comfy-kitchen

This project publishes no release notes. Between v0.2.30 and v0.2.31 there were 9 commits, 9 of them substantive:

• Small comfy kitchen attention optimizations.
• HIP port of #103: int8 attention (#104)
• Some more tiny comfy kitchen optimizations.
• Revert bad optimization.
• Revert potentially problematic optimization.
• Exempt size-one extents from the int8 attention stride checks (#112)
• add minimal bf16 flash attention for seqlen capability (#111)
• Add function to check if flash attention decode is available.
• comfy kitchen v0.2.31

This project publishes no release notes. Between v0.2.28 and v0.2.30 there were 6 commits, 6 of them substantive:

• HIP ports of #99 (fused W4A8 requantize) and #95 (na3d / na2d) (#100)
• Implement optimized int8 attention. (#103)
• Comfy kitchen v0.2.29
• Remove stuff that isn't actually used for smaller kernel sizes.
• Ruff fix.
• Comfy kitchen v0.2.30

This project publishes no release notes. Between v0.2.27 and v0.2.28 there were 3 commits, 3 of them substantive:

• w4a8: stochastic rounding for LoRA merge + fused requant kernel (#99)
• Add na3d / na2d neighborhood attention (#95)
• comfy-kitchen v0.2.28

This project publishes no release notes. Between v0.2.26 and v0.2.27 there were 6 commits, 6 of them substantive:

• Fix HIP RoPE frequency length mismatch (#91)
• HIP port of #88: swiglu input_act and partial-rotary rot_dim (#93)
• Add optimized w4a8 with int8 codebook (#90)
• a4w8: use frozen codebook LUT + row-chunked quantize, improve LoRA application (#96)
• HIP port of #90: grouped W4A8 over the INT8 GEMM (#97)
• comfy-kitchen v0.2.27

This project publishes no release notes. Between v0.2.25 and v0.2.26 there were 4 commits, 4 of them substantive:

• Fix windows CI builds not caching properly.
• Add swiglu input_act and partial-rotary rot_dim for rms_rope (#88)
• Fix HIP backend device detection on ROCm lazy initialization (#89)
• comfy-kitchen v0.2.26

This project publishes no release notes. Between v0.2.24 and v0.2.25 there were 7 commits, 7 of them substantive:

• Add HIP backend for AMD RDNA2-RDNA4 to the existing wheels (#74)
• various cleanups to previous PR.
• AMD arch related build cleanup.
• Use AMD int8/int4 ops in the right intermediate dtype instead of bf16.
• Forgot this.
• Implement some missing hip ops.
• comfy-kitchen v0.2.25

This project publishes no release notes. Between v0.2.23 and v0.2.24 there were 2 commits, 2 of them substantive:

• Fix rope regression on the cuda path.
• comfy-kitchen v0.2.24

This project publishes no release notes. Between v0.2.22 and v0.2.23 there were 6 commits, 6 of them substantive:

• Add rms_rope to Triton backend. (#79)
• Support int8 convrot embedding lookup (#83)
• Add support for WoA (#82)
• Refactor RoPE kernels and add in-place variants (#85)
• rms_adaln and input_act kernels (#86)
• comfy-kitchen v0.2.23

This project publishes no release notes. Between v0.2.21 and v0.2.22 there were 3 commits, 3 of them substantive:

• Optimize int8 cutlass kernel for more shapes.
• Try to fix CI.
• comfy-kitchen v0.2.22