From 761908ffd7acac342fcb34633ceb982ddbe2042f Mon Sep 17 00:00:00 2001
From: Connor Baker <ConnorBaker01@gmail.com>
Date: Mon, 12 Oct 2026 22:09:14 -0700
Subject: [PATCH] cudaPackages.cccl: backport cuda::make_tma_descriptor fix

cuda::make_tma_descriptor reverses the box sizes into driver order
(innermost first), then checks the last element, the outermost
dimension, against the 16-byte requirement of tma_interleave_layout::none.
It rejects valid descriptors whose outermost box dimension isn't 16-byte
aligned and accepts invalid ones whose innermost isn't. Fixed in CCCL
3.5.0; CUDA 13.4 ships 3.3.4. The upstream commit applies cleanly only to
CUDA 13.4's CCCL, so the backport is limited to it.

Assisted-by: Claude Code (Claude Opus 5.5, claude-opus-5-5)
---
 pkgs/development/cuda-modules/packages/cccl.nix | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/pkgs/development/cuda-modules/packages/cccl.nix b/pkgs/development/cuda-modules/packages/cccl.nix
index 7b86ad2e0aec..292920bf0762 100644
--- a/pkgs/development/cuda-modules/packages/cccl.nix
+++ b/pkgs/development/cuda-modules/packages/cccl.nix
@@ -47,6 +47,18 @@ buildRedist {
         extraPrefix = "include/";
         hash = "sha256-eko1GSD2NPSLtbQ3diwgKMbuS6wU9lHDOajLJy5lBwM=";
       })
+    ]
+    ++ lib.optionals (cudaAtLeast "13.4" && cudaOlder "13.5") [
+      # cuda::make_tma_descriptor checks the outermost box dimension instead of the innermost against the 16-byte
+      # requirement, rejecting valid descriptors; fixed in CCCL 3.5.0.
+      # Backport: https://github.com/NVIDIA/cccl/pull/10546
+      (fetchpatch {
+        name = "fix-make_tma_descriptor";
+        url = "https://github.com/NVIDIA/cccl/commit/c95095c8d5faf1d8e978057a4f5cc0c27bccab70.patch";
+        relative = "libcudacxx/include";
+        extraPrefix = "include/";
+        hash = "sha256-QmbzcRMHNX2PuUihaXNrbB2kWwAOUArC4OZ6v9+Tf/g=";
+      })
     ];
 
   # NVIDIA, in their wisdom, expect CCCL to be a directory inside include.