You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Qwen-Image-Edit-2509 (cortiq imagine --image, the file)
runs every transformer block of a step as one resident chain over CUDA (gpu_cuda/qwen_image.rs, qwen_image.cu): each stream's LayerNorm with its shift and scale, the q/k/v products, per-head RMS norm of q and
k with the interleaved rotation, joint attention over [text | image] on h3_flash, the gated residuals and the
tanh-GELU MLPs, the weights kept on the card in their codec under CMF_GPU_VRAM_MB. The Qwen VAE's convolutions
and its single 512-wide attention head run on the card too (gpu_cuda::h3_vae_conv2d; heads wider than the
flash kernel take TF32 products around an f32 softmax). RTX A5000, release build, a 1024 x 1024 edit at 8 steps
with CFG 4: 105.8 s (0.8.79: 464 s), 115.1 s within an 8 GB ceiling (peak 7.4 GB); the card's 40 steps in
371.6 s. A seed gives the same picture twice. CMF_QWEN_IMAGE_CUDA=0 keeps the per-operation path.
Changed
Qwen-Image-Edit over Vulkan runs its blocks on the H3 chain's parts (gpu_wgpu/ltx.rs::qwen_chain): the q4tp
products read in place on the 16x16 matrix units into one joint [q | k | v] panel (text rows first), q/k norm and
rotation in place on it, vit_flash over all rows, the gated residuals and MLPs between. A transformer pass
takes 4.3 s at 1024 x 1024 (0.8.79: 21.5 s): the edit above in 148.1 s (0.8.79: 442.5 s), 30.5 dB from the
CUDA picture. CMF_QWEN_WGPU_H3=0 keeps the previous chain.
The Qwen-Image resident chain is on by default under CMF_GPU=cuda (it needed a discrete Vulkan adapter with
enough memory before).