What's New

🧹 Streamlined Interface Removed non-functional options that caused crashes with Anima DiT models:

• Removed weight_dtype selector — Anima models are natively bfloat16. Forcing FP8 caused LayerNormKernelImpl not implemented for Float8_e4m3fn crashes because PyTorch does not support FP8 in RMSNorm operations used by DiT architectures. The loader now always uses the optimal native dtype.

• Removed compile_mode selector — Modes reduce-overhead and max-autotune use CUDA Graphs internally, which conflict with Anima's in-place tensor operations, causing CUDAGraphs tensor overwritten errors. The loader now always uses the stable default compilation mode.

🔬 SageAttention Fix
Fixed a critical bug where selecting specific quantization kernels (sageattn_qk_int8_pv_fp16_cuda / fp8) silently failed due to API changes in sageattention >= 1.0.6. The library migrated from C++ extensions to Triton kernels and removed those functions from the public API. The node now correctly uses the auto mode which delegates to ComfyUI's built-in SageAttention integration — this automatically selects the optimal quantization strategy for your GPU.

📦 New: Checkpoint Loader Added Anima Booster Checkpoint Loader — a new node for loading bundled Anima checkpoints that contain Model + CLIP + VAE in a single file. Outputs all three components while applying the same SageAttention and torch.compile optimizations as the standard Anima Booster Loader.

⚡ Clean Node Design Both loaders now feature a minimal, crash-proof interface:

• Model selector

• SageAttention toggle (auto / disabled)

• torch.compile toggle (on / off)

No unnecessary options that could lead to errors. Just plug in and generate.