@ZyphraAI: Zyphra Research releases PUFFER, a novel incremental fuzzy deduplication system for LLM-scale datasets. PUFFER runs ent…
Summary
Zyphra Research releases PUFFER, an incremental fuzzy deduplication system for LLM-scale datasets that runs entirely on CPU with 11-35x speedup over existing methods, open-sourced under Apache 2.0.
View Cached Full Text
Cached at: 09/02/26, 01:57 PM
Zyphra Research releases PUFFER, a novel incremental fuzzy deduplication system for LLM-scale datasets.
PUFFER runs entirely on CPU and achieves 11-35x speedup over existing methods.
We use PUFFER internally for our own training dataset and release it under Apache 2.0 license. https://t.co/I4XygVZkFi
Similar Articles
@Zai_org: https://x.com/Zai_org/status/2057216685040443743
This paper presents ZCube, a novel network architecture developed by Z.ai, Harnets.AI, and Tsinghua University to address topology-induced congestion in Prefill-Decode disaggregated LLM inference clusters. Production deployments on GLM-5.1 coding workloads achieved a 33% reduction in network CapEx, 15% throughput improvement, and 40.6% reduction in TTFT P99 latency.
Scalable Frequency- and Length-Aware Subdocument Deduplication for Large Language Model Pretraining
Proposes a scalable subdocument deduplication framework for LLM pretraining that separates duplicate detection from copy retention, using frequency- and length-aware policies. Experiments on FineWeb-Edu and a code web corpus show improved model performance.
OpenZL
OpenZL is a compression library that generates specialized compressors for specific data formats, achieving high compression ratios at high speeds suitable for datacenter workloads like AI processing.
@robertnishihara: Some intuition about PD disaggregation from the blog - PD doesn't speed up prefill and can actually hurt TTFT - PD's re…
This blog post from Anyscale explains the intuition behind Prefill-Decode (PD) disaggregation for LLM serving, showing how separating prefill and decode phases onto dedicated GPUs can achieve up to 2.7x better goodput and 67% cost savings when using Ray and vLLM on AMD MI325X, while also discussing when PD disaggregation does not help.
I reverse-engineered an NPU vendor's engine format (int8 weights stored as two nibble planes) to run GGUFs with no model conversion — now 1.5× faster than the vendor's own runtime
Reverse-engineered the Axera AX8850 NPU's int8 weight format to enable direct GGUF inference in llama.cpp, achieving up to 24.5 t/s decode and 716 t/s prefill on a Raspberry Pi 5, outperforming the vendor's runtime by 1.5×.