Giuseppe Franco

Giuseppe Franco, Ian Colbert, Pablo Monteagudo-Lago et al.

Quantizing large language models (LLMs) to low-precision floating-point representations is central to efficient deployment, yet applying a single bit-width uniformly across all layers is sub-optimal in terms of both performance and accuracy. This work introduces dMX, a differentiable mixed-precision quantization framework for learnable floating-point bit-width assignment. We study its application for the microscaling floating-point (MXFP) family of data types defined by the Open Compute Project (OCP) standard. The per-layer bit-width assignment is formulated as a continuous optimization problem in which each layer's floating-point format format is parameterized by a scalar parameter, folding the multi-variate design space into a single learnable offset. During training this offset takes continuous values, avoiding sudden oscillations between discrete quantization formats. A temperature-based annealing schedule progressively discretizes the learned offsets, ensuring that the final configuration maps to hardware-compatible MXFP formats without abrupt transitions between training and inference behavior. A target-aware regularization term steers the average bit-width toward a user-specified budget, serving as a coarse-grained proxy for inference cost and balancing model quality against deployment efficiency. We performed experiments on different families of LLM, such as Llama, Qwen3, and SmolLM2, evaluating perplexity on WikiText-2 and accuracy on four zero-shot reasoning benchmarks. Across these settings, dMX consistently yields Pareto-dominating models and improves over Kullback-Leibler (KL) divergence-based layer-selection heuristics, efficiently navigating trade-offs between model quality and average bit-width.

Ian Colbert, Giuseppe Franco, Fabian Grob et al.

When quantizing weights and activations to increasingly narrower representations, the cost of additions begins to dominate that of multiplications in multiply-accumulate (MAC) units. Recent studies show that reducing addition costs via low-precision accumulation improves throughput, power, and area across inference platforms, albeit with an increased risk of overflow. Accumulator-aware quantization research has so far only considered the quantization-aware training (QAT) paradigm, in which models are fine-tuned or trained from scratch with quantization in the loop. As models and datasets continue to grow in size, QAT techniques become increasingly more expensive, which has motivated the recent surge in post-training quantization (PTQ) research. To bridge this gap, we introduce AXE, the first accumulator-aware quantization framework explicitly designed to endow overflow avoidance guarantees to PTQ algorithms. We present theoretical motivation for AXE and demonstrate its flexibility by implementing it on top of two existing algorithms: GPFQ and OPTQ. We design AXE to support multi-stage accumulation, opening the door to full datapath optimization for the first time. We evaluate AXE using recent language generation models; when quantizing Llama3 8B for a 16-bit multi-stage accumulation datapath, AXE maintains up to 98% of the FP16 perplexity, surpassing naive bit width manipulation by up to 15%.

Shivam Aggarwal, Hans Jakob Damsgaard, Alessandro Pappalardo et al.

Post-training quantization (PTQ) is a powerful technique for model compression, reducing the numerical precision in neural networks without additional training overhead. Recent works have investigated adopting 8-bit floating-point formats(FP8) in the context of PTQ for model inference. However, floating-point formats smaller than 8 bits and their relative comparison in terms of accuracy-hardware cost with integers remains unexplored on FPGAs. In this work, we present minifloats, which are reduced-precision floating-point formats capable of further reducing the memory footprint, latency, and energy cost of a model while approaching full-precision model accuracy. We implement a custom FPGA-based multiply-accumulate operator library and explore the vast design space, comparing minifloat and integer representations across 3 to 8 bits for both weights and activations. We also examine the applicability of various integerbased quantization techniques to minifloats. Our experiments show that minifloats offer a promising alternative for emerging workloads such as vision transformers.

Sai Sanjeet, Ian Colbert, Pablo Monteagudo-Lago et al.

Recent post-training quantization (PTQ) methods have adopted block rotations to diffuse outliers prior to rounding. While this reduces the overhead of full-vector rotations, the effect of block structure on outlier suppression remains poorly understood. To fill this gap, we present the first systematic, non-asymptotic analysis of outlier suppression for block Hadamard rotations. Our analysis reveals that outlier suppression is fundamentally limited by the geometry of the input vector. In particular, post-rotation outliers are deterministically minimized when the pre-rotation $\ell_1$ norm mass is evenly distributed across blocks. Guided by these insights, we introduce MixQuant, a block rotation-aware PTQ framework that redistributes activation mass via permutations prior to rotation. We propose a greedy mass diffusion algorithm to calibrate permutations by equalizing the expected blockwise $\ell_1$ norms. To avoid adding inference overhead, we identify permutation-equivariant regions in transformer architectures to merge the resulting permutations into model weights before deployment. Experiments show that MixQuant consistently improves accuracy across all block sizes, recovering up to 90% of the full-vector rotation perplexity when quantizing Llama3 1B to INT4 with block size 16, compared to 46% without permutations.

Giuseppe Franco, Pablo Monteagudo-Lago, Ian Colbert et al.

The size of a model has been a strong predictor of its quality, as well as its cost. As such, the trade-off between model cost and quality has been well-studied. Post-training optimizations like quantization and pruning have typically focused on reducing the overall volume of pre-trained models to reduce inference costs while maintaining model quality. However, recent advancements have introduced optimization techniques that, interestingly, expand models post-training, increasing model size to improve quality when reducing volume. For instance, to enable 4-bit weight and activation quantization, incoherence processing often necessitates inserting online Hadamard rotations in the compute graph, and preserving highly sensitive weights often calls for additional higher precision computations. However, if application requirements cannot be met, the prevailing solution is to relax quantization constraints. In contrast, we demonstrate post-training model expansion is a viable strategy to improve model quality within a quantization co-design space, and provide theoretical justification. We show it is possible to progressively and selectively expand the size of a pre-trained large language model (LLM) to improve model quality without end-to-end retraining. In particular, when quantizing the weights and activations to 4 bits for Llama3 1B, we reduce the gap to full-precision perplexity by an average of 9% relative to both QuaRot and SpinQuant with only 5% more parameters, which is still a 3.8% reduction in volume relative to a BF16 reference model.