Issues

• First scaled_mm_nvfp4 inside a CUDA graph capture fails: GetScalarZero's one-time cudaMemcpyToSymbol

  Comfy-Org/comfy-kitchen#305 · Filter by author PaulFidika opened on Nov 6, 2026

• eager backend convrot_w4a4 int4_linear computes int4 MMA in fp16 — overflows to Inf/NaN on fully-black outputs (RDNA2/ROCm and any host without the fused kernel)

  Comfy-Org/comfy-kitchen#299 · Filter by author Stanley-MD opened on Nov 2, 2026

• Windows/RDNA3 perf data point: --use-ck-attention is 1.30-1.37x vs default attention at 1024² (Qwen-Image-2.1, RX 7900 XTX)

  Comfy-Org/comfy-kitchen#292 · Filter by author AIwork4me opened on Nov 1, 2026

• int8_linear rounds to bf16 around the activation division: error on bf16 inputs, fp32 division fixes it at no cost

  Comfy-Org/comfy-kitchen#289 · Filter by author sorryhyun opened on Oct 31, 2026

• INT8 linears on sm_121 (GB10 / DGX Spark) run at ~37% of tensor-core peak via cutlass_int8_dequant; a fused-dequant Triton GEMM reaches ~88%, bit-identical

  Comfy-Org/comfy-kitchen#280 · Filter by author kanthai opened on Oct 26, 2026

• int8_linear silently falls back to a DP4A-emulated INT8 GEMM on Volta (sm_70), which is slower than the FP16 tensor cores

  Comfy-Org/comfy-kitchen#278 · Filter by author seningpd opened on Oct 25, 2026
  2

• flash_attention_decode_gqa rejects single-token decode (S == 1): output stride check fails after DLPack normalizes size-1 strides

  Comfy-Org/comfy-kitchen#274 · Filter by author ntsrigaud opened on Oct 23, 2026
  2

• Integrating EntroPack for flexible-bitrate weight compression

  Comfy-Org/comfy-kitchen#273 · Filter by author mi804 opened on Oct 23, 2026

• triton rms_rope kernel reads its input twice and launches 95k single-row programs

  Comfy-Org/comfy-kitchen#261 · Filter by author okolenmi opened on Oct 19, 2026

• SM75: MiniMax H3 VAE fp16_accumulation launches unsupported CUTLASS MMA and aborts

  Comfy-Org/comfy-kitchen#258 · Filter by author ZanePoe opened on Oct 19, 2026
  1

• Sol-Attn repeated call invalid resource handle on Blackwell with CUDA 13 source wheel

  Comfy-Org/comfy-kitchen#245 · Filter by author xmarre opened on Oct 12, 2026
  7

• INT8 GEMM not supported on this GPU for these dimensions (requires SM >= 7.5, and dimensions multiple of 4).

  Comfy-Org/comfy-kitchen#243 · Filter by author dranzerashi opened on Oct 12, 2026

• CUDA sol_attn token_aug overflow: atomic reservations select nondeterministic retained IDs despite histogram parity

  Comfy-Org/comfy-kitchen#241 · Filter by author xmarre opened on Oct 12, 2026
  1

• [Bug] comfy kitchen attention (INT8) returns pure noise on AMD gfx1100 once text conditioning exceeds ~150 tokens (Qwen-Image 2.1 int8_convrot)

  Comfy-Org/comfy-kitchen#230 · Filter by author snow930 opened on Oct 1, 2026
  1

• how to disable comfy-kitchen, making bug multi gpu for cuda 128 blackwell

  Comfy-Org/comfy-kitchen#228 · Filter by author EKI-INDRADI opened on Oct 1, 2026
  1

• [AMD RDNA3 Regression] CK Attention produces corrupted Qwen Image 2.1 output after recent update

  Comfy-Org/comfy-kitchen#226 · Filter by author dkspwndj opened on Sep 30, 2026
  1

• dequantize_int8_simple_dtype: cuda/hip drop leading dims when a one-element scale has more dims than q

  Comfy-Org/comfy-kitchen#204 · Filter by author arcusbuilds opened on Sep 26, 2026

• INT8 ConvRot stalls on model re-entry on Windows/NVIDIA — 100% GPU at ~80W, high VRAM, sometimes OOM in int8_linear

  Comfy-Org/comfy-kitchen#196 · Filter by author Epervios opened on Sep 24, 2026
  2

• Add CUDA dequantize_mxfp8

  Comfy-Org/comfy-kitchen#190 · Filter by author arcusbuilds opened on Sep 21, 2026

• [gfx1201] hip backend missing bare (non-_dtype) dequant functions — silently falls back to slow single-threaded eager backend during model load

  Comfy-Org/comfy-kitchen#188 · Filter by author CS6 opened on Sep 20, 2026
  1

• int8_attention: OverflowError past 299,520 key rows (int32 strides), reachable on long-video workloads

  Comfy-Org/comfy-kitchen#178 · Filter by author shootthesound opened on Sep 15, 2026
  2

• int8_attention: masked path costs ~1.8x unmasked, and a request for optional return_lse

  Comfy-Org/comfy-kitchen#177 · Filter by author shootthesound opened on Sep 14, 2026

• HIP backend: kernels taking __bf16 never launch on ROCm 7.1 / clang 21 (host stub and device symbol mangled differently)

  Comfy-Org/comfy-kitchen#170 · Filter by author dyc-dev opened on Sep 10, 2026
  1

• [Bug] Triton rms_rope kernel exceeds CUDA gridDim.z limit (65535) when seq_len >= 65536

  Comfy-Org/comfy-kitchen#169 · Filter by author trojanGoat opened on Sep 10, 2026

• hello, anyone interested in corrections for Apple Silicon? I had codex correct a new issue after updating comfy.

  Comfy-Org/comfy-kitchen#164 · Filter by author buggz2k opened on Sep 7, 2026
  1