llama : suppress misleading Gemma4Assistant error during memory fitting

[leotm]

Overview

Thank you @am17an for Gemma4 MTP support ^ it introduced

llama.cpp/src/llama-context.cpp

Lines 93 to 101 in 4988f6e

	// TODO: more generic
	if (model.arch == LLM_ARCH_GEMMA4_ASSISTANT) {
	if (params.ctx_other == nullptr) {
	// TODO: change from runtime_error to llama_exception to avoid printing error message
	throw std::runtime_error("Gemma4Assistant requires ctx_other to be set (this warning is normal during memory fitting)");
	}
	cparams.ctx_other = params.ctx_other;
	}

which sent me down the wrong track earlier (my fault)

• Fix Gemma 4 MTP on llama-server (Windows) #24480

and was noted in couple follow-ups

• mtp: support for gemma-4 E2B and E4B assistants #24282 (comment)
• kv-cache: follow the source cache size when sharing cells #24267 (comment)

and couple issues

• Misc. bug: E llama_init_from_model: failed to initialize the context: Gemma4Assistant #24343 🐧
• Eval bug: Gemma4 MTP #24350 (Windows)

so i addressed the 2nd TODO only (for minimality)

if this looks as intended, happy to address the 1st TODO post-merge
(RE both LLM_ARCH_GEMMA4_ASSISTANT and LLM_ARCH_EAGLE3)

Fix: #24343
Fix: #24350

Additional information

I've tested on

• Windows 11
• unsloth/gemma-4-31B-it-qat-GGUF
• local build (unsloth.ai/docs/models/gemma-4/qat#llama.cpp-guide adapted for Win11 CMD, yup - old skool)

Before

After

NB: i noted shedrachokonofua/aether@b0d4bca but idk what it's doing and the mentioned #24376 is unrelated

P.S: i've not touched C++ in a while

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[leotm]

can update to a struct or move to src\llama-impl.h if preferred

[leotm]

could add e.g. LLAMA_LOG_INFO("%s: successfully initialized the context: %s\n", __func__); if preferred

[leotm]

*LLAMA_LOG_DEBUG

[leotm]

can rename err vars to e if preferred (seems the more common convention)

[sanmai]

This is indeed misleading; this warning should not be normal during memory fitting - it hides a bug (fitting is broken)

[ggerganov]

Also throw in the EAGLE3 case below.

[ggerganov]

This is indeed misleading; this warning should not be normal during memory fitting - it hides a bug (fitting is broken)

We can't fit before loading the target model because the assistants and eagles require a target model to be already loaded.

[sanmai]

It sounds like we can stub them with no_alloc = true

[leotm]

Also throw in the EAGLE3 case below.

updated ^