MPBRQ - A Framework for Mixed-Precision Quantization for Large Language Models
Bibliographic record
Abstract
State-of-the-art Large Language Models (LLMs) have achieved impressive performance on varioustasks, yet their sizes are also increasing rapidly, requiring large amounts of computing and memory resources. Quantization has emerged as a promising method to reduce these costs and enable highly efficient inference. Based on the hypothesis that different parts of the model have varying quantization sensitivity and precision requirements, we introduce MPBRQ, a mixed-precision quantization framework that efficiently quantizes each network layer to appropriate precision. Evaluated on the Llama2-7B model, MPBRQ produces quantized models with better performance than the Omniquant baseline — one of the best-performing uniform precision methods — does under similar model footprints. MPBRQ achieves 10% improvement in average accuracy on zero-shot tasks for around 4 bits average bitlength and closes the accuracy gap with the unquantized model to only 0.75% at 6 bits average bitlength. Furthermore, MPBRQ provides custom CUDA kernels that enable efficient execution of such mixed precision models on conventional GPUs. With the mixed precision bitlengths found for Llama2-7B, these custom kernels result in at most 1.97 times speedup for an individual kernel invocation, up to 1.42 times overall acceleration for all Linear layers, and an estimated upper-bound of 1.33 times speedup for the entire network, with respect to the unquantized layers in FP16.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.014 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.003 | 0.005 |
| Open science | 0.005 | 0.004 |
| Research integrity | 0.002 | 0.005 |
| Insufficient payload (model declined to judge) | 0.012 | 0.006 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".