Skip to content

native: the CUDA lane backend forwards every entry a family's backend gives #742

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
ashhart merged 1 commit into from
Oct 13, 2026
Merged

native: the CUDA lane backend forwards every entry a family's backend gives #742

Show file tree
Hide file tree
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
107 changes: 12 additions & 95 deletions zig/src/native/cuda.zig
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,7 @@ const lanes = @import("lanes");
const nemotron = @import("nemotron");
const Allocator = std.mem.Allocator;
const budget = @import("cuda_memory.zig");
const lane_backend = @import("cuda_bound.zig");
const nemotron_slide = @import("nemotron_slide.zig");

/// CUDA families provide metadata, open their lane backend and explain their refusals.
Expand Down Expand Up @@ -174,7 +175,7 @@ const Host = struct {
h.gpa.destroy(h);
}

fn bind(p: *anyopaque) *Host {
pub fn bind(p: *anyopaque) *Host {
const h: *Host = @ptrCast(@alignCast(p));
if (h.gpu) |g| if (bound != &g.ctx) {
g.ctx.makeCurrent() catch |e| std.log.err("cuCtxSetCurrent on the lane thread: {s}", .{@errorName(e)});
Expand Down Expand Up @@ -232,100 +233,7 @@ const Host = struct {

/// The family's backend, each call made with the context current on the calling thread.
fn backend(h: *Host) lanes.backend.Backend {
const v = h.inner.vtable;
h.vtable = .{
.prefill = struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!void {
const x = bind(p);
return x.inner.vtable.prefill(x.inner.ptr, s);
}
}.f,
.first = struct {
fn f(p: *anyopaque, s: *lanes.Stream, position: u64) anyerror!u64 {
const x = bind(p);
return x.inner.vtable.first(x.inner.ptr, s, position);
}
}.f,
.queue = struct {
fn f(p: *anyopaque, s: *lanes.Stream, feed: lanes.backend.Feed, position: u64) anyerror!u64 {
const x = bind(p);
return x.inner.vtable.queue(x.inner.ptr, s, feed, position);
}
}.f,
.read = struct {
fn f(p: *anyopaque, handle: u64) anyerror!u32 {
const x = bind(p);
return x.inner.vtable.read(x.inner.ptr, handle);
}
}.f,
.verify = struct {
fn f(p: *anyopaque, w: []const lanes.backend.Window, out: []lanes.backend.Verified) anyerror!void {
const x = bind(p);
return x.inner.vtable.verify(x.inner.ptr, w, out);
}
}.f,
.keep = struct {
fn f(p: *anyopaque, w: []const lanes.backend.Window, paths: []const []const u32) anyerror!void {
const x = bind(p);
return x.inner.vtable.keep(x.inner.ptr, w, paths);
}
}.f,
.draft = struct {
fn f(p: *anyopaque, r: []const lanes.backend.DraftRequest) anyerror!void {
const x = bind(p);
return x.inner.vtable.draft(x.inner.ptr, r);
}
}.f,
.unspeculate = if (v.unspeculate != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!void {
const x = bind(p);
return x.inner.vtable.unspeculate.?(x.inner.ptr, s);
}
}.f else null,
.probabilities = if (v.probabilities != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []f64) anyerror!bool {
const x = bind(p);
return x.inner.vtable.probabilities.?(x.inner.ptr, s, out);
}
}.f else null,
.tree = if (v.tree != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, gpa: Allocator) anyerror!?lanes.stream.Held {
const x = bind(p);
return x.inner.vtable.tree.?(x.inner.ptr, s, gpa);
}
}.f else null,
.alternatives = if (v.alternatives != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []lanes.backend.Alternative) anyerror!usize {
const x = bind(p);
return x.inner.vtable.alternatives.?(x.inner.ptr, s, out);
}
}.f else null,
.first_row = if (v.first_row != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!lanes.LogprobRow {
const x = bind(p);
return x.inner.vtable.first_row.?(x.inner.ptr, s);
}
}.f else null,
.first_masked = if (v.first_masked != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, position: u64, mask: []const u32) anyerror!u64 {
const x = bind(p);
return x.inner.vtable.first_masked.?(x.inner.ptr, s, position, mask);
}
}.f else null,
.held = if (v.held != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []u32) anyerror!void {
const x = bind(p);
return x.inner.vtable.held.?(x.inner.ptr, s, out);
}
}.f else null,
.prompt_rows = v.prompt_rows,
.release = struct {
fn f(p: *anyopaque, s: *lanes.Stream) void {
const x = bind(p);
x.inner.vtable.release(x.inner.ptr, s);
}
}.f,
};
h.vtable = lane_backend.vtable(Host, h.inner.vtable.*);
return .{ .ptr = h, .vtable = &h.vtable };
}
};
Expand Down Expand Up @@ -485,6 +393,15 @@ fn toGib(bytes: u64) f64 {

test {
_ = budget; // cuda_memory.zig's own tests, which no other test binary reaches
_ = lane_backend;
}

test "the lane core sees a family's prefill_step through the lane backend, so a served prompt can step" {
var target: lanes.fake.Fake = .{ .gpa = std.testing.allocator, .prefill_chunks = 3 };
defer target.deinit();
var h: Host = undefined;
h.inner = target.stepped();
try std.testing.expect(h.backend().vtable.prefill_step != null);
}

test "--learn is refused whatever the family, and --slide or a learned-disk size alone is not" {
Expand Down
139 changes: 139 additions & 0 deletions zig/src/native/cuda_bound.zig
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,139 @@
//! The lane backend a CUDA family hands the lane core, each entry called with the device context current.
const std = @import("std");
const lanes = @import("lanes");
const be = lanes.backend;

