Performance Discrepancy with -ctv q4_0 compared to q8_0

[TkskKurumi]

Found that -ctv q4_0 is much slower than q8_0 (more than 10x slower on pp512), with following environments.
OS: Windows 10
llama.cpp build: prebuilt b7154 windows cuda12.4 from GitHub releases page. (llama-b7154-bin-win-cuda-12.4-x64.zip + cudart-llama-bin-win-cuda-12.4-x64.zip)
device: RTX 3090 24GB.
model: https://huggingface.co/unsloth/Qwen3-VL-30B-A3B-Instruct-GGUF/blob/main/Qwen3-VL-30B-A3B-Instruct-UD-Q4_K_XL.gguf
llama-bench results:

model	size	params	backend	ngl	n_ubatch	type_k	type_v	fa	dev	test	t/s
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q8_0	1	CUDA1	pp512	2915.41 ± 300.53
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q8_0	1	CUDA1	tg128	151.21 ± 0.48
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q4_0	1	CUDA1	pp512	164.29 ± 15.78
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q4_0	1	CUDA1	tg128	29.07 ± 0.90

launching command line:

.\llama-bench.exe --model "E:\LLM\Qwen3-VL-30B-A3B-Instruct\Qwen3-VL-30B-A3B-Instruct-UD-Q4_K_XL.gguf" `
                   -ctk q8_0 `
                   -ctv q8_0,q4_0 `
                   -fa 1 `
                   -ngl 99 `
                   --threads 8 `
                   --device CUDA1 `
                   -ub 4096 `
                   -b 2048 

Is this performance discrepancy expected or not?
I can think of the following possibilities:

1. The 4-bit quantization introduces additional overhead.
2. The 8-bit operations are much more optimized at the hardware level.

I would be happy to provide any additional information needed. Thank you for your time and for the continued work on this project.

[TkskKurumi]

Found that when -ctk is also Q4_0, the performance boost up back.

model	size	params	backend	ngl	n_ubatch	type_k	type_v	fa	dev	test	t/s
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q8_0	1	CUDA1	pp512	3240.28 ± 142.19
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q8_0	1	CUDA1	tg128	150.97 ± 1.08
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q4_0	1	CUDA1	pp512	164.68 ± 16.57
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q8_0	q4_0	1	CUDA1	tg128	30.97 ± 0.77
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q4_0	q8_0	1	CUDA1	pp512	172.30 ± 5.63
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q4_0	q8_0	1	CUDA1	tg128	31.18 ± 0.46
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q4_0	q4_0	1	CUDA1	pp512	3480.12 ± 29.99
qwen3vlmoe 30B.A3B Q4_K - Medium	16.49 GiB	30.53 B	CUDA	99	4096	q4_0	q4_0	1	CUDA1	tg128	150.12 ± 0.93