@jax-js/jax 0.1.2 → 0.1.4

package/dist/index.js

@@ -1,28 +1,36 @@
 import { __export } from "./chunk-Cl8Af3a2.js";
-import { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, FpHash, Kernel, PPrint, Reduction, ShapeTracker, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, deepEqual, defaultDevice, devices, dtypedArray, dtypedJsArray, generalBroadcast, getBackend, init, invertPermutation, isFloatDtype, isNumberPair, isPermutation, normalizeAxis, partitionList, prod, promoteTypes, range, recursiveFlatten, rep, runWithCache, setDebug, toposort, unravelAlu, unzip2, zip, zipn } from "./backend-BqymqzuU.js";
+import { AluExp, AluGroup, AluOp, AluVar, DEBUG, DType, FpHash, Kernel, PPrint, Reduction, Routine, Routines, ShapeTracker, accessorAluExp, accessorGlobal, assertNonNull, byteWidth, checkAxis, deepEqual, defaultDevice, devices, dtypedArray, dtypedJsArray, generalBroadcast, getBackend, init, invertPermutation, isFloatDtype, isNumberPair, isPermutation, normalizeAxis, partitionList, prod, promoteTypes, range, recursiveFlatten, rep, runWithCache, setDebug, toposort, unravelAlu, unzip2, zip, zipn } from "./backend-tngXtWe4.js";
 
 //#region src/frontend/convolution.ts
 /**
 * Check that the shapes and parameters passed to convolution are valid.
+* Expected shapes of the lhs and rhs of the convolution are:
+*
+* - `lhsShape = [*vmapDims, batchSize, inChannels, spatialDims...]`
+* - `rhsShape = [*vmapDims, outChannels, inChannels, kernelSize...]`
 *
 * If the check succeeds, returns the output shape.
 */
-function checkConvShape(lhsShape, rhsShape, { strides, padding, lhsDilation, rhsDilation }) {
+function checkConvShape(lhsShape, rhsShape, { vmapDims, strides, padding, lhsDilation, rhsDilation }) {
 	if (lhsShape.length !== rhsShape.length) throw new Error(`conv() requires inputs with the same number of dimensions, got ${lhsShape.length} and ${rhsShape.length}`);
-	const n = lhsShape.length - 2;
+	const n = lhsShape.length - 2 - vmapDims;
 	if (n < 0) throw new Error("conv() requires at least 2D inputs");
 	if (strides.length !== n) throw new Error("conv() strides != spatial dims");
 	if (padding.length !== n) throw new Error("conv() padding != spatial dims");
 	if (lhsDilation.length !== n) throw new Error("conv() lhsDilation != spatial dimensions");
 	if (rhsDilation.length !== n) throw new Error("conv() rhsDilation != spatial dimensions");
-	if (lhsShape[1] !== rhsShape[1]) throw new Error(`conv() input channels: ${lhsShape[1]} != ${rhsShape[1]}`);
-	const outShape = [lhsShape[0], rhsShape[0]];
+	if (lhsShape[vmapDims + 1] !== rhsShape[vmapDims + 1]) throw new Error(`conv() input channels: ${lhsShape[1]} != ${rhsShape[1]}`);
+	const outShape = [
+		...generalBroadcast(lhsShape.slice(0, vmapDims), rhsShape.slice(0, vmapDims)),
+		lhsShape[vmapDims],
+		rhsShape[vmapDims]
+	];
 	for (let i = 0; i < n; i++) {
 		if (strides[i] <= 0 || !Number.isInteger(strides[i])) throw new Error(`conv() strides[${i}] must be a positive integer`);
 		if (padding[i].length !== 2 || !padding[i].every(Number.isInteger)) throw new Error(`conv() padding[${i}] must be a 2-tuple of integers`);
 		if (lhsDilation[i] <= 0 || !Number.isInteger(lhsDilation[i])) throw new Error(`conv() lhsDilation[${i}] must be a positive integer`);
 		if (rhsDilation[i] <= 0 || !Number.isInteger(rhsDilation[i])) throw new Error(`conv() rhsDilation[${i}] must be a positive integer`);
-		const [x, k] = [lhsShape[i + 2], rhsShape[i + 2]];
+		const [x, k] = [lhsShape[i + vmapDims + 2], rhsShape[i + vmapDims + 2]];
