Skip to content
AI NEWS JOURNAL
AI, IN PERSPECTIVE

Search AI news

Search headlines, summaries and articles to connect the dots.

Newest first

Does the Ranking Hold? A Practitioner-Facing Audit of Hallucination Detectors on Small, Quantized, On-Device LLMs

arXiv:2611.17339v1 Announce Type: new Abstract: Hallucination detection for large language models is dominated by methods evaluated on 7B-70B research models with abundant sampling budgets and exact-match or ROUGE correctness labels. The small models practitioners actually ship on-device, Phi-3-mini, Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, and Gemma-2-2B-Instruct, typically at 4-bit precision, sit outside this evidence base. We present a reproducible one-GPU audit of four detectors spanning three paradigms: length-normalized token-margin calibration, a lexical n-gram baseline, and two sampling-based consistency methods (semantic entropy and SelfCheckGPT-NLI) on two small open instruction-tuned models (Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct) loaded at 4-bit NF4 quantization on a single Colab T4 GPU, evaluated on closed-book TriviaQA. Cross-model ranking transfer is poor: Spearman rho is -0.80 and Kendall tau-b is -0.667, meaning the detector ranking actively inverts between the two models, and the cheap lexical baseline is statistically indistinguishable from the sampling-based consistency detectors on both models under bootstrap 95 percent confidence intervals. As a second, methodological contribution, we show that the standard token-overlap-to-multiple-choice correctness labeler for TruthfulQA-Generation produces near-chance AUROC (0.46-0.59) for all four detectors on both models, a labeler failure, not a detector failure, since the same detectors reach 0.69-0.89 AUROC on TriviaQA under the identical pipeline. We release the harness, prompts, and per-item scores and frame the detector choice explicitly as a fairness-of-error-rates allocation problem for practitioners deploying on resource-constrained, on-device hardware.

b12377

The update refactors DFlash-based speculative decoding methods in llama.cpp by restructuring code organization, sharing components across models, splitting decoder graphs per head, and introducing dedicated classes for DFlash2 and LiliCorr drafts. It also adds support for new draft types, improves documentation, and removes unused code.

Inkling 2 Small on vLLM: Day-0 Support

vLLM supports Thinking Machines' Inkling 2 Small from day 0: attention residuals, a new decode kernel for relative-bias attention, all-reduce fused with the ops around it, and DSpark speculative decoding.

ForeSpark: Joint Draft and Verification Budget Scheduling for Parallel Speculative Decoding

arXiv:2611.13442v1 Announce Type: new Abstract: Parallel drafters accelerate LLM decoding by proposing a whole block of draft tokens in a single forward pass, making them attractive for LLM serving. Recent systems optimize goodput (i.e., accepted tokens per second) because rejected drafts steal compute from other requests. However, these systems control only how many tokens are verified, so every request still drafts a full block. At high serving load, we find that drafting consumes up to 40% of step time, of which a majority is discarded because it is never verified. Yet, draft-length control from autoregressive drafters does not carry over, because a parallel drafter commits to its length before its single pass. Moreover, draft and verification lengths are coupled across the batch, so they must be selected jointly for every request while keeping the decision computationally cheap at every step. In this paper, we propose work-conserving speculation (WCS), a design that jointly allocates draft and verification budgets per request in LLM serving, and present ForeSpark as its prototype. ForeSpark realizes WCS with (i) a context-aware pre-draft estimator that predicts acceptance for every candidate length from the target model's context features, (ii) a lightweight joint draft-verify scheduler that selects both lengths through a global sort, and (iii) a post-draft budget distribution that spends verification on the most promising tokens. Implemented on vLLM, ForeSpark improves throughput by up to 1.77x over the state-of-the-art DSpark and by 2.49x over autoregressive decoding at a matched per-request generation rate, without changing the output distribution. It also tolerates higher load, sustaining up to 1.51x the request rate of DSpark under latency service level objectives (SLOs).

Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU

TL;DR We serve Meta’s Muse Spark Open, an 816B-parameter mixture-of-experts model, on a single TPU v7 host of four chips, with SGLang running on the TorchTPU PyTorch backend. At the... The post Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU appeared first on PyTorch.

ReKV: Overlapping Quantized KV Transfer and Suffix Recomputation

arXiv:2611.08770v1 Announce Type: new Abstract: Prefill--decode disaggregation moves a prompt's key--value (KV) cache between workers before the first token can be generated. Reducing this handoff to either cache compression or recomputation misses a useful choice: which layers to transmit, at what precision, and which layers to execute again on the decode worker. ReKV transfers a quantized prefix KV cache and one BF16 boundary state, then reconstructs a contiguous suffix while the prefix is in flight. In a model-shaped synthetic benchmark on cross-node H100s, the Q8 configuration meeting separately evaluated quality criteria reduces 8K handoff latency by $2.36\times$, $2.84\times$, and $3.65\times$ over BF16 transfer for Qwen2.5-7B, Qwen2.5-3B, and Qwen3-8B. It is also faster than the best measured transfer-only or recompute-only action at those points. A separate real-weight BF16 recomputation experiment reduces prefill-to-token-return time by $1.16$--$1.63\times$. The preferred action changes with prompt length and link bandwidth.

b12180

The release adds the Kimi-K3.1-Pro text model to llama.cpp, replacing full-attention MLA layers in K3 with DSA (Dynamic Sparse Attention) layers featuring a lightning indexer. The indexer uses an additive mask over the cache and reuses sparse flash attention paths, while preserving existing KDA layers and MoE components. The model's architecture includes four indexer tensors per DSA layer, hybrid memory for caching, and maintains quantization integrity by excluding indexer parameters from quant overrides. Verified against HF checkpoint with identical PPL beyond context length, and supports multiple platforms including macOS, Linux, Windows, Android, and iOS.

b12184

The llama.cpp release b12184 includes a fix to add a missing link to pthread and privately link Threads::Threads, addressing potential runtime issues. The release provides precompiled binaries for multiple platforms including macOS, Linux, Windows, Android, and openEuler, supporting various backends such as CUDA, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon-specific hardware. Additional UI components are also available.

b12131

The update removes the u8path() function and replaces it with generic_string() in the server component, as part of a broader effort to streamline and standardize string handling across platforms.

b12132

The latest release of llama.cpp, version b12132, includes the removal of u8path() in favor of generic_string(), along with updated binaries for multiple platforms including macOS, Linux, Windows, Android, and openEuler, supporting various backends like CUDA, Vulkan, ROCm, OpenVINO, and SYCL.

b12137

A recent update to llama.cpp addresses a critical issue by rejecting non-object session updates in the server, preventing malformed realtime session data from causing uncaught exceptions and server termination. The release includes binaries for multiple platforms including macOS, Linux, Windows, Android, and iOS, with support for various backends like CUDA, Vulkan, ROCm, OpenVINO, and SYCL.

b12174

The llama.cpp release b12174 introduces the ability to generate RC files directly without requiring a win-app.rc.in file, alongside a comprehensive set of precompiled binaries for macOS, Linux, Windows, Android, and openEuler across various architectures and backends including CUDA, Vulkan, ROCm, OpenVINO, SYCL, and Snapdragon-specific hardware. The release also includes a standalone UI package.

LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time

arXiv:2611.07025v1 Announce Type: new Abstract: Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression.