A survey on post-training quantization for Vision Transformers: Failure modes and deployment constraints

Deploying Vision Transformers (ViTs) under low-precision constraints is a system-level problem as much as an algorithmic one. Post-training quantization (PTQ) is a widely used deployment path for ViTs because it avoids retraining, but it introduces accuracy degradation and execution overhead that originate at specific operators in the ViT computation graph. This paper presents a system-level analysis of the PTQ for ViT literature, organized around a Driver–Strategy framework that links operator-level failure modes to mitigation techniques and their deployment implications. Eight Drivers are identified, including residual accumulation across encoder depth, nonlinear operator sensitivity in Softmax and GELU, coupled attention-path fragility in the Q/K/V and Q K ⊤ computation, LayerNorm instability at normalization boundaries, outlier-dominated quantization ranges, low-bit brittleness below six bits, mixed-precision allocation costs, and backend execution limits arising from kernel availability and quantize–dequantize conversion overhead. Six corresponding Strategy families describe how methods address these failure modes, with explicit attention to integer-only execution paths, operator fallback behavior, and the cost of precision conversions in inference pipelines. Explicit assignment criteria and paper-level justifications support the Driver–Strategy mappings across the complete 83-study corpus. Beyond taxonomy, the paper introduces a graph-level perspective explaining why PTQ failures concentrate at computational bottlenecks rather than distributing uniformly across layers, a failure-mode characterization by precision regime from W8A8 to sub-4-bit, five distilled design principles, and a diagnostic selection workflow for targeting specific bit-widths and hardware platforms. A corpus-wide deployment analysis shows that only 31 of the 83 studies provide deployment or hardware evidence; among these, 21 report measured latency or throughput, only five verify end-to-end integer execution on a named physical backend, and none reports a complete numerical count of quantize/dequantize (Q/DQ) boundaries. Open challenges include deployable integer-only formulations of LayerNorm and Softmax, backend-verified execution claims, and reproducible low-bit calibration protocols. This analysis serves as a conceptual reference and a practical guide for engineers targeting reliable low-precision ViT deployment on hardware stacks.