 		if (k <= 0) throw new Error("conv() kernel size must be positive");
 		const [pl, pr] = padding[i];
 		if (pl < -x || pr < -x || pl + pr < -x) throw new Error(`conv() padding[${i}]=(${pl},${pr}) is too negative for input size ${x}`);
@@ -147,27 +155,13 @@ function poolTranspose(st, inShape, ks, strides = 1, dilation = 1) {
 function applyDilation(st, dilation) {
 	if (dilation.every((s) => s === 1)) return st;
 	const s_ = dilation;
-	const [a, b, ...k_] = st.shape;
-	st = st.reshape([
-		a,
-		b,
-		...k_.flatMap((k) => [k, 1])
-	]);
-	st = st.pad([
-		[0, 0],
-		[0, 0],
-		...s_.flatMap((s) => [[0, 0], [0, s - 1]])
-	]);
-	st = st.reshape([
-		a,
-		b,
-		...k_.map((k, i) => k * s_[i])
-	]);
-	st = st.shrink([
-		[0, a],
-		[0, b],
-		...k_.map((k, i) => [0, (k - 1) * s_[i] + 1])
-	]);
+	const n = s_.length;
+	const prefix = st.shape.slice(0, -n);
+	const k_ = st.shape.slice(-n);
+	st = st.reshape([...prefix, ...k_.flatMap((k) => [k, 1])]);
+	st = st.pad([...prefix.map(() => [0, 0]), ...s_.flatMap((s) => [[0, 0], [0, s - 1]])]);
+	st = st.reshape([...prefix, ...k_.map((k, i) => k * s_[i])]);
+	st = st.shrink([...prefix.map((p) => [0, p]), ...k_.map((k, i) => [0, (k - 1) * s_[i] + 1])]);
 	return st;
 }
 /**
@@ -177,25 +171,26 @@ function applyDilation(st, dilation) {
 * beforehand using `checkConvShape()`.
 */
 function prepareConv(stX, stY, params) {
-	const n = stX.shape.length - 2;
+	const v = params.vmapDims;
+	const n = stX.shape.length - 2 - v;
+	const vmapShape = stX.shape.slice(0, v);
 	stX = applyDilation(stX, params.lhsDilation);
-	const ks = stY.shape.slice(2);
-	stX = stX.padOrShrink([
-		[0, 0],
-		[0, 0],
-		...params.padding
-	]);
+	const ks = stY.shape.slice(v + 2);
+	stX = stX.padOrShrink([...rep(v + 2, [0, 0]), ...params.padding]);
 	stX = pool(stX, ks, params.strides, params.rhsDilation);
-	stX = stX.moveaxis(1, n + 1).reshape([
-		stX.shape[0],
+	stX = stX.moveaxis(v + 1, v + n + 1).reshape([
+		...vmapShape,
+		stX.shape[v],
 		1,
-		...stX.shape.slice(2, n + 2),
-		stX.shape[1] * prod(ks)
+		...stX.shape.slice(v + 2, v + n + 2),
+		stX.shape[v + 1] * prod(ks)
 	]);
 	stY = stY.reshape([
-		stY.shape[0],
+		...vmapShape,
+		1,
+		stY.shape[v],
 		...rep(n, 1),
-		stY.shape[1] * prod(ks)
+		stY.shape[v + 1] * prod(ks)
 	]);
 	return [stX, stY];
 }
@@ -336,6 +331,8 @@ let Primitive = /* @__PURE__ */ function(Primitive$1) {
 	Primitive$1["Mul"] = "mul";
 	Primitive$1["Idiv"] = "idiv";
 	Primitive$1["Mod"] = "mod";
+	Primitive$1["Min"] = "min";
+	Primitive$1["Max"] = "max";
 	Primitive$1["Neg"] = "neg";
 	Primitive$1["Reciprocal"] = "reciprocal";
 	Primitive$1["Floor"] = "floor";
@@ -343,7 +340,6 @@ let Primitive = /* @__PURE__ */ function(Primitive$1) {
 	Primitive$1["StopGradient"] = "stop_gradient";
 	Primitive$1["Cast"] = "cast";
 	Primitive$1["Bitcast"] = "bitcast";
-	Primitive$1["RandomBits"] = "random_bits";
 	Primitive$1["Sin"] = "sin";
 	Primitive$1["Cos"] = "cos";
 	Primitive$1["Asin"] = "asin";
@@ -353,8 +349,6 @@ let Primitive = /* @__PURE__ */ function(Primitive$1) {
 	Primitive$1["Erf"] = "erf";
 	Primitive$1["Erfc"] = "erfc";
 	Primitive$1["Sqrt"] = "sqrt";
-	Primitive$1["Min"] = "min";
-	Primitive$1["Max"] = "max";
 	Primitive$1["Reduce"] = "reduce";
 	Primitive$1["Dot"] = "dot";
 	Primitive$1["Conv"] = "conv";
@@ -362,14 +356,19 @@ let Primitive = /* @__PURE__ */ function(Primitive$1) {
 	Primitive$1["PoolTranspose"] = "pool_transpose";
 	Primitive$1["Compare"] = "compare";
 	Primitive$1["Where"] = "where";
+	Primitive$1["RandomBits"] = "random_bits";
