The paper investigates how exporting ternary language models (BitNet, Falcon‑E, BitCPM) through a bf16 cast step can introduce significant discrepancies between the fine‑tuned latent weights and the deployed ternary codes. In three lab pipelines, the authors find that fp32 quantization of shipped latents disagrees with the deployed codes on up to 1.77% of codes, and that the export step can drastically reduce strict accuracy on GSM8K (e.g., from 58.79% to 0.78% for Falcon‑E‑1B‑Base). They propose two compatibility remedies—directly writing the training quantizer’s codes or adjusting bf16 inputs—to meet a 4‑point strict‑accuracy non‑inferiority criterion across all models.
arXiv:2607. 26515v1 Announce Type: new Abstract: We present, to our knowledge, the first end-to-end FP4 RL post-training, in which both the rollout and training policies, including their forward and backward passes, operate at 4-bit precision.
The paper reports a large‑scale post‑training ternarisation of the Qwen3 language model, extending a conversion pipeline from the 4B to the 8B variant. Using KOTMS rotation, E2M‑ATQ adaptive ternarisation, and GPTQ‑style error compensation, the authors achieve a 1.361× perplexity ratio across three corpora and retain 78.5% of the FP16 accuracy on zero‑shot tasks, with the 8B model outperforming the 4B by 8.9 percentage points. The study also demonstrates lossless lattice‑aware packing, producing an 8.24 GiB checkpoint that preserves perplexity, and shows that direct packed execution can reach 15.52 tokens/s in 7.35 GiB, though packed GEMV remains slower than FP16 cuBLAS.
The paper reports a post‑training ternarization of the 4‑billion‑parameter Qwen model, achieving an effective 1.641‑bit representation for 81.62 % of its weights while keeping activations at 16‑bit precision. Accuracy drops from 64.5 % to 54.7 % across ten capability tests, with uneven degradation (e.g., BoolQ 84.6 % of teacher performance, ARC‑Challenge 43.8 %). After packing the ternary planes, the model size shrinks from 8.29 GiB to 3.96 GiB with negligible change in perplexity, though inference speed is not improved.
arXiv:2609.39223v2 Announce Type: new Abstract: Large language model (LLM) inference is increasingly moving toward lower precision to realize the throughput of hardware accelerators, but aggressive p...
TriPLU is a Trilinear Product Linear Unit that replaces the gated feed‑forward branch in tiny decoder‑only language models with a degree‑3 product‑only branch that multiplies three projected streams coordinate‑wise. In a character‑level TinyStories 1M‑byte prefix study, TriPLU achieves a mean best validation loss of 1.0637, outperforming closely matched SwiGLU (1.1017), a degree‑4 product control (1.0780), and a degree‑2 control (1.1026). In train‑only Byte‑BPE experiments, TriPLU also lowers validation and held‑out bits per byte on TinyStories and WikiText‑2 raw under low‑learning‑rate settings, with PMI‑slice evidence indicating gains on seen middle‑ and high‑PMI adjacent‑token pairs.
arXiv:2609.00224v1 Announce Type: cross Abstract: Weight-only post-training quantization (PTQ) can alleviate the computational burden of serving large language models (LLMs) at scale. However, existi...
The paper demonstrates that greedy decoding from large language models is not precision‑invariant: the same model, prompt, and decoding algorithm can produce different outputs when run in BF16 versus FP16 on identical hardware. Across six models (1.1B–7B parameters, four families, and 12B) and three benchmarks, 49–100 % of prompts diverge, with a single token flip often cascading into trajectory‑level divergence. The authors develop an empirical error‑propagation analysis that identifies the top‑two logit margin at the LM head as the key factor, and they propose a low‑overhead intervention—selective FP32 LM head recomputation—that improves exact agreement by 22–36 percentage points with less than 4 % latency overhead. "whyItMatters":"The findings reveal that precision choices can fundamentally alter model outputs, challenging the assumption of deterministic greedy decoding and highlighting the need for precision‑aware inference strategies."
arXiv:2606. 09686v1 Announce Type: cross Abstract: Numeric format proliferation in machine learning hardware -- FP8 (E4M3 and E5M2), BF16, MXFP4, microscaling block formats, and dozens of research variants -- has outpaced the availability of vendor-neutral, bit-exact reference material.
The study re‑examines a reported advantage of a routed ternary (1.58‑bit) language model over a full‑precision transformer at 60K parameters. By running controlled experiments with multiple seeds and a fixed training recipe, the authors find that the apparent benefit largely stems from the choice of baseline model shape rather than the ternary architecture itself. While the routed model does outperform other shapes at a larger 130M‑byte budget, its advantage diminishes when a plain gated diagonal‑SSM block is used, and the ternary penalty varies with architecture and quantization details.
arXiv:2608.28911v1 Announce Type: new Abstract: The key-value (KV) cache is the dominant memory bottleneck of long-context large language model (LLM) inference, growing linearly with context length....
The paper introduces format‑aware fusion, a method that co‑designs quantization producers with their scale domains and consumer layouts to fully exploit four‑bit floating‑point (FP4) Tensor Cores. Using this approach, the authors pretrain the Llama‑3‑family 8B model on 160 billion tokens, achieving up to 37.9 K tokens/s/GPU—significantly higher than standard bfloat16 or Transformer Engine FP4 baselines. The study demonstrates that FP4 performance depends on the interplay of scaling, operand packing, layout, and execution path, with downstream task rankings diverging from training‑loss rankings.
arXiv:2607. 02893v1 Announce Type: new Abstract: Low-bit quantization shrinks language models but treats precision as a single global hyper-parameter: every weight uses the same bit-width.