生成式 AI 工具
ComfyUI 0.33.1 Adds Native MiniMax Music 3 and CUDA Graph Support, Fixes Non-Dynamic VRAM Mode
ComfyUI 0.33.1 brings MiniMax Music 3 model loading and inference nodes into the core and adds CUDA Graph support for backend use. The release also fixes Music 3 failing to run with fixed VRAM configurations and issues detecting QKV layouts in some MiniMax weights.
The highlight of ComfyUI 0.33.1 is not the addition of another remote API node, but the direct integration of MiniMax Music 3’s local model path and foundational CUDA Graph support into the core. Users can build music-generation workflows using the text encoders, VAE, and GGUF weights already published by the community. The core now also performs early detection of fused QKV weights and separate Q, K, and V weights, reducing the likelihood that different converted versions will fail during loading because of incompatible tensor layouts.
CUDA Graph works by recording a sequence of GPU kernel calls and replaying it with lower CPU dispatch overhead. For workloads such as diffusion or audio generation that repeatedly execute similar computation graphs, it may reduce idle gaps caused by numerous small kernel launches. However, this release’s changelog confirms only “core support” and does not provide comparisons of Music 3’s end-to-end latency, VRAM usage, or throughput. Whether workflows involving dynamic shapes, custom nodes, or changing control flow can be captured reliably must also be tested individually; merging the feature should not be interpreted as guaranteeing a fixed performance gain.
The release also fixes MiniMax Music 3 not working in non-dynamic-VRAM mode and an issue that caused dynamic VRAM to be disabled automatically under WSL. These fixes are particularly relevant to fixed VRAM strategies, Windows/WSL deployments, and memory-constrained local workstations. The same release also addresses nested latent issues, a `float64` device issue in the LTX diffusion decoder, and Gemma 4 E2B/E4B ignoring `thinking=false`. The community also published Music 3 GGUF assets soon after the release, indicating that an ecosystem of workflows is already taking shape. Before upgrading production environments, users should still pin version 0.33.1, preserve existing workflows, and separately validate model weight layouts, audio duration, peak VRAM usage, and CUDA Graph fallback behavior; an official comprehensive compatibility matrix is not yet available.