[o7si]

When the model is a hybrid/recurrent model, llama_memory_seq_rm may fail. In this case, the slot's status is not SLOT_STATE_IDLE.

llama.cpp/tools/server/server-context.cpp

Lines 2220 to 2227 in 10b4f82

	if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, p0, -1)) {
	SLT_WRN(slot, "failed to truncate tokens with position >= %d - clearing the memory\n", p0);
	clear_slot(slot);
	// there is no common part left
	slot.n_prompt_tokens_cache = 0;
	}

So calling clear_slot will trigger GGML_ASSERT, causing the program to crash.

llama.cpp/tools/server/server-context.cpp

Lines 1010 to 1017 in af3be13

	void clear_slot(server_slot & slot) const {
	GGML_ASSERT(!slot.is_processing());
	SLT_WRN(slot, "clearing slot with %zu tokens\n", slot.prompt.tokens.size());
	llama_memory_seq_rm(llama_get_memory(ctx), slot.id, -1, -1);
	slot.prompt.tokens.clear();
	}

Related issue:

• Eval bug: GGML_ASSERT(!slot.is_processing()) for empty response with hybrid models (granite-4-h) #18235

[ngxson]

instead of duplicating this code, simply add a second argument in clear_slot:

clear_slot(server_slot & slot, bool allow_processing)

[o7si]

Should allow_processing have a default value? I think it should default to false to maintain the existing behavior and safety check for most cases:

void clear_slot(server_slot & slot, bool allow_processing = false);