Open Access
Review
Table 3
Comparison of representative inference framework for MLLMs deploymenta)
| Framework | Core optimization | Key MLLM support | Multi-modal | Primary use case | Target platforms |
|---|---|---|---|---|---|
| vLLM [158] | PagedAttention, throughput | HF Hub seamless integration | √ | High-throughput batch processing / Stateless APIs | NV, AMD, TPU, AWS |
| SGLang [159] | RadixAttention, graph opt | Llama, DeepSeek, LLaVA | √ | Latent-sensitive chat / Complex agentic systems | NV, AMD GPUs |
| llama.cpp [160] | GGUF quantization, lightweight kernels | LLaMA series, Falcon, Qwen2-VL, LLaVA | √ | CPU-heavy inference / local prototyping | Apple, ARM, x86, NV, Mobile GPU |
| TensorRT-LLM [161] | Deep operator fusion, speculative decoding | GPT, MoE, quantized LLMs | – | Industrial-grade production on NVIDIA clusters | NV GPUs, Clusters |
| MLC-LLM [162] | TVM compiler-based, universal deployment | Llama, Mistral, GPT-2, Phi-3.5, Qwen3 | – | Universal cross-platform “write-once-run-anywhere" | iOS, Android, ARM, NV/AMD/Intel |
| MNN [163] | Heterogeneous computing, mobile-specific kernels | Qwen, Yi, InternLM, SmolM, MiMo | √ | Ultra-resource-constrained mobile NPU/GPU tasks | Mobile NPU/GPU, Vulkan, Metal |
a) NV: NVIDIA; HF Hub: Hugging Face Model Hub; GGUF: GPT-Generated Unified Format; TVM: Tensor Virtual Machine; NPU: Neural Processing Unit; AWS: Amazon Web Services; TPU: Tensor Processing Unit.
Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.
Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.
Initial download of the metrics may take a while.