Pipeline-Native Transformers: Co-Designing Model Architecture and CPU Inference for Bandwidth-Efficient Autoregressive Decode

• URL: http://arxiv.org/abs/2608.23841v1
• Date: Mon, 24 Aug 2026 21:28:05 GMT
• Title: Pipeline-Native Transformers: Co-Designing Model Architecture and CPU Inference for Bandwidth-Efficient Autoregressive Decode
• Abstract summary: Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic.<n>This report argues that the most effective response is to co-design the model architecture and the inference together.<n>It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Single-token autoregressive decode on CPUs is bound by memory bandwidth, not arithmetic: a modern CPU sustains roughly 1 TFLOP/s of compute but only about 50 GB/s from main memory, and each generated token must stream every active weight once. This report argues that the most effective response is to co-design the model architecture and the inference runtime together. It presents cflow, a CPU-first streaming engine, alongside a family of pipeline-native transformer architectures whose inter-layer dependency graphs are constructed to permit a vertical, stage-major execution schedule. cflow stores weights as L2-sized tiles in compute-consumption order, reads only the top-k experts of each mixture-of-experts layer, fuses projections, and executes a delay-aware schedule from per-model dependency parameters. Across five architectures trained on TinyStories, one (arch2_4_combined) achieves a 2.00x reduction in critical-path weight bandwidth (9.00 to 4.50 MB/token) within 0.24 perplexity of the best candidate, and the tile layout incurs 7.29x fewer L1-data read misses than a row-major baseline. On a 30.9-billion-parameter pipeline-native MoE, cflow decodes at 5.94 tokens/s (tok/s) on a 32-vCPU Ice Lake server, ahead of llama.cpp (4.75) and the vLLM CPU backend (1.65) on comparably sized dense models. Realizing the expert-delay window as asynchronous I/O overlap on a disk-resident expert tier yields a further net win of up to 1.68x, matching the overlap model within 1%. Measurement refutes one of the eight design claims and leaves a second inconclusive; both are reported in full, with the conditions under which they would hold.

Related papers

• Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PC Fleets [0.0]
  We show that a handful of AIPCs can serve models beyond the capability of any single one.<n>A four-node deployment of Lunar Lake AI PCs on Intel Tiber Cloud serves a single user at interactive speed.
  arXiv  Detail & Related papers  (2026-08-19T17:33:28Z)
• Hidden Decoding at Scale: Latent Computation Scaling for Large Language Models [75.80275843320511]
  We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.<n>We propose Hidden Decoding, a sequence-length scaling method applied during continued pretraining.
  arXiv  Detail & Related papers  (2026-07-09T07:37:59Z)
• Think Before You Grid-Search: Floor-First Triage for LLM Serving [0.0]
  LLM serving optimization typically benchmarks many configurations and reaches for heavy profilers when latency targets are missed.<n>We argue for the reverse discipline: estimation is the analytical layer of profiling.
  arXiv  Detail & Related papers  (2026-07-07T06:11:54Z)
• VisualClaw: A Real-Time, Personalized Agent for the Physical World [76.89211120690028]
  We present VisualClaw, a self-evolving multimodal agent built around two principles.<n>First, hybrid encoding reduces deployment cost by filtering less informative streaming frames with a cascaded gate.<n>Second, skill evolution lets the agent learn from failures, producing skill-bank updates that help future questions.
  arXiv  Detail & Related papers  (2026-06-15T06:58:22Z)
• Pushing the Limits of On-Device Streaming ASR: A Compact, High-Accuracy English Model for Low-Latency Inference [5.685908474026849]
  High-quality automatic speech recognition (ASR) on edge devices requires models that jointly optimize accuracy, latency, and memory footprint while operating entirely on CPU without GPU acceleration.<n>We conduct a systematic empirical study of state-of-the-art ASR architectures, encompassing encoder-decoder, transducer, and LLM-based paradigms, evaluated across batch, chunked, and streaming inference modes.<n>Our recommended configuration, the int4 k-quant variant, achieves 8.20% average streaming WER across eight standard benchmarks, running comfortably faster than real-time on CPU with 0.56 s algorithmic latency.
  arXiv  Detail & Related papers  (2026-04-16T00:04:32Z)
• Scaling DoRA: High-Rank Adaptation via Factored Norms and Fused Kernels [83.99688944263843]
  Weight-De Low-Rank Adaptation (DoRA) extends LoRA by decoupling weight magnitude from direction.<n>At d_in = 8192 and rank r = 384, a single module's norm requires about 512 MB of transient working memory in bf16.<n>A factored norm decomposes the squared norm into base, cross, and Gram terms computable through O(d_out r + r2) intermediates, eliminating the dense product.
  arXiv  Detail & Related papers  (2026-03-23T17:57:24Z)
• A Persistent-State Dataflow Accelerator for Memory-Bound Linear Attention Decode on FPGA [10.452946241750562]
  Gated DeltaNet (GDN) is a linear attention mechanism that replaces the growing KV cache with a fixed-size recurrent state.<n>We show that this bottleneck is architectural, not algorithmic, as all subquadratic sequence models exhibit arithmetic intensities below 1 FLOP/B at decode time.<n>We present an FPGA accelerator that eliminates this bottleneck by holding the full 2 MB recurrent state persistently in on-chip BRAM.
  arXiv  Detail & Related papers  (2026-03-06T06:03:38Z)
• dInfer: An Efficient Inference Framework for Diffusion Language Models [54.80918957287927]
  Diffusion-based large language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) LLMs.<n>We present dInfer, an efficient and framework for dLLM inference.
  arXiv  Detail & Related papers  (2025-10-09T16:19:42Z)
• 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float [52.079202872069835]
  Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs) have grown rapidly in size.<n>We introduce Dynamic-Length Float (DFloat11), a compression framework that reduces LLM and DM size by 30% while preserving outputs that are bit-for-bit identical to the original model.
  arXiv  Detail & Related papers  (2025-04-15T22:38:38Z)
• TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices [36.714057078457195]
  We present TPI-LLM, a compute- and memory-efficient tensor parallel inference system for 70B-scale models. TPI-LLM keeps sensitive raw data local in the users' devices and introduces a sliding window memory scheduler. We show that TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate.
  arXiv  Detail & Related papers  (2024-10-01T09:18:56Z)
• AI and Memory Wall [81.06494558184049]
  We show how memory bandwidth can become the dominant bottleneck for decoder models. We argue for a redesign in model architecture, training, and deployment strategies to overcome this memory limitation.
  arXiv  Detail & Related papers  (2024-03-21T04:31:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.