crash setting the llm_sampler_init_xxx parameters

[danielebriggi]

            SELECT llm_sampler_init_temp(1.0);
            SELECT llm_sampler_init_top_k(64);
            SELECT llm_sampler_init_top_p(0.95, 1);
            SELECT llm_sampler_init_min_p(0.0, 1);
            SELECT llm_sampler_init_dist(-1);
            SELECT llm_sampler_init_penalties(1024, 1.1, 0.0, 0.0);

If you try to load one after the other, you get a combination of case where the connection crashes. I think that this is related to the dist seed number (llm_sampler_init_dist();). When not called after these llm_sampler_init_XX connection crashes.

Play with llm_sampler_init_xxx:

SELECT llm_model_load('/home/sqlitecloud/repositories/sqlite-rag/models/unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q8_0.gguf');

SELECT llm_context_create('context_size=32000');
SELECT llm_chat_create();

SELECT llm_sampler_init_temp(1.0);
SELECT llm_sampler_init_top_k(64);
SELECT llm_sampler_init_top_p(0.95, 1);
SELECT llm_sampler_init_min_p(0.0, 1);
-- SELECT llm_sampler_init_dist(-1);
SELECT llm_sampler_init_penalties(1024, 1.1, 0.0, 0.0);

SELECT llm_chat_respond('my question is..');

[andinux]

You must add (tipically at the end of the chain) at least one sampler that performs the final selection step; llm_sampler_init_dist is the usual choice, optionally seeded for determinism.

from llama.h

// typically, the chain should end with a sampler such as "greedy", "dist" or "mirostat"
// this sampler will be responsible to select the actual token

if a dist or greedy step are not added at the end of the sampler chain than the following call to llama_sampler_sample(sampler, ctx, -1); will crash.

[danielebriggi]

but why I have to set it after the other init_xx? If the default for us as sqlite-ai is -1, it should be the default even after I change the settings of the model via init_xxx. Isn't it?
I assume the default for us is -1 because the inference works