Misc. bug: The Llama server starts with 4 slots by default.

[sxch775-work]

Name and Version

llama server
llama-b7079-bin-win-cpu-x64

Operating systems

No response

Which llama.cpp modules do you know to be affected?

No response

Command line

C:\Users\21\Downloads\llama-b7079-bin-win-cpu-x64\llama-server.exe -m "C:\Users\21\Downloads\Qwen3-VL-4B-Instruct-Q8_0.gguf" -c 8000 --port 1280 --api-key 123 --mmproj C:\Users\21\Downloads\mmproj-Qwen3-VL-4B-Instruct-Q8_0.gguf -np 1

Problem description & steps to reproduce

I used the following command to start the llama server, expecting to launch one slot with a context of 8000, but the llama server started four slots, each with a context of 8000.
this is the command
C:\Users\21\Downloads\llama-b7079-bin-win-cpu-x64\llama-server.exe -m "C:\Users\21\Downloads\Qwen3-VL-4B-Instruct-Q8_0.gguf" -c 8000 --port 1280 --api-key 123 --mmproj C:\Users\21\Downloads\mmproj-Qwen3-VL-4B-Instruct-Q8_0.gguf -np 1

this is the start log
load_hparams: model size: 800.43 MiB
load_hparams: metadata size: 0.11 MiB
alloc_compute_meta: warmup with image size = 1472 x 1472
alloc_compute_meta: CPU compute buffer size = 322.49 MiB
alloc_compute_meta: graph splits = 1, nodes = 766
warmup: flash attention is enabled
srv load_model: loaded multimodal model, 'C:\Users\21\Downloads\mmproj-Qwen3-VL-4B-Instruct-Q8_0.gguf'
srv init: initializing slots, n_slots = 4
slot init: id 0 | task -1 | new slot, n_ctx = 8192
slot init: id 1 | task -1 | new slot, n_ctx = 8192
slot init: id 2 | task -1 | new slot, n_ctx = 8192
slot init: id 3 | task -1 | new slot, n_ctx = 8192
srv init: prompt cache is enabled, size limit: 8192 MiB
srv init: use --cache-ram 0 to disable the prompt cache
srv init: for more info see #16391
srv init: thinking = 0

First Bad Commit

No response

Relevant log output

[o7si]

@sxch775-work Hi, the number of slots is bound to the n_parallel parameter, as shown in the code below:

llama.cpp/tools/server/server.cpp

Lines 2513 to 2514 in 416e7c7

	for (int i = 0; i < params_base.n_parallel; i++) {
	server_slot slot;

In #16736, this code was added:

llama.cpp/tools/server/server.cpp

Lines 4477 to 4482 in 416e7c7

	if (params.n_parallel == 1 && params.kv_unified == false && !params.has_speculative()) {
	LOG_WRN("%s: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)\n", __func__);
	params.n_parallel = 4;
	params.kv_unified = true;
	}

I haven't read the PR carefully, but you can use the -kvu parameter to ensure that the n_parallel value doesn't get modified to 4.

[dark-penguin]

Interesting. I wonder what is the rationale behind changing those values in this case?

EDIT: I guess this is to avoid extra swapping with RAM as long as all contexts fit in VRAM?

[sxch775-work]

@sxch775-work Hi, the number of slots is bound to the n_parallel parameter, as shown in the code below:

llama.cpp/tools/server/server.cpp

Lines 2513 to 2514 in 416e7c7

for (int i = 0; i < params_base.n_parallel; i++) {
server_slot slot;
In #16736, this code was added:

llama.cpp/tools/server/server.cpp

Lines 4477 to 4482 in 416e7c7

if (params.n_parallel == 1 && params.kv_unified == false && !params.has_speculative()) {
LOG_WRN("%s: setting n_parallel = 4 and kv_unified = true (add -kvu to disable this)\n", func);

 params.n_parallel = 4; 
 params.kv_unified = true; 

}
I haven't read the PR carefully, but you can use the -kvu parameter to ensure that the n_parallel value doesn't get modified to 4.

thank you