github llama.cpp · 1h ago · inference

llama.cpp b10437 adds support for MiniMax-Text-01 and MiniMax-M1 models

from English

The llama.cpp project released version b10437, which introduces native support for the MiniMax-Text-01 and MiniMax-M1 causal language models. This update includes the necessary conversion scripts, chat templates, and model architecture implementations to run these specific models within the llama.cpp framework.

Added `MiniMaxText01ForCausalLM` and `MiniMaxM1ForCausalLM` support to the model layer.
Implemented token suppression for embeddings with zero values to prevent disruption of the sampling process.
Optimized performance by removing state transpose operations and using common functions for concise implementation.
Added a Jinja chat template specifically for MiniMax-M1.
Included build artifacts for macOS, Linux, Windows, Android, and openEuler across CPU, GPU, and various accelerators.

This release enables users to locally run MiniMax models using llama.cpp's existing inference infrastructure.

Importance 1/3 Trust 1/3 llama.cpp MiniMax Inference efficiency

Read original

github llama.cpp · 19d ago · 4 views

llama.cpp b10142 adds preliminary MiniMax-M3 vision support

The llama.cpp project released build b10142, which introduces preliminary vision support for the MiniMax-M3 model. This update implements a text-only port that reuses existing components from MiniMax-M2, including GQA with per-head QK-norm and partial rotary, DeepSeek-V3 style experts, and swigluoai activation.

github llama.cpp · 8h ago

llama.cpp b10435 fixes quadratic cost in jinja gather_string_parts

The llama.cpp project has released version b10435, which addresses a performance issue within the Jinja template engine. The primary change fixes a quadratic time complexity bug in the `gather_string_parts` function.

github llama.cpp · 11h ago · 1 view

llama.cpp b10434 passes reasoning_effort to templates

The llama.cpp project released build b10434, which adds support for passing the `reasoning_effort` parameter through to Jinja templates. This change ensures that OpenAI Chat Completions `reasoning_effort` values are stored and made available during generation.

lab NVIDIA Research · 12h ago · 6 views

Nvidia presents FusionRelight for real-time portrait relighting via hybrid domain knowledge fusion

Researchers from Nvidia have introduced FusionRelight, a method for portrait relighting that combines physically plausible illumination transfer with identity preservation and compact real-time inference. The approach utilizes Hybrid Domain Knowledge Fusion (HDKF), a training framework that distills physics, reflectance, and realism priors from synthetic, One-Light-at-a-Time (OLAT), and in-the-wild data into a student model.

github llama.cpp · 13h ago · 1 view

llama.cpp b10431 adds recurrent state rollback for SSM models

The llama.cpp project released version b10431, which introduces support for recurrent state (RS) rollback in the ggml_ssm_scan operation. This change enables RS rollback for Nemotron models on both CPU and CUDA backends.