Skip to content

Install a CUDA wheel that exists on a pre-12.4 driver #650

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
thcp merged 1 commit into from
Sep 21, 2026
Merged

Install a CUDA wheel that exists on a pre-12.4 driver #650

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
106 changes: 94 additions & 12 deletions desktop/src-tauri/src/main.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2241,8 +2241,14 @@ fn cuda_tag(cuda_version: &str) -> &'static str {
// not just Blackwell (#502). CUDA is backward compatible, so cu128 is
// the right floor for a 13.x driver.
[major, _] if *major >= 13 => "cu128",
[12, minor] if *minor >= 4 => "cu124",
[12, _] => "cu121",
// Every CUDA 12 driver, not just 12.4 and newer. A 12.0-12.3 driver
// used to be handed cu121, whose index stops at torch 2.5.1 -- so the
// 2.6.0 line this app installs was asked for at a tag that has never
// published it, pip answered "No matching distribution found", and an
// RTX 4060 dropped to CPU with a GPU sitting right there (#644).
// cu124 is the right answer for all of them: CUDA 12 is minor-version
// compatible, so a cu124 build runs on any 12.x driver.
[12, _] => "cu124",
[11, _] => "cu118",
_ => "cu124",
}
Expand Down Expand Up @@ -2281,9 +2287,17 @@ fn wheel_candidates(compute_cap: Option<&str>, cuda_version: &str) -> Vec<&'stat
.and_then(|cap| cap.split('.').next()?.parse::<u32>().ok())
.is_some_and(|major| major >= 10);
if blackwell {
vec!["cu128"]
} else {
vec![cuda_tag(cuda_version)]
return vec!["cu128"];
}
match cuda_tag(cuda_version) {
// cu118 as a second chance for a CUDA 12 driver. cu124 is the right
// first answer for all of them (see cuda_tag), but minor-version
// compatibility is the thing being relied on there, and when it does
// not hold the alternative used to be CPU. cu118 runs on every 12.x
// driver and publishes the same torch 2.6.0 line, so the fallthrough
// this list was built for (#502) finally has somewhere to go (#644).
"cu124" => vec!["cu124", "cu118"],
tag => vec![tag],
}
}

Expand Down Expand Up @@ -5896,7 +5910,7 @@ mod tests {
// torchaudio 2.9 removed the soundfile backend. Any tag this maps to
// 2.9+ ships a torchaudio whose save() needs torchcodec, which is not
// a StemDeck dependency, so demucs' ta.save() would fail at runtime.
for tag in ["cu128", "cu124", "cu121", "cu118"] {
for tag in ["cu128", "cu124", "cu118"] {
let v = super::torch_version_for_tag(tag);
let minor: u32 = v.split('.').nth(1).unwrap().parse().unwrap();
assert!(minor < 9, "{tag} maps to torch {v}, which is 2.9+");
Expand All @@ -5909,18 +5923,86 @@ mod tests {
// catch-all and got cu124 -- two major versions behind, on every card.
assert_eq!(super::cuda_tag("13.0"), "cu128");
assert_eq!(super::cuda_tag("14.2"), "cu128");
// Older drivers keep their existing mapping.
// A CUDA 12 driver gets cu124 whatever its minor version: 12.1 used
// to get cu121, whose index has no torch 2.6.0 to install (#644).
assert_eq!(super::cuda_tag("12.8"), "cu124");
assert_eq!(super::cuda_tag("12.1"), "cu121");
assert_eq!(super::cuda_tag("12.1"), "cu124");
// An 11.x driver keeps its existing mapping.
assert_eq!(super::cuda_tag("11.8"), "cu118");
}

#[test]
fn a_non_blackwell_card_still_follows_the_driver() {
assert_eq!(super::wheel_candidates(Some("8.9"), "12.4"), vec!["cu124"]);
assert_eq!(
super::wheel_candidates(Some("8.9"), "12.4"),
vec!["cu124", "cu118"]
);
assert_eq!(super::wheel_candidates(None, "11.8"), vec!["cu118"]);
}

