PulseAugur
EN
LIVE 01:51:31

llama.cpp bug causes non-deterministic results for M-RoPE embedding batches

By PulseAugur Editorial · [2 sources] ·

A bug in the llama.cpp library causes incorrect results when processing embedding batches for M-RoPE models like Qwen3.5 and Qwen2.5-VL. The issue stems from a heap buffer overflow where the library reads past the allocated memory for position arrays. This can lead to non-deterministic logits and, in some cases, incorrect token predictions, though it is often subtle and hard to detect. The problem occurs when using the embedding mode for these models, while standard token processing remains unaffected. AI

IMPACT This bug in llama.cpp could lead to subtle inaccuracies in multimodal AI applications using M-RoPE models, potentially impacting reproducibility and output quality.

RANK_REASON The cluster describes a bug in a specific software tool (llama.cpp) that affects model inference, rather than a new model release or research breakthrough.

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →

llama.cpp bug causes non-deterministic results for M-RoPE embedding batches

How we ranked this

Signal score
0 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster describes a bug in a specific software tool (llama.cpp) that affects model inference, rather than a new model release or research breakthrough.
Source corroboration
2 independent sources
Multiple independent publishers reporting the same story raises confidence that it's real and newsworthy.
Topics
infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
59 days old
Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.

Full methodology in our editorial standards.

COVERAGE [2]

  1. dev.to — LLM tag TIER_1 English(EN) · The Homelab Postmortem ·

    llama.cpp reads past your pos array on embedding batches for M-RoPE models, and the logits change with the heap

    <p><strong>TL;DR</strong>: decode the same <code>llama_batch.embd</code> batch twice on an M-RoPE model — Qwen3.5, Qwen2.5-VL, Qwen3-VL, anything whose <code>llama_model_rope_type()</code> is <code>MROPE</code> or <code>IMROPE</code> — and the logits differ, on the CPU, one threa…

  2. dev.to — LLM tag TIER_1 English(EN) · The Homelab Postmortem ·

    llama.cpp reads past its own pos buffer for M-RoPE embeddings, and nothing reports it

    <p><strong>TL;DR</strong>: <code>include/llama.h</code> says the <code>pos</code> array "must have size of n_tokens", and that if you pass <code>NULL</code> "the token position will be tracked automatically". For a model that uses multiple positions per embedding — M-RoPE, which …

RELATED ENTITIES

RELATED TOPICS