+	Primitive$1["Gather"] = "gather";
 	Primitive$1["Transpose"] = "transpose";
 	Primitive$1["Broadcast"] = "broadcast";
 	Primitive$1["Reshape"] = "reshape";
 	Primitive$1["Flip"] = "flip";
 	Primitive$1["Shrink"] = "shrink";
 	Primitive$1["Pad"] = "pad";
-	Primitive$1["Gather"] = "gather";
-	Primitive$1["JitCall"] = "jit_call";
+	Primitive$1["Sort"] = "sort";
+	Primitive$1["Argsort"] = "argsort";
+	Primitive$1["TriangularSolve"] = "triangular_solve";
+	Primitive$1["Cholesky"] = "cholesky";
+	Primitive$1["Jit"] = "jit";
 	return Primitive$1;
 }({});
 let CompareOp = /* @__PURE__ */ function(CompareOp$1) {
@@ -391,6 +390,12 @@ function idiv(x, y) {
 function mod(x, y) {
 	return bind1(Primitive.Mod, [x, y]);
 }
+function min$1(x, y) {
+	return bind1(Primitive.Min, [x, y]);
+}
+function max$1(x, y) {
+	return bind1(Primitive.Max, [x, y]);
+}
 function neg(x) {
 	return bind1(Primitive.Neg, [x]);
 }
@@ -412,12 +417,6 @@ function cast(x, dtype) {
 function bitcast(x, dtype) {
 	return bind1(Primitive.Bitcast, [x], { dtype });
 }
-function randomBits(k0, k1, shape$1, mode = "xor") {
-	return bind1(Primitive.RandomBits, [k0, k1], {
-		shape: shape$1,
-		mode
-	});
-}
 function sin$1(x) {
 	return bind1(Primitive.Sin, [x]);
 }
@@ -445,12 +444,6 @@ function erfc$1(x) {
 function sqrt$1(x) {
 	return bind1(Primitive.Sqrt, [x]);
 }
-function min$1(x, y) {
-	return bind1(Primitive.Min, [x, y]);
-}
-function max$1(x, y) {
-	return bind1(Primitive.Max, [x, y]);
-}
 function reduce(x, op, axis = null, opts) {
 	if (!AluGroup.Reduce.has(op)) throw new TypeError(`Invalid reduce operation: ${op}`);
 	axis = normalizeAxis(axis, ndim$1(x));
@@ -467,9 +460,11 @@ function dot$2(x, y) {
 }
 function conv$1(x, y, params = {}) {
 	if (x.ndim !== y.ndim) throw new Error(`conv() requires inputs with the same number of dimensions, got ${x.ndim} and ${y.ndim}`);
-	const n = x.ndim - 2;
+	const vmapDims = params.vmapDims ?? 0;
+	const n = x.ndim - 2 - vmapDims;
 	if (n < 0) throw new Error("conv() requires at least 2D inputs");
 	return bind1(Primitive.Conv, [x, y], {
+		vmapDims,
 		strides: params.strides ?? rep(n, 1),
 		padding: params.padding ?? rep(n, [0, 0]),
 		lhsDilation: params.lhsDilation ?? rep(n, 1),
@@ -504,6 +499,23 @@ function where$1(cond, x, y) {
 		y
 	]);
 }
+function randomBits(k0, k1, shape$1, mode = "xor") {
+	return bind1(Primitive.RandomBits, [k0, k1], {
+		shape: shape$1,
+		mode
+	});
+}
+function gather(x, indices, axis, outDim) {
+	if (indices.length === 0) throw new Error("gather() requires at least one index");
+	if (!Array.isArray(axis) || axis.length !== indices.length) throw new Error(`Invalid gather() axis: expected ${indices.length} axes, got ${JSON.stringify(axis)}`);
+	axis = axis.map((a) => checkAxis(a, ndim$1(x)));
+	if (new Set(axis).size !== axis.length) throw new Error(`Invalid gather() axis: duplicate axes ${JSON.stringify(axis)}`);
+	outDim = checkAxis(outDim, ndim$1(x) - axis.length + 1);
+	return bind1(Primitive.Gather, [x, ...indices], {
+		axis,
+		outDim
+	});
+}
 function transpose$1(x, perm) {
 	perm = perm ? perm.map((a) => checkAxis(a, ndim$1(x))) : range(ndim$1(x)).reverse();
 	if (!isPermutation(perm, ndim$1(x))) throw new Error(`Invalid transpose permutation for ${ndim$1(x)} axes: ${JSON.stringify(perm)}`);
@@ -553,16 +565,27 @@ function pad$1(x, width) {
 	} else if (width.length !== nd) throw new Error(`Invalid pad(): expected ${nd} axes, got ${width.length}`);
 	return bind1(Primitive.Pad, [x], { width });
 }
