Add Q2_K reordered MMVQ and ESIMD kernels

[malsbat]

Overview

Continuing the work in #25942 to add the Q2_K kernel. In order to support the reordered ESIMD it is necessary to support reordered MMVQ as the MMVQ path is enabled in multi-token calls.

Measured tg improvement of 1.54x to 1.90x (depending on hardware) using Qwen35-27B-Q2_K_M.

Additional information

Note: all measurements were captured with #25741 applied locally as the parent PR does not include the newer upstream fix. This only affects pp numbers which are not the important part of this PR.

PVC performance

Baseline (unreordered MMVQ):

export GGML_SYCL_ENABLE_OPT=0
export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=0

model	size	params	backend	ngl	sm	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	pp512	886.33 ± 4.67
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	tg128	11.48 ± 0.00

Reordered MMVQ

export GGML_SYCL_ENABLE_OPT=1
export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=0

model	size	params	backend	ngl	sm	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	pp512	883.92 ± 3.91
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	tg128	14.06 ± 0.00

DMMV

export GGML_SYCL_ENABLE_OPT=1
export GGML_SYCL_PRIORITIZE_DMMV=1
export GGML_SYCL_ENABLE_ESIMD=0

model	size	params	backend	ngl	sm	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	pp512	888.42 ± 3.82
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	tg128	14.09 ± 0.00

ESIMD

export GGML_SYCL_ENABLE_OPT=1
export GGML_SYCL_PRIORITIZE_DMMV=0
export GGML_SYCL_ENABLE_ESIMD=1

model	size	params	backend	ngl	sm	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	pp512	888.90 ± 1.99
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	none	tg128	17.63 ± 0.13

Perplexity tests across wikitext, hellaswag, and winogrande showed no significant changes between unreordered MMVQ, reordered MMVQ, and ESIMD.

B70 for same 4 scenarios

model	size	params	backend	ngl	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	pp512	668.22 ± 1.60
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	tg128	10.74 ± 0.03

model	size	params	backend	ngl	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	pp512	678.17 ± 1.58
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	tg128	15.19 ± 0.07

model	size	params	backend	ngl	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	pp512	674.22 ± 1.05
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	tg128	17.70 ± 0.09

model	size	params	backend	ngl	test	t/s
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	pp512	672.23 ± 1.92
qwen35 27B Q2_K - Medium	11.21 GiB	27.32 B	SYCL	100	tg128	20.43 ± 0.13

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, used Claude heavily to author new kernels based on existing patterns.

[malsbat]

Same as the Q3 PR, I will rebase and mark ready for review once #25942 is merged.

[arthw]

With Qwen3.5-27B-UD-Q2_K_XL.gguf on B60.

Test	fa	Base t/s	Primary t/s	Increase Rate (Primary vs Base)
pp512	0	172.46	172.44	-0.01%
pp512	1	173.57	169.84	-2.15%
tg128	0	10.82	12.27	13.40%
tg128	1	10.83	12.28	13.39%

[arthw]

It's good job!

Thank you!

[CISC]

https://github.com/ggml-org/llama.cpp/actions/runs/32461271351/job/96735363159#step:7:311

[ggerganov]

Weird that the CI passed before merging

[CISC]

Weird that the CI passed before merging

It wasn't rebased after #26779 merged.