/// `v` (the family's entries, held as `inner` by `H`) behind `H.bind`: every entry the family gives, and only those.
pub fn vtable(comptime H: type, v: be.Backend.VTable) be.Backend.VTable {
return .{
.prefill = struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.prefill(x.inner.ptr, s);
}
}.f,
.prefill_step = if (v.prefill_step != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!bool {
const x = H.bind(p);
return x.inner.vtable.prefill_step.?(x.inner.ptr, s);
}
}.f else null,
.first = struct {
fn f(p: *anyopaque, s: *lanes.Stream, position: u64) anyerror!u64 {
const x = H.bind(p);
return x.inner.vtable.first(x.inner.ptr, s, position);
}
}.f,
.queue = struct {
fn f(p: *anyopaque, s: *lanes.Stream, feed: be.Feed, position: u64) anyerror!u64 {
const x = H.bind(p);
return x.inner.vtable.queue(x.inner.ptr, s, feed, position);
}
}.f,
.read = struct {
fn f(p: *anyopaque, handle: u64) anyerror!u32 {
const x = H.bind(p);
return x.inner.vtable.read(x.inner.ptr, handle);
}
}.f,
.verify = struct {
fn f(p: *anyopaque, w: []const be.Window, out: []be.Verified) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.verify(x.inner.ptr, w, out);
}
}.f,
.keep = struct {
fn f(p: *anyopaque, w: []const be.Window, paths: []const []const u32) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.keep(x.inner.ptr, w, paths);
}
}.f,
.draft = struct {
fn f(p: *anyopaque, r: []const be.DraftRequest) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.draft(x.inner.ptr, r);
}
}.f,
.unspeculate = if (v.unspeculate != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.unspeculate.?(x.inner.ptr, s);
}
}.f else null,
.probabilities = if (v.probabilities != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []f64) anyerror!bool {
const x = H.bind(p);
return x.inner.vtable.probabilities.?(x.inner.ptr, s, out);
}
}.f else null,
.tree = if (v.tree != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, gpa: std.mem.Allocator) anyerror!?lanes.stream.Held {
const x = H.bind(p);
return x.inner.vtable.tree.?(x.inner.ptr, s, gpa);
}
}.f else null,
.alternatives = if (v.alternatives != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []be.Alternative) anyerror!usize {
const x = H.bind(p);
return x.inner.vtable.alternatives.?(x.inner.ptr, s, out);
}
}.f else null,
.prepare_features = if (v.prepare_features != null) struct {
fn f(p: *anyopaque, taps: []const u32) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.prepare_features.?(x.inner.ptr, taps);
}
}.f else null,
.features = if (v.features != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, taps: []const u32, start: u64, count: u32) anyerror!be.Features {
const x = H.bind(p);
return x.inner.vtable.features.?(x.inner.ptr, s, taps, start, count);
}
}.f else null,
.first_row = if (v.first_row != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream) anyerror!lanes.LogprobRow {
const x = H.bind(p);
return x.inner.vtable.first_row.?(x.inner.ptr, s);
}
}.f else null,
.first_masked = if (v.first_masked != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, position: u64, mask: []const u32) anyerror!u64 {
const x = H.bind(p);
return x.inner.vtable.first_masked.?(x.inner.ptr, s, position, mask);
}
}.f else null,
.held = if (v.held != null) struct {
fn f(p: *anyopaque, s: *lanes.Stream, out: []u32) anyerror!void {
const x = H.bind(p);
return x.inner.vtable.held.?(x.inner.ptr, s, out);
}
}.f else null,
.prompt_rows = v.prompt_rows,
.release = struct {
fn f(p: *anyopaque, s: *lanes.Stream) void {
const x = H.bind(p);
x.inner.vtable.release(x.inner.ptr, s);
}
}.f,
};
}

test "every entry the family's backend gives is forwarded, and none it leaves out (a stepped prompt pass included)" {
const H = struct {
inner: be.Backend,
pub fn bind(p: *anyopaque) *@This() {
return @ptrCast(@alignCast(p));
}
};
var fake: lanes.fake.Fake = .{ .gpa = std.testing.allocator, .prefill_chunks = 3 };
defer fake.deinit();
for ([_]be.Backend{ fake.backend(), fake.stepped() }) |b| {
const w = vtable(H, b.vtable.*);
const info = @typeInfo(be.Backend.VTable).@"struct";
inline for (info.field_names, info.field_types) |name, T| {
if (@typeInfo(T) == .optional) try std.testing.expectEqual(@field(b.vtable, name) != null, @field(w, name) != null);
if (T == bool) try std.testing.expectEqual(@field(b.vtable, name), @field(w, name));
}
}
try std.testing.expect(vtable(H, fake.stepped().vtable.*).prefill_step != null);
}
Loading