-function gather(x, indices, axis, outDim) {
-	if (indices.length === 0) throw new Error("gather() requires at least one index");
-	if (!Array.isArray(axis) || axis.length !== indices.length) throw new Error(`Invalid gather() axis: expected ${indices.length} axes, got ${JSON.stringify(axis)}`);
-	axis = axis.map((a) => checkAxis(a, ndim$1(x)));
-	if (new Set(axis).size !== axis.length) throw new Error(`Invalid gather() axis: duplicate axes ${JSON.stringify(axis)}`);
-	outDim = checkAxis(outDim, ndim$1(x) - axis.length + 1);
-	return bind1(Primitive.Gather, [x, ...indices], {
-		axis,
-		outDim
-	});
+function triangularSolve$1(a, b, { lower = false, unitDiagonal = false } = {}) {
+	if (lower) {
+		a = flip$1(a, [-2, -1]);
+		b = flip$1(b, [-1]);
+	}
+	let x = bind1(Primitive.TriangularSolve, [a, b], { unitDiagonal });
+	if (lower) x = flip$1(x, [-1]);
+	return x;
+}
+function cholesky$2(x) {
+	return bind1(Primitive.Cholesky, [x]);
+}
+function sort$1(x) {
+	const nd = ndim$1(x);
+	if (nd === 0) throw new Error("sort: requires at least 1D input");
+	return bind1(Primitive.Sort, [x]);
+}
+function argsort$1(x) {
+	const nd = ndim$1(x);
+	if (nd === 0) throw new Error("argsort: requires at least 1D input");
+	return bind(Primitive.Argsort, [x]);
 }
 function bind1(prim, args, params = {}) {
 	const [results] = bind(prim, args, params);
@@ -693,8 +716,10 @@ var Tracer = class Tracer {
 		axis = normalizeAxis(axis, this.ndim);
 		const n = axis.reduce((acc, a) => acc * this.shape[a], 1);
 		if (n === 0) throw new Error("mean: cannot compute mean over zero-length axis");
-		const result = reduce(this, AluOp.Add, axis, opts);
-		return result.mul(1 / n);
+		const originalDtype = this.dtype;
+		const castDtype = promoteTypes(originalDtype, DType.Float32);
+		const result = reduce(this.astype(castDtype), AluOp.Add, axis, opts);
+		return result.mul(1 / n).astype(originalDtype);
 	}
 	/** Permute the dimensions of an array. Defaults to reversing the axis order. */
 	transpose(perm) {
@@ -723,7 +748,7 @@ var Tracer = class Tracer {
 		if (isFloatDtype(this.dtype)) return this.mul(reciprocal$1(other));
 		return idiv(this, other);
 	}
-	/** Return specified diagonals. See `numpy.diagonal` for full docs. */
+	/** Return specified diagonals. See `jax.numpy.diagonal` for full docs. */
 	diagonal(offset = 0, axis1 = 0, axis2 = 1) {
 		if (!Number.isInteger(offset)) throw new TypeError(`offset must be an integer, got ${offset}`);
 		if (offset < 0) return this.diagonal(-offset, axis2, axis1);
@@ -776,6 +801,34 @@ var Tracer = class Tracer {
 		this.dispose();
 	}
 	/**
+	* Return a sorted copy of an array in ascending order.
+	*
+	* See `jax.numpy.sort` for full docs.
+	*/
+	sort(axis = -1) {
+		axis = checkAxis(axis, this.ndim);
+		if (this.shape[axis] <= 1) return this;
+		if (axis === this.ndim - 1) return sort$1(this);
+		const perm = range(this.ndim);
+		perm.splice(axis, 1);
+		perm.push(axis);
+		return sort$1(this.transpose(perm)).transpose(invertPermutation(perm));
+	}
+	/**
+	* Return the indices that would sort an array. This may not be a stable
+	* sorting algorithm; it need not preserve order of indices in ties.
+	*
+	* See `jax.numpy.argsort` for full docs.
+	*/
+	argsort(axis = -1) {
+		axis = checkAxis(axis, this.ndim);
+		if (axis === this.ndim - 1) return argsort$1(this)[1];
+		const perm = range(this.ndim);
+		perm.splice(axis, 1);
+		perm.push(axis);
+		return argsort$1(this.transpose(perm))[1].transpose(invertPermutation(perm));
+	}
+	/**
 	* Slice an array along one or more axes.
