Q4_0 quantization support for mmproj

[tdakhran]

On arm64 devices, Q4_0 delivers the best speed.

mmproj can be converted into the datatypes available in the conversion script {f32,f16,bf16,q8_0,tq1_0,tq2_0}.

However, when I try

bin/llama-quantize /data/playground/vlm2/LFM2-VL-1.6B/mmproj-LFM2-VL-1.6B-Q8_0.gguf /tmp/out.gguf Q4_0

an error is thrown

llama_model_quantize: failed to quantize: unknown model architecture: 'clip'

Is there a way to quantize mmproj into Q4_0?

Here are comparison results between ExecuTorch (mmproj in Q4_0) and llama.cpp on Raspberry Pi 5.

Executorch

        Prompt Tokens: 158    Generated Tokens: 33
        Model Load Time:                  2.11 (seconds)
        Total inference time:             3.46 (seconds)                 Rate:    9.52 (tokens/second)
                Prompt evaluation:        2.55 (seconds)                 Rate:   61.98 (tokens/second)
                Generated 33 tokens:      0.92 (seconds)                 Rate:   36.04 (tokens/second)
        Time to first generated token:    2.55 (seconds)

llama.cpp

llama_perf_context_print:        load time =     234.01 ms
llama_perf_context_print: prompt eval time =    4146.94 ms /   158 tokens (   26.25 ms per token,    38.10 tokens per second)
llama_perf_context_print:        eval time =     646.18 ms /    27 runs   (   23.93 ms per token,    41.78 tokens per second)
llama_perf_context_print:       total time =    4932.17 ms /   185 tokens
llama_perf_context_print:    graphs reused =          0

Prompt evaluation time is 2.55s vs 4.15s.

cc: @ngxson

[CISC]

The problem is that llama_model_quantize_impl calls model.load_arch:

llama.cpp/src/llama-quant.cpp

Line 604 in 657b8a7

model.load_arch (ml);

which does this:

llama.cpp/src/llama-model.cpp

Lines 447 to 451 in 657b8a7

	void llama_model::load_arch(llama_model_loader & ml) {
	arch = ml.get_arch();
	if (arch == LLM_ARCH_UNKNOWN) {
	throw std::runtime_error("unknown model architecture: '" + ml.get_arch_name() + "'");
	}

which only knows of these (text models):

