Skip to content

npu-pack: cache the 1.9 canonical .fconst on disk, not in the temp directory (it is RAM here: 3.6 GB per h4096 pack) #1801

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
bong-water-water-bong merged 1 commit into from
Nov 10, 2026
Merged

npu-pack: cache the 1.9 canonical .fconst on disk, not in the temp directory (it is RAM here: 3.6 GB per h4096 pack) #1801

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
19 changes: 18 additions & 1 deletion npu/pack_import.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -717,8 +717,19 @@ std::string materialise_token_fconst(const std::string& pack_dir, const Dims& di
// 1.9 pack failed with `filesystem error: temp_directory_path: No such file or directory [/tmp/x]` and no hint that
// the pack was fine. Fall back rather than fail: /tmp, then the current directory. The copy is scratch, and a note
// says where it went so nobody wonders what that 291 MB file is.
// Where the canonical copy lives. This is a CACHE (reused on size+mtime+rules version), not a leak, but its size is the
// pack's fconst -- 3.6 GB for an h4096 pack -- and on a box whose /tmp is a RAM disk that is RAM, per pack, held until
// reboot: three of them from one afternoon's imports were 11.4 GB of tmpfs (measured 2026-11-10). So the order is:
// an explicit override, then the on-disk cache directory, and only then the temp directory this used to prefer.
std::vector<fs::path> dirs;
if (const char* scratch = std::getenv("ONEBIT_FCONST_SCRATCH"); scratch && *scratch)
dirs.emplace_back(scratch);
if (const char* xdg = std::getenv("XDG_CACHE_HOME"); xdg && *xdg)
dirs.emplace_back(fs::path(xdg) / "1bit/fconst19");
else if (const char* home = std::getenv("HOME"); home && *home)
dirs.emplace_back(fs::path(home) / ".cache/1bit/fconst19");
try {
// Last resort only: on strixhalo /tmp is tmpfs, so this costs RAM equal to the pack's fconst.
dirs.push_back(fs::temp_directory_path());
} catch (const std::exception& e) {
std::fprintf(stderr,
Expand All @@ -730,7 +741,13 @@ std::string materialise_token_fconst(const std::string& pack_dir, const Dims& di
std::string work;
for (const auto& d : dirs) {
std::error_code ec;
if (!fs::is_directory(d, ec) || ec) continue;
// The cache directory is ours to create: without this the new first choice never wins, because it does not exist on a
// fresh box and the loop walked straight past it to /tmp (caught by testing the default, not the override).
if (!fs::is_directory(d, ec) || ec) {
std::error_code made;
fs::create_directories(d, made);
if (made || !fs::is_directory(d)) continue;
}
const fs::path cand = d / name;
// Writable in fact, not just in theory: the copy is written immediately after this.
std::ofstream probe(cand, std::ios::binary | std::ios::app);
Expand Down
Loading