/// Every wheel tag setup can offer publishes the torch line it installs.
///
/// The tag comes from the driver and the version from a table, and nothing
/// held the two together. A CUDA 12.0-12.3 driver was handed cu121, whose
/// index stops at torch 2.5.1, so setup asked pip for torch==2.6.0+cu121 --
/// never published, on any platform -- and an RTX 4060 dropped to CPU with
/// "No matching distribution found" (#644).
#[cfg(not(target_os = "macos"))]
#[test]
fn every_wheel_tag_offered_has_a_published_torch_wheel() {
// Checked against download.pytorch.org/whl/<tag>/ on 2026-09-21, for
// torch, torchaudio and torchvision, cp312, win_amd64 and
// linux_x86_64. A tag with no row here is one nobody confirmed
// publishes what setup is about to ask it for, so add the row before
// offering the tag.
const PUBLISHED: [(&str, &str); 3] =
[("cu128", "2.8.0"), ("cu124", "2.6.0"), ("cu118", "2.6.0")];

let drivers = [
"11.8", "12.0", "12.1", "12.3", "12.4", "12.8", "13.0", "14.2", "unknown",
];
let caps = [
None,
Some("7.5"),
Some("8.6"),
Some("8.9"),
Some("10.0"),
Some("12.0"),
Some("N/A"),
];
for driver in drivers {
for cap in caps {
for tag in super::wheel_candidates(cap, driver) {
let published = PUBLISHED
.iter()
.find(|(t, _)| *t == tag)
.map(|(_, version)| *version);
let Some(published) = published else {
panic!("{tag} (driver {driver}, cap {cap:?}) has no verified torch line");
};
assert_eq!(
super::torch_version_for_tag(tag),
published,
"{tag} installs a torch its index does not publish",
);
}
}
}
}

/// The reporter's machine: an RTX 4060 behind a pre-12.4 driver (#644).
#[cfg(not(target_os = "macos"))]
#[test]
fn a_pre_12_4_driver_gets_a_gpu_wheel_and_a_second_chance() {
for driver in ["12.0", "12.1", "12.2", "12.3"] {
assert_eq!(
super::wheel_candidates(Some("8.9"), driver),
vec!["cu124", "cu118"],
"driver {driver}",
);
}
}

#[test]
fn legacy_migration_preserves_user_settings_when_data_dir_already_exists() {
// setup() creates the destination before ensure_workspace() invokes
Expand Down Expand Up @@ -6642,10 +6724,10 @@ b6052160df96b31c9b1e33854a4dcda3d4b57641b880270f31736fb9f445d384 ffmpeg-n7.1-la
assert_eq!(super::wheel_tag(Some("10.0"), "12.4"), "cu128");
// Non-Blackwell cards fall back to the CUDA-version heuristic.
assert_eq!(super::wheel_tag(Some("8.9"), "12.4"), "cu124");
assert_eq!(super::wheel_tag(Some("8.6"), "12.1"), "cu121");
assert_eq!(super::wheel_tag(Some("8.6"), "12.1"), "cu124");
assert_eq!(super::wheel_tag(Some("7.5"), "11.8"), "cu118");
// Missing / unparseable compute capability also falls back.
assert_eq!(super::wheel_tag(None, "12.1"), "cu121");
assert_eq!(super::wheel_tag(None, "12.1"), "cu124");
assert_eq!(super::wheel_tag(Some("N/A"), "12.4"), "cu124");
}

Expand All @@ -6662,7 +6744,7 @@ b6052160df96b31c9b1e33854a4dcda3d4b57641b880270f31736fb9f445d384 ffmpeg-n7.1-la
assert_eq!(super::torch_version_for_tag("cu128"), "2.8.0");
assert_eq!(super::torchvision_version_for_tag("cu128"), "0.23.0");

for tag in ["cu124", "cu121", "cu118"] {
for tag in ["cu124", "cu118"] {
assert_eq!(super::torch_version_for_tag(tag), super::CPU_TORCH_VERSION);
assert_eq!(
super::torchvision_version_for_tag(tag),
Expand Down
Loading