ gate/up and down format.

Prepared with AI assistance (Claude Code): the code, the measurements and this text were produced by Claude on the submitter's two Arc Pro B70s, at the submitter's request.