nvexec(continues_on): basic transition

[romintomasetti]

On the machines I could test, very basic transitions with nvexec::strean_context never work (it segfaults).

I already flagged this in:

• Transitioning from a nvexec::stream_context to another seg faults #1563

@ericniebler I haven't looked for a fix, but this bug seems quite important to fix. Feel free to work with this branch, I won't touch anymore 😉

For the record, here is the stack trace reported by cuda-gdb: output.log.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

thanks for the test cases. the stream scheduler has a curious history. it was written several years ago by a coworker who is a CUDA ninja, but was new to std::execution. i am not a CUDA guru (although i'm getting better). the stream scheduler code is ... challenging for me. when i try to fix bugs there i lose whole days. and sometimes my "fixes" cause other issues, or performance regressions.

all this is to say, if anybody has an interest in digging into this, they would have my undying gratitude. i agree the current situation is unsatisfactory.

[maartenarnst]

The issue appears to be caused by a double call to schedule_from.

Once in the continues_on sender expression:

stdexec/include/stdexec/__detail/__continues_on.hpp

Line 421 in 27d64cd

schedule_from(static_cast<_Sender&&>(__sndr)));

And then once more in nvexec's implementation:

stdexec/include/nvexec/stream/continues_on.cuh

Line 178 in 27d64cd

: sndr_(schedule_from(static_cast<Sender&&>(sndr)))

This patch avoids the seg fault in my build:

From eb0868543dc7a742418661675270dcb4eb7af76b Mon Sep 17 00:00:00 2001
From: Maarten Arnst <maarten.arnst@uliege.be>
Date: Sun, 19 Apr 2026 09:42:25 +0200
Subject: [PATCH] Avoid double call to schedule_from

Signed-off-by: Maarten Arnst <maarten.arnst@uliege.be>
---
 include/nvexec/stream/continues_on.cuh | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/include/nvexec/stream/continues_on.cuh b/include/nvexec/stream/continues_on.cuh
index a7ee8f26..eb254959 100644
--- a/include/nvexec/stream/continues_on.cuh
+++ b/include/nvexec/stream/continues_on.cuh
@@ -172,15 +172,13 @@ namespace nv::execution::_strm
     template <class Sender>
     struct source_sender : stream_sender_base
     {
-      using schedule_from_sender_t = __result_of<schedule_from, Sender>;
-
       explicit source_sender(Sender sndr)
-        : sndr_(schedule_from(static_cast<Sender&&>(sndr)))
+        : sndr_(static_cast<Sender&&>(sndr))
       {}

       template <__decay_copyable Self, STDEXEC::receiver Receiver>
       STDEXEC_EXPLICIT_THIS_BEGIN(auto connect)(this Self&& self, Receiver rcvr)
-        -> connect_result_t<__copy_cvref_t<Self, schedule_from_sender_t>, Receiver>
+        -> connect_result_t<__copy_cvref_t<Self, Sender>, Receiver>
       {
         return STDEXEC::connect(static_cast<Self&&>(self).sndr_, static_cast<Receiver&&>(rcvr));
       }
@@ -201,7 +199,7 @@ namespace nv::execution::_strm
       }

      private:
-      __result_of<schedule_from, Sender> sndr_;
+      Sender sndr_;
     };

     template <class... Ty>
--
2.39.2 (Apple Git-143)

(It seems I can't push directly with a commit to the branch).

[romintomasetti]

I confirm that @maartenarnst 's fix works. I'll push it.

[romintomasetti]

The following still segfaults (switching from one context to another):

  TEST_CASE("continues on a different context instance", "[cuda][stream][adaptors][continues_on]")
  {
    nvexec::stream_context ctx_A, ctx_B;

    auto sndr = STDEXEC::just()
              | STDEXEC::continues_on(ctx_A.get_scheduler())
              | STDEXEC::continues_on(ctx_B.get_scheduler());

    STDEXEC::sync_wait(std::move(sndr));
  }

@ericniebler It's probably fine to merge this PR as a first step, i.e. it fixes the transition "to itself".

[ericniebler]

/ok to test ace3bba