 	*
 	* This is the equivalent of slicing in Python, e.g. `x[1:3, 2, :, None]`. To
@@ -892,6 +945,9 @@ var ShapedArray = class ShapedArray {
 	get ndim() {
 		return this.shape.length;
 	}
+	get size() {
+		return prod(this.shape);
+	}
 	toString() {
 		return `${this.dtype}[${this.shape.join(",")}]`;
 	}
@@ -1170,7 +1226,7 @@ var Jaxpr = class Jaxpr {
 			} else if (eqn.primitive === Primitive.Idiv) {
 				const [a, b] = inputs;
 				const c = eqn.outBinders[0];
-				if (atomIsLit(b, 1)) context.set(c, a);
+				if (atomIsLit(b, 1) && !isFloatDtype(a.aval.dtype)) context.set(c, a);
 				else newEqns.push(eqn);
 			} else if ((eqn.primitive === Primitive.Broadcast || eqn.primitive === Primitive.Reshape) && deepEqual(eqn.params.shape, eqn.inputs[0].aval.shape) || eqn.primitive === Primitive.Transpose && eqn.params.perm.every((p, i) => p === i) || eqn.primitive === Primitive.Flip && eqn.params.axis.length === 0 || eqn.primitive === Primitive.Shrink && eqn.params.slice.every(([s, e$2], i) => s === 0 && e$2 === eqn.inputs[0].aval.shape[i]) || eqn.primitive === Primitive.Pad && eqn.params.width.every(([w0, w1]) => w0 === 0 && w1 === 0)) context.set(eqn.outBinders[0], eqn.inputs[0]);
 			else newEqns.push(eqn);
@@ -1187,13 +1243,13 @@ var Jaxpr = class Jaxpr {
 		}
 		return new Jaxpr(this.inBinders, liveEqns.reverse(), outs);
 	}
-	/** Flattens nested JitCall in a Jaxpr. Useful for handling jit-of-jit. */
+	/** Flattens nested Jit in a Jaxpr. Useful for handling jit-of-jit. */
 	flatten() {
-		if (!this.eqns.some((eqn) => eqn.primitive === Primitive.JitCall)) return this;
+		if (!this.eqns.some((eqn) => eqn.primitive === Primitive.Jit)) return this;
 		const newEqns = [];
 		const varMap = /* @__PURE__ */ new Map();
 		const varMapF = (x) => x instanceof Var ? varMap.get(x) ?? x : x;
-		for (const eqn of this.eqns) if (eqn.primitive === Primitive.JitCall) {
+		for (const eqn of this.eqns) if (eqn.primitive === Primitive.Jit) {
 			const jaxpr = eqn.params.jaxpr.flatten();
 			const translation = /* @__PURE__ */ new Map();
 			const translationF = (x) => x instanceof Var ? translation.get(x) : x;
@@ -1294,19 +1350,48 @@ function evalJaxpr(jaxpr, args) {
 function jaxprAsFun(jaxpr) {
 	return (...args) => evalJaxpr(jaxpr, args);
 }
+/** Jaxpr with a collection of associated, traced constants. */
+var ClosedJaxpr = class ClosedJaxpr {
+	constructor(jaxpr, consts) {
+		this.jaxpr = jaxpr;
+		this.consts = consts;
+	}
+	/** String representation of this Jaxpr. */
+	toString() {
+		return this.jaxpr.toString();
+	}
+	/** Apply a function to the underlying Jaxpr. */
+	mapJaxpr(f) {
+		return new ClosedJaxpr(f(this.jaxpr), this.consts);
+	}
+	/** Dispose of the constants in this Jaxpr. */
+	dispose() {
+		for (const c of this.consts) c.dispose();
+	}
+};
 /** Tracer that records its operations to dynamically construct a Jaxpr. */
 var JaxprTracer = class extends Tracer {
+	#rc;
 	constructor(trace$1, aval) {
 		super(trace$1);
 		this.aval = aval;
+		this.#rc = 1;
 	}
 	toString() {
 		return `JaxprTracer(${this.aval.toString()})`;
 	}
 	get ref() {
+		if (this.#rc <= 0) throw new UseAfterFreeError(this);
+		this.#rc++;
 		return this;
 	}
-	dispose() {}
+	dispose() {
+		if (this.#rc <= 0) throw new UseAfterFreeError(this);
+		this.#rc--;
+	}
+	trackLiftedConstant() {
+		this.#rc++;
+	}
 };
 /** Analogous to the 'DynamicJaxprTrace' class in JAX. */
 var JaxprTrace = class extends Trace {
@@ -1319,17 +1404,24 @@ var JaxprTrace = class extends Trace {
 	}
 	/** Register a constant / literal in this Jaxpr. */
 	getOrMakeConstTracer(val) {
+		if (!(val instanceof Tracer)) val = pureArray(val);
 		let tracer = this.builder.constTracers.get(val);
