Repository navigation

Issues

New issue

Dynamic VRAM streaming crashes all generations with HostBuffer.read_file_slice failed → CUDA OOM (regression after Aug 3 2026 update) (CORE-398)
#15255 · roumani76-arch opened on Aug 3, 2026
76

Hong
Comfy-Org/ComfyUI#17021 · opened on Oct 14, 2026
[WARNING] On windows we are currently forcing single GPU mode in ComfyUI due to a Nvidia related issue
Comfy-Org/ComfyUI#17017 · opened on Oct 14, 2026
5
Slow generation H3 Oct 12
Comfy-Org/ComfyUI#17015 · opened on Oct 14, 2026
SeedVR2 FP8 outputs a black image with --fast fp16_accumulation: FP16 overflows to infinity
Comfy-Org/ComfyUI#17014 · opened on Oct 14, 2026
1
TrimAudioDuration retains the entire source waveform allocation
Comfy-Org/ComfyUI#17012 · opened on Oct 14, 2026
[Intel XPU] Qwen3-VL DeepStack multi-image boolean indexing selects wrong token count (1059 vs 2061); CPU-index workaround
Comfy-Org/ComfyUI#16991 · opened on Oct 13, 2026
4
Alert Failed to initialize 3D Viewer
Comfy-Org/ComfyUI#16980 · opened on Oct 12, 2026
Loop iterations with cache_iterations off are kept in the output cache for the whole prompt, filling RAM until model pins are freed
Comfy-Org/ComfyUI#16978 · opened on Oct 12, 2026
DynamicPrompts don't apply on string inputs in a DynamicCombo option
Comfy-Org/ComfyUI#16977 · opened on Oct 12, 2026
Color Transfer Fails in Comfy Core but Works in KJNodes ColorMatch
Comfy-Org/ComfyUI#16973 · opened on Oct 12, 2026
Download fails with “EISDIR: illegal operation on a directory, link” when the models directory is on an exFAT drive (hard link used to finalize download, no fallback)
Comfy-Org/ComfyUI#16972 · opened on Oct 12, 2026
1
Detect Edges (Canny): whole batch in one kornia call - 134 MB VRAM per 720p frame, and slower per frame the larger the batch
Comfy-Org/ComfyUI#16971 · opened on Oct 12, 2026
Generation speed is slower when updated to the latest version
Comfy-Org/ComfyUI#16969 · opened on Oct 12, 2026
4
w
Comfy-Org/ComfyUI#16966 · opened on Oct 11, 2026
fsvdxz
Comfy-Org/ComfyUI#16965 · opened on Oct 11, 2026
Select CLIP Device: text encoder still runs on the old offloaded model, so generation on gpu:1 is ~7x slower
Comfy-Org/ComfyUI#16964 · opened on Oct 11, 2026
SAM 3D Body: BVH export writes a garbled rest skeleton (parent-local OFFSETs), breaking retargeting; plus fp16 hand-refinement dtype crash
Comfy-Org/ComfyUI#16963 · opened on Oct 11, 2026
Per-checkpoint default sampler settings (auto-fill KSampler and latent size on model selection)
Comfy-Org/ComfyUI#16961 · opened on Oct 11, 2026
About an **Error saving setting Comfy.CustomColorPalettes: TimeoutError: Fetch timeout** during startup
Comfy-Org/ComfyUI#16959 · opened on Oct 11, 2026
3
Duplicate in-flight prompt IDs can break heap insertion for equal priorities
Comfy-Org/ComfyUI#16952 · opened on Oct 11, 2026
1
RuntimeError "Boolean value of Tensor with more than one value is ambiguous" in select() (model_animate2.py) when using Qwen Image 2.1
Comfy-Org/ComfyUI#16947 · opened on Oct 10, 2026
[Windows][RTX 3060 12GB] MiniMax-H3 repeatedly freezes entire PC; no recent dump, NVFP4 encoder/offload log ends abruptly
Comfy-Org/ComfyUI#16946 · opened on Oct 10, 2026
5
[Feature Request] Persistent Model RAM Cache & Coordinated VRAM/RAM Swapping with Asynchronous Prefetching
Comfy-Org/ComfyUI#16939 · opened on Oct 10, 2026
1
Save video / Save Image - add option to disable prompt embedding
Comfy-Org/ComfyUI#16932 · opened on Oct 10, 2026
1
KSampler Doesnt Run
Comfy-Org/ComfyUI#16925 · opened on Oct 10, 2026
1

Navigation Menu

Repository navigation

Issues

Issues

Hong

[WARNING] On windows we are currently forcing single GPU mode in ComfyUI due to a Nvidia related issue

Slow generation H3 Oct 12

SeedVR2 FP8 outputs a black image with --fast fp16_accumulation: FP16 overflows to infinity

TrimAudioDuration retains the entire source waveform allocation

[Intel XPU] Qwen3-VL DeepStack multi-image boolean indexing selects wrong token count (1059 vs 2061); CPU-index workaround

Alert Failed to initialize 3D Viewer

Loop iterations with cache_iterations off are kept in the output cache for the whole prompt, filling RAM until model pins are freed

DynamicPrompts don't apply on string inputs in a DynamicCombo option

Color Transfer Fails in Comfy Core but Works in KJNodes ColorMatch

Download fails with “EISDIR: illegal operation on a directory, link” when the models directory is on an exFAT drive (hard link used to finalize download, no fallback)

Detect Edges (Canny): whole batch in one kornia call - 134 MB VRAM per 720p frame, and slower per frame the larger the batch

Generation speed is slower when updated to the latest version

w

fsvdxz

Select CLIP Device: text encoder still runs on the old offloaded model, so generation on gpu:1 is ~7x slower

SAM 3D Body: BVH export writes a garbled rest skeleton (parent-local OFFSETs), breaking retargeting; plus fp16 hand-refinement dtype crash

Per-checkpoint default sampler settings (auto-fill KSampler and latent size on model selection)

About an Error saving setting Comfy.CustomColorPalettes: TimeoutError: Fetch timeout during startup

Duplicate in-flight prompt IDs can break heap insertion for equal priorities

RuntimeError "Boolean value of Tensor with more than one value is ambiguous" in `select()` (`model_animate2.py`) when using Qwen Image 2.1

[Windows][RTX 3060 12GB] MiniMax-H3 repeatedly freezes entire PC; no recent dump, NVFP4 encoder/offload log ends abruptly

[Feature Request] Persistent Model RAM Cache & Coordinated VRAM/RAM Swapping with Asynchronous Prefetching

Save video / Save Image - add option to disable prompt embedding

KSampler Doesnt Run