v0.2.19 - FLUX.1 Image Generation

Highlights

FLUX.1 Image Generation

Text-to-image generation with Black Forest Labs' FLUX.1 model:

• Full FLUX.1-schnell transformer (19 joint + 38 single blocks)
• Flow matching Euler scheduler
• GPU-native operations (transpose, batched matmul, RoPE)

Lazy Model Loading with Streaming

Memory-efficient model loading strategies:

• StreamingStrategy.EAGER - Load all at once (default)
• StreamingStrategy.PROGRESSIVE - Load during first forward
• StreamingStrategy.LAYER_BY_LAYER - Minimal memory usage

cuBLAS Dynamic Loader

• Runtime DLL loading without compile-time CUDA Toolkit
• Auto-detection of cuBLASLt versions (13/12/11)
• Graceful fallback to native kernels

C++ Kernel Profiler

• Built-in CUDA kernel profiling with minimal overhead
• Per-kernel timing statistics

HuggingFace T5 Encoder

• Sharded safetensors support
• Full T5 encoder for FLUX/SD3

DiT Architecture

• PixArt transformer with AdaLN-Zero
• Self/cross attention with GQA
• GEGLU FFN

New GPU Operations

• transpose_4d_0213, transpose_3d_012
• gpu_batched_matmul, gpu_softmax, gpu_apply_rope
• cross_attention, conv2d, group_norm

Known Issues

• FLUX.1 performance needs optimization (#187)

Full Changelog

v0.2.18...v0.2.19