 		if (tracer === void 0) {
 			tracer = this.builder.newTracer(this, ShapedArray.fromAval(getAval(val)));
-			this.builder.addConst(tracer, val instanceof Tracer ? val.ref : array(val));
+			this.builder.addConst(tracer, val);
+		} else {
+			val.dispose();
+			tracer.trackLiftedConstant();
 		}
 		return tracer;
 	}
 	pure = this.getOrMakeConstTracer;
 	lift = this.getOrMakeConstTracer;
 	processPrimitive(primitive, tracers, params) {
-		const avalsIn = tracers.map((t) => t.aval);
+		const avalsIn = tracers.map((t) => {
+			t.dispose();
+			return t.aval;
+		});
 		const avalsOut = abstractEvalRules[primitive](avalsIn, params);
 		const outTracers = avalsOut.map((aval) => this.builder.newTracer(this, aval));
 		this.builder.addEqn(new JaxprEqn(primitive, tracers.map((t) => this.builder.getVar(t)), params, outTracers.map((t) => this.builder.addVar(t))));
@@ -1372,20 +1464,17 @@ var JaxprBuilder = class {
 		return v;
 	}
 	build(inTracers, outTracers) {
-		let [constVars, consts] = unzip2(this.constVals.entries());
+		const [constVars, consts] = unzip2(this.constVals.entries());
 		const t2v = this.getVar.bind(this);
 		const inBinders = [...constVars, ...inTracers.map(t2v)];
 		const outVars = outTracers.map(t2v);
-		let jaxpr = new Jaxpr(inBinders, this.eqns, outVars);
+		const jaxpr = new Jaxpr(inBinders, this.eqns, outVars);
 		typecheckJaxpr(jaxpr);
-		[jaxpr, consts] = _inlineLiterals(jaxpr, consts);
-		return {
-			jaxpr,
-			consts
-		};
+		const cjaxpr = new ClosedJaxpr(jaxpr, consts);
+		return _inlineLiterals(cjaxpr);
 	}
 };
-function _inlineLiterals(jaxpr, consts) {
+function _inlineLiterals({ jaxpr, consts }) {
 	const literals = /* @__PURE__ */ new Map();
 	const constBinders = [];
 	const newConsts = [];
@@ -1400,7 +1489,7 @@ function _inlineLiterals(jaxpr, consts) {
 	const newOuts = jaxpr.outs.map((x) => literals.get(x) ?? x);
 	const newJaxpr = new Jaxpr([...constBinders, ...jaxpr.inBinders.slice(consts.length)], newEqns, newOuts);
 	typecheckJaxpr(newJaxpr);
-	return [newJaxpr, newConsts];
+	return new ClosedJaxpr(newJaxpr, newConsts);
 }
 function binopAbstractEval([x, y]) {
 	if (!(x instanceof ShapedArray) || !(y instanceof ShapedArray)) throw new TypeError("binopAbstractEval expects ShapedArray inputs");
@@ -1419,6 +1508,8 @@ const abstractEvalRules = {
 	[Primitive.Mul]: binopAbstractEval,
 	[Primitive.Idiv]: binopAbstractEval,
 	[Primitive.Mod]: binopAbstractEval,
+	[Primitive.Min]: binopAbstractEval,
+	[Primitive.Max]: binopAbstractEval,
 	[Primitive.Neg]: vectorizedUnopAbstractEval,
 	[Primitive.Reciprocal]: vectorizedUnopAbstractEval,
 	[Primitive.Floor]: vectorizedUnopAbstractEval,
@@ -1432,12 +1523,6 @@ const abstractEvalRules = {
 		if (byteWidth(x.dtype) !== byteWidth(dtype)) throw new TypeError(`Bitcast from ${x.dtype} to ${dtype} with different byte width`);
 		return [new ShapedArray(x.shape, dtype, false)];
 	},
-	[Primitive.RandomBits]([k0, k1], { shape: shape$1 }) {
-		if (k0.dtype !== DType.Uint32 || k1.dtype !== DType.Uint32) throw new TypeError(`RandomBits requires uint32 keys, got ${k0.dtype} and ${k1.dtype}`);
-		const keyShape = generalBroadcast(k0.shape, k1.shape);
-		if (!deepEqual(generalBroadcast(keyShape, shape$1), shape$1)) throw new TypeError(`Keys of shapes ${k0.shape} and ${k1.shape} cannot be broadcast to shape ${shape$1}`);
-		return [new ShapedArray(shape$1, DType.Uint32, false)];
-	},
 	[Primitive.Sin]: vectorizedUnopAbstractEval,
 	[Primitive.Cos]: vectorizedUnopAbstractEval,
 	[Primitive.Asin]: vectorizedUnopAbstractEval,
@@ -1447,8 +1532,6 @@ const abstractEvalRules = {
 	[Primitive.Erf]: vectorizedUnopAbstractEval,
 	[Primitive.Erfc]: vectorizedUnopAbstractEval,
