[PR #211] [MERGED] fix: decline fused DeltaNet kernels below SM75 #223

New issue

Closed

opened 2026-10-26 22:21:45 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:21:45 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/211
Author: @itssutch
Created: 9/28/2026
Status: ✅ Merged
Merged: 9/28/2026
Merged by: @comfyanonymous

Base: main ← Head: fix/gated-delta-sm75-floor

📝 Commits (1)

f9e6df3 fix: decline fused DeltaNet kernels below SM75

📊 Changes

2 files changed (+32 additions, -0 deletions)

View changed files

📝 comfy_kitchen/gated_delta.py (+6 -0)
📝 tests/test_gated_delta_fused.py (+26 -0)

📄 Description

Summary

gated_delta.is_available() treated the CUDA extension and shared-memory optin as enough. Published wheels compile 75-virtual as the floor, so Tesla V100 (sm_70) has no cubin.
V100 still reports ~96 KB opt-in shared memory, which is enough for the fused decode budget. ComfyUI 0.37 therefore took the fused Qwen 3.5 TextGenerate path and deltanet_conv_step raised launch rejected instead of using the existing PyTorch fallback.
Gate on (7, 5), matching sage_attention._NATIVE_MINIMUM_CAPABILITY.

Fixes #210

Test plan

tests/test_gated_delta_fused.py::test_gated_delta_capability_dispatch — (7, 0) is False, (7, 5) and newer are True (HIP skipped)
Kitchen CUDA CI on SM75+ (existing fused tests unchanged)

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/comfy-kitchen/pull/211 **Author:** [@itssutch](https://github.com/itssutch) **Created:** 9/28/2026 **Status:** ✅ Merged **Merged:** 9/28/2026 **Merged by:** [@comfyanonymous](https://github.com/comfyanonymous) **Base:** `main` ← **Head:** `fix/gated-delta-sm75-floor` --- ### 📝 Commits (1) - [`f9e6df3`](https://github.com/Comfy-Org/comfy-kitchen/commit/f9e6df3ba470cf27e7c77d084e49be50dc099575) fix: decline fused DeltaNet kernels below SM75 ### 📊 Changes **2 files changed** (+32 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `comfy_kitchen/gated_delta.py` (+6 -0) 📝 `tests/test_gated_delta_fused.py` (+26 -0) </details> ### 📄 Description ## Summary - `gated_delta.is_available()` treated the CUDA extension and shared-memory optin as enough. Published wheels compile `75-virtual` as the floor, so Tesla V100 (`sm_70`) has no cubin. - V100 still reports ~96 KB opt-in shared memory, which is enough for the fused decode budget. ComfyUI 0.37 therefore took the fused Qwen 3.5 `TextGenerate` path and `deltanet_conv_step` raised `launch rejected` instead of using the existing PyTorch fallback. - Gate on `(7, 5)`, matching `sage_attention._NATIVE_MINIMUM_CAPABILITY`. Fixes #210 ## Test plan - [x] `tests/test_gated_delta_fused.py::test_gated_delta_capability_dispatch` — `(7, 0)` is False, `(7, 5)` and newer are True (HIP skipped) - [ ] Kitchen CUDA CI on SM75+ (existing fused tests unchanged) --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-26 22:21:45 +00:00