PulseAugur
EN
LIVE 18:41:44

llama.cpp update fixes NaN output issue in model activations

A recent update to the llama.cpp project, tagged b12227, addresses an issue where specific model activations could lead to NaN (Not a Number) outputs. The problem arose from the `blk.N.ffn.act.scales` tensor in the mpt graph, which, with a particular seed, caused activations to overflow when processed by Metal stages. This overflow resulted in NaN logits and failed checks, which has now been corrected by ensuring the magnitude of `ffn.act` is kept away from zero. AI

IMPACT Ensures stability and correctness for models running on the llama.cpp framework.

RANK_REASON This is a code commit for an open-source project, not a new model release or significant research.

Read on llama.cpp — Releases →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

llama.cpp update fixes NaN output issue in model activations

How we ranked this

Signal score
1 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
This is a code commit for an open-source project, not a new model release or significant research.
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
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
1 days old
Coverage has settled into its steady-state source set.

Full methodology in our editorial standards.

COVERAGE [1]

  1. llama.cpp — Releases TIER_1 English(EN) · ggerganov ·

    b12227: tests : keep dummy ffn.act scales away from zero (#31940)

    <p>The dummy-model generator fills every tensor from a plain normal pool.<br /> <code>blk.N.ffn.act.scales</code> is a divisor in the mpt graph, so with seed<br /> 1224489258 it draws |v| ~ 1e-6, and <code>ffn_act = gelu(...) / scales</code> reaches<br /> 2.7e5. Metal stages both…