 	[Primitive.Sqrt]: vectorizedUnopAbstractEval,
-	[Primitive.Min]: binopAbstractEval,
-	[Primitive.Max]: binopAbstractEval,
 	[Primitive.Reduce]([x], { axis }) {
 		const axisSet = new Set(axis);
 		const newShape = x.shape.filter((_, i) => !axisSet.has(i));
@@ -1481,6 +1564,25 @@ const abstractEvalRules = {
 		const shape$1 = generalBroadcast(cond.shape, xy.shape);
 		return [new ShapedArray(shape$1, xy.dtype, xy.weakType)];
 	},
+	[Primitive.RandomBits]([k0, k1], { shape: shape$1 }) {
+		if (k0.dtype !== DType.Uint32 || k1.dtype !== DType.Uint32) throw new TypeError(`RandomBits requires uint32 keys, got ${k0.dtype} and ${k1.dtype}`);
+		const keyShape = generalBroadcast(k0.shape, k1.shape);
+		if (!deepEqual(generalBroadcast(keyShape, shape$1), shape$1)) throw new TypeError(`Keys of shapes ${k0.shape} and ${k1.shape} cannot be broadcast to shape ${shape$1}`);
+		return [new ShapedArray(shape$1, DType.Uint32, false)];
+	},
+	[Primitive.Gather]([x, ...indices], { axis, outDim }) {
+		for (const a of indices) if (a.dtype !== DType.Int32 && a.dtype !== DType.Uint32) throw new TypeError(`Gather indices must be Int32 or Uint32, got ${a.dtype}`);
+		if (axis.length !== indices.length) throw new TypeError(`Gather: ${axis} axes but ${indices.length} indices`);
+		if (indices.length === 0) throw new TypeError("Gather must have 1+ indices with same shape");
+		if (axis.some((a) => a < 0 || a >= x.shape.length)) throw new TypeError("Gather axis out of bounds");
+		if (outDim < 0 || outDim > x.shape.length - axis.length) throw new TypeError("Gather outDim out of bounds");
+		const axisSet = new Set(axis);
+		if (axisSet.size !== axis.length) throw new TypeError("Gather axes are not unique");
+		const gatherShape = indices.reduce((shape$1, a) => generalBroadcast(shape$1, a.shape), []);
+		const newShape = x.shape.filter((_, i) => !axisSet.has(i));
+		newShape.splice(outDim, 0, ...gatherShape);
+		return [new ShapedArray(newShape, x.dtype, x.weakType)];
+	},
 	[Primitive.Transpose]([x], { perm }) {
 		return [new ShapedArray(perm.map((i) => x.shape[i]), x.dtype, x.weakType)];
 	},
@@ -1501,23 +1603,31 @@ const abstractEvalRules = {
 		const newShape = x.shape.map((dim, i) => dim + width[i][0] + width[i][1]);
 		return [new ShapedArray(newShape, x.dtype, x.weakType)];
 	},
-	[Primitive.Gather]([x, ...indices], { axis, outDim }) {
-		for (const a of indices) if (a.dtype !== DType.Int32 && a.dtype !== DType.Uint32) throw new TypeError(`Gather indices must be Int32 or Uint32, got ${a.dtype}`);
-		if (axis.length !== indices.length) throw new TypeError(`Gather: ${axis} axes but ${indices.length} indices`);
-		if (indices.length === 0) throw new TypeError("Gather must have 1+ indices with same shape");
-		if (axis.some((a) => a < 0 || a >= x.shape.length)) throw new TypeError("Gather axis out of bounds");
-		if (outDim < 0 || outDim > x.shape.length - axis.length) throw new TypeError("Gather outDim out of bounds");
-		const axisSet = new Set(axis);
-		if (axisSet.size !== axis.length) throw new TypeError("Gather axes are not unique");
-		const gatherShape = indices.reduce((shape$1, a) => generalBroadcast(shape$1, a.shape), []);
-		const newShape = x.shape.filter((_, i) => !axisSet.has(i));
-		newShape.splice(outDim, 0, ...gatherShape);
-		return [new ShapedArray(newShape, x.dtype, x.weakType)];
+	[Primitive.Sort]([x]) {
+		if (x.ndim === 0) throw new TypeError("sort: requires at least 1D input");
+		return [ShapedArray.fromAval(x)];
+	},
+	[Primitive.Argsort]([x]) {
+		if (x.ndim === 0) throw new TypeError("argsort: requires at least 1D input");
+		return [ShapedArray.fromAval(x), new ShapedArray(x.shape, DType.Int32, false)];
+	},
+	[Primitive.TriangularSolve]([a, b]) {
