vulkan: use graphics queue on AMD

[0cc4m]

I'm not sure why, but the graphics queue is slightly faster in tg on AMD than the compute queue, and this also fixes the partial offload issue I fixed in #19976, so the second queue no longer has to be enabled by default. I got the idea from @zedbytes reporting that tg goes up when running with RADV_DEBUG=nocompute.

AMD RX 9070 XT

model	size	params	ngl	fa	test	t/s (before)	t/s (after)	diff
llama 8B Q4_0	4.33 GiB	8.03 B	20	1	pp512	2288.04 ± 2.42	2225.76 ± 2.31	-2.7%
llama 8B Q4_0	4.33 GiB	8.03 B	20	1	tg128	24.33 ± 0.04	24.58 ± 0.05	+1.0%
llama 8B Q4_0	4.33 GiB	8.03 B	99	1	pp512	4886.26 ± 105.08	4901.77 ± 102.66	+0.3%
llama 8B Q4_0	4.33 GiB	8.03 B	99	1	tg128	115.78 ± 0.02	121.39 ± 0.02	+4.8%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	20	1	pp512	736.21 ± 9.37	735.19 ± 7.51	-0.1%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	20	1	tg128	39.53 ± 0.10	40.36 ± 0.21	+2.1%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	99	1	pp512	3383.58 ± 29.26	3425.38 ± 28.68	+1.2%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	99	1	tg128	200.45 ± 1.89	220.41 ± 1.46	+10.0%

AMD Radeon Pro VII

model	size	params	ngl	fa	test	t/s (before)	t/s (after)	diff
llama 8B Q4_0	4.33 GiB	8.03 B	20	1	pp512	636.62 ± 9.07	615.62 ± 0.79	-3.3%
llama 8B Q4_0	4.33 GiB	8.03 B	20	1	tg128	38.35 ± 0.09	38.20 ± 0.01	-0.4%
llama 8B Q4_0	4.33 GiB	8.03 B	99	1	pp512	830.30 ± 1.51	834.44 ± 1.05	+0.5%
llama 8B Q4_0	4.33 GiB	8.03 B	99	1	tg128	102.45 ± 0.64	100.28 ± 0.24	-2.1%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	20	1	pp512	289.76 ± 3.59	287.75 ± 3.10	-0.7%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	20	1	tg128	34.57 ± 0.32	34.05 ± 1.20	-1.5%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	99	1	pp512	749.65 ± 5.42	762.52 ± 5.89	+1.7%
qwen3moe 30B.A3B Q2_K - Medium	10.48 GiB	30.53 B	99	1	tg128	94.70 ± 0.46	97.55 ± 0.20	+3.0%

[zedbytes]

nice one , i am guessing that after this change RADV_DEBUG=nocompute is set by default for AMD GPUs ?

[0cc4m]

It won't be needed after this change. With nocompute, it disabled the compute queue, so the backend used the graphics queue instead. That is what made it faster (but don't ask me why). This PR changes it to use the graphics queue even if a compute-only queue is available.

[winstonma]

I just installed this version and I am using llama-server with immersive translate addon to translate the web article for me. I set the addon to send an API request to the llama-server every 10 seconds just to make some translation.

With this feature now my KDE desktop lag when the LLM is processing. I also tried playing youtube video on Firefox and it drop 30% of the frame while playing 1080p video. Does this happen on your side?

Sorry I forgot to provide more info. My laptop is running AMD Ryzen AI 360 with 880M iGPU.

[Neutralized]

I downloaded latest release and my tg/s dropped about 40% compared to release b8333 ! I use 2 Radeon cards 9070XT and 6700 XT, i would suggest removing changes from this update until further tested

[0cc4m]

I downloaded latest release and my tg/s dropped about 40% compared to release b8333 ! I use 2 Radeon cards 9070XT and 6700 XT, i would suggest removing changes from this update until further tested

Linux or Windows?

[Neutralized]

Windows

[Neutralized]

#20597