Set max prefill parallelization independent of -np?

[Kangaroux]

I normally use -np 1 because of memory constraints but switched to -np 2 for a session that was using parallel subagents. I noticed when restoring the session with -np 2, prefill took half as much time.

With -np 1 it takes 79.25s to process 41%

[59129] 0.10.682.876 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   4096, progress = 0.10, t =   6.27 s / 653.43 tokens per second
[59129] 0.17.297.685 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =   8192, progress = 0.21, t =  12.88 s / 635.86 tokens per second
[59129] 0.17.297.889 I slot update_slots: id  0 | task 0 | 8192 tokens since last checkpoint at 0, creating new checkpoint during processing at position 12288
[59129] 0.17.402.711 I slot create_check: id  0 | task 0 | created context checkpoint 1 of 32 (pos_min = 8191, pos_max = 8191, n_tokens = 8192, size = 163.782 MiB)
[59129] 0.44.577.827 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =  12288, progress = 0.31, t =  40.16 s / 305.95 tokens per second
[59129] 1.23.668.616 I slot print_timing: id  0 | task 0 | prompt processing, n_tokens =  16384, progress = 0.41, t =  79.25 s / 206.73 tokens per second

With -np 2 that drops to 33.65s

[51947] 0.14.734.173 I slot print_timing: id  1 | task 0 | prompt processing, n_tokens =   4096, progress = 0.10, t =   7.76 s / 527.84 tokens per second
[51947] 0.22.856.786 I slot print_timing: id  1 | task 0 | prompt processing, n_tokens =   8192, progress = 0.21, t =  15.88 s / 515.79 tokens per second
[51947] 0.22.856.947 I slot update_slots: id  1 | task 0 | 8192 tokens since last checkpoint at 0, creating new checkpoint during processing at position 12288
[51947] 0.22.960.893 I slot create_check: id  1 | task 0 | created context checkpoint 1 of 32 (pos_min = 8191, pos_max = 8191, n_tokens = 8192, size = 163.782 MiB)
[51947] 0.31.560.681 I slot print_timing: id  1 | task 0 | prompt processing, n_tokens =  12288, progress = 0.31, t =  24.59 s / 499.79 tokens per second
[51947] 0.40.619.602 I slot print_timing: id  1 | task 0 | prompt processing, n_tokens =  16384, progress = 0.41, t =  33.65 s / 486.96 tokens per second

Is this something that can be split out from the -np behavior to avoid splitting up the context size?

[OPS-NeoRetro]

What is your hardware?

[Kangaroux]

7900XTX 5700x3d 32gb ddr4. I use qwen3.6-27B so I try to keep it entirely in vram (24gb)

[OPS-NeoRetro]

So, RX 7900 XTX, and what type of Qwen 3.6 do you use, because I think you might use Q4_K_M or Q6_K? What command did you use and what IDE do you use, because it seems like you're using Claude Code, Cursor IDE or something similar. Are you using the ROCm version for local AI agents on your code?