llama.cpp/src/llama-arch.cpp

Lines 7 to 97 in 657b8a7

	static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
	{ LLM_ARCH_LLAMA, "llama" },
	{ LLM_ARCH_LLAMA4, "llama4" },
	{ LLM_ARCH_DECI, "deci" },
	{ LLM_ARCH_FALCON, "falcon" },
	{ LLM_ARCH_GROK, "grok" },
	{ LLM_ARCH_GPT2, "gpt2" },
	{ LLM_ARCH_GPTJ, "gptj" },
	{ LLM_ARCH_GPTNEOX, "gptneox" },
	{ LLM_ARCH_MPT, "mpt" },
	{ LLM_ARCH_BAICHUAN, "baichuan" },
	{ LLM_ARCH_STARCODER, "starcoder" },
	{ LLM_ARCH_REFACT, "refact" },
	{ LLM_ARCH_BERT, "bert" },
	{ LLM_ARCH_NOMIC_BERT, "nomic-bert" },
	{ LLM_ARCH_NOMIC_BERT_MOE, "nomic-bert-moe" },
	{ LLM_ARCH_NEO_BERT, "neo-bert" },
	{ LLM_ARCH_JINA_BERT_V2, "jina-bert-v2" },
	{ LLM_ARCH_BLOOM, "bloom" },
	{ LLM_ARCH_STABLELM, "stablelm" },
	{ LLM_ARCH_QWEN, "qwen" },
	{ LLM_ARCH_QWEN2, "qwen2" },
	{ LLM_ARCH_QWEN2MOE, "qwen2moe" },
	{ LLM_ARCH_QWEN2VL, "qwen2vl" },
	{ LLM_ARCH_QWEN3, "qwen3" },
	{ LLM_ARCH_QWEN3MOE, "qwen3moe" },
	{ LLM_ARCH_PHI2, "phi2" },
	{ LLM_ARCH_PHI3, "phi3" },
	{ LLM_ARCH_PHIMOE, "phimoe" },
	{ LLM_ARCH_PLAMO, "plamo" },
	{ LLM_ARCH_PLAMO2, "plamo2" },
	{ LLM_ARCH_CODESHELL, "codeshell" },
	{ LLM_ARCH_ORION, "orion" },
	{ LLM_ARCH_INTERNLM2, "internlm2" },
	{ LLM_ARCH_MINICPM, "minicpm" },
	{ LLM_ARCH_MINICPM3, "minicpm3" },
	{ LLM_ARCH_GEMMA, "gemma" },
	{ LLM_ARCH_GEMMA2, "gemma2" },
	{ LLM_ARCH_GEMMA3, "gemma3" },
	{ LLM_ARCH_GEMMA3N, "gemma3n" },
	{ LLM_ARCH_STARCODER2, "starcoder2" },
	{ LLM_ARCH_MAMBA, "mamba" },
	{ LLM_ARCH_MAMBA2, "mamba2" },
	{ LLM_ARCH_JAMBA, "jamba" },
	{ LLM_ARCH_FALCON_H1, "falcon-h1" },
	{ LLM_ARCH_XVERSE, "xverse" },
	{ LLM_ARCH_COMMAND_R, "command-r" },
	{ LLM_ARCH_COHERE2, "cohere2" },
	{ LLM_ARCH_DBRX, "dbrx" },
	{ LLM_ARCH_OLMO, "olmo" },
	{ LLM_ARCH_OLMO2, "olmo2" },
	{ LLM_ARCH_OLMOE, "olmoe" },
	{ LLM_ARCH_OPENELM, "openelm" },
	{ LLM_ARCH_ARCTIC, "arctic" },
	{ LLM_ARCH_DEEPSEEK, "deepseek" },
	{ LLM_ARCH_DEEPSEEK2, "deepseek2" },
	{ LLM_ARCH_CHATGLM, "chatglm" },
	{ LLM_ARCH_GLM4, "glm4" },
	{ LLM_ARCH_GLM4_MOE, "glm4moe" },
	{ LLM_ARCH_BITNET, "bitnet" },
	{ LLM_ARCH_T5, "t5" },
	{ LLM_ARCH_T5ENCODER, "t5encoder" },
	{ LLM_ARCH_JAIS, "jais" },
	{ LLM_ARCH_NEMOTRON, "nemotron" },
	{ LLM_ARCH_EXAONE, "exaone" },
	{ LLM_ARCH_EXAONE4, "exaone4" },
	{ LLM_ARCH_RWKV6, "rwkv6" },
	{ LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" },
	{ LLM_ARCH_RWKV7, "rwkv7" },
	{ LLM_ARCH_ARWKV7, "arwkv7" },
	{ LLM_ARCH_GRANITE, "granite" },
	{ LLM_ARCH_GRANITE_MOE, "granitemoe" },
	{ LLM_ARCH_GRANITE_HYBRID, "granitehybrid" },
	{ LLM_ARCH_CHAMELEON, "chameleon" },
	{ LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
	{ LLM_ARCH_PLM, "plm" },
	{ LLM_ARCH_BAILINGMOE, "bailingmoe" },
	{ LLM_ARCH_DOTS1, "dots1" },
	{ LLM_ARCH_ARCEE, "arcee" },
	{ LLM_ARCH_ERNIE4_5, "ernie4_5" },
	{ LLM_ARCH_ERNIE4_5_MOE, "ernie4_5-moe" },
	{ LLM_ARCH_HUNYUAN_MOE, "hunyuan-moe" },
	{ LLM_ARCH_HUNYUAN_DENSE, "hunyuan-dense" },
	{ LLM_ARCH_SMOLLM3, "smollm3" },
	{ LLM_ARCH_OPENAI_MOE, "gpt-oss" },
	{ LLM_ARCH_LFM2, "lfm2" },
	{ LLM_ARCH_DREAM, "dream" },
	{ LLM_ARCH_SMALLTHINKER, "smallthinker" },
	{ LLM_ARCH_LLADA, "llada" },
	{ LLM_ARCH_UNKNOWN, "(unknown)" },
	};

Support can be added by bypassing this for clip, but you'd have to make sure you don't quantize tensors that should not be quantized.

[ngxson]

I planned to rewrite the tool to quantize mmproj GGUF, but still haven't yet had time. It would be nicer to add it directly to llama-quantize, although I don't know if it is complicated. Will have a look when I have more time.

[tdakhran]

Agree, supporting clip in llama-quantize would be the cleanest solution.

[ngxson]

I'll have a look next week if no one else is planning on this