+		if (a.ndim < 2) throw new TypeError(`triangular_solve: a must be at least 2D, got ${a}`);
+		if (b.ndim < 2) throw new TypeError(`triangular_solve: b must be at least 2D, got ${b}`);
+		const [m, n] = a.shape.slice(-2);
+		const [_batch, q] = b.shape.slice(-2);
+		if (!deepEqual(a.shape.slice(0, -2), b.shape.slice(0, -2)) || a.dtype !== b.dtype || m !== n || n !== q) throw new TypeError(`triangular_solve: mismatch ${a} vs ${b}`);
+		return [new ShapedArray(b.shape, b.dtype, a.weakType && b.weakType)];
+	},
+	[Primitive.Cholesky]([a]) {
+		if (a.ndim < 2) throw new TypeError(`cholesky: requires at least 2D input, got ${a}`);
+		if (a.shape[a.ndim - 2] !== a.shape[a.ndim - 1]) throw new TypeError(`cholesky: must be square, got ${a}`);
+		return [ShapedArray.fromAval(a)];
 	},
-	[Primitive.JitCall](args, { jaxpr }) {
+	[Primitive.Jit](args, { jaxpr }) {
 		const { inTypes, outTypes } = typecheckJaxpr(jaxpr);
-		if (args.length !== inTypes.length) throw new TypeError(`jit_call expected ${inTypes.length} arguments, got ${args.length}`);
-		for (let i = 0; i < inTypes.length; i++) if (!args[i].equals(inTypes[i])) throw new TypeError(`jit_call argument ${i} has type ${args[i]}, expected ${inTypes[i]}`);
+		if (args.length !== inTypes.length) throw new TypeError(`jit expected ${inTypes.length} arguments, got ${args.length}`);
+		for (let i = 0; i < inTypes.length; i++) if (!args[i].equals(inTypes[i])) throw new TypeError(`jit argument ${i} has type ${args[i]}, expected ${inTypes[i]}`);
 		return outTypes;
 	}
 };
@@ -1553,11 +1663,10 @@ function makeJaxpr$1(f, opts) {
 			const tracersIn = avalsIn.map((aval) => trace$1.newArg(typeof aval === "object" ? aval : pureArray(aval)));
 			const outs = fFlat(...tracersIn);
 			const tracersOut = outs.map((out) => fullRaise(trace$1, out));
-			const { jaxpr, consts } = builder.build(tracersIn, tracersOut);
+			const jaxpr = builder.build(tracersIn, tracersOut);
 			if (outTree.value === void 0) throw new Error("outTree was not set in makeJaxpr");
 			return {
-				jaxpr: jaxpr.simplify(),
-				consts,
+				jaxpr: jaxpr.mapJaxpr((j) => j.simplify()),
 				treedef: outTree.value
 			};
 		} catch (_) {
@@ -1576,22 +1685,28 @@ function jit$1(f, opts) {
 		const avalsInFlat = argsFlat.map((x) => ShapedArray.fromAval(getAval(x)));
 		const avalsIn = unflatten(inTree, avalsInFlat);
 		const jaxprArgs = joinIdx(args.length, staticArgs, avalsIn, staticArgnums);
-		const { jaxpr, consts, treedef: outTree } = runWithCache(cache, jaxprArgs, () => makeJaxpr$1(f, opts)(...jaxprArgs));
-		const outs = bind(Primitive.JitCall, [...consts.map((c) => c.ref), ...argsFlat], {
+		const { jaxpr, treedef: outTree } = runWithCache(cache, jaxprArgs, () => makeJaxpr$1(f, opts)(...jaxprArgs));
+		const outs = bind(Primitive.Jit, [...jaxpr.consts.map((c) => c.ref), ...argsFlat], {
 			name: f.name || "closure",
-			jaxpr,
-			numConsts: consts.length
+			jaxpr: jaxpr.jaxpr,
+			numConsts: jaxpr.consts.length
 		});
 		return unflatten(outTree, outs);
 	});
 	result.dispose = () => {
-		for (const { consts } of cache.values()) for (const c of consts) c.dispose();
+		for (const { jaxpr } of cache.values()) jaxpr.dispose();
 	};
 	return result;
 }
 
 //#endregion
 //#region src/frontend/jit.ts
+const routinePrimitives = new Map([
+	[Primitive.Sort, Routines.Sort],
+	[Primitive.Argsort, Routines.Argsort],
+	[Primitive.TriangularSolve, Routines.TriangularSolve],
+	[Primitive.Cholesky, Routines.Cholesky]
+]);
 /** Result of compiling a Jaxpr. Can be evaluated on a series of inputs. */
 var JitProgram = class {
 	constructor(backend, steps, inputs, outputs) {
@@ -1606,9 +1721,14 @@ var JitProgram = class {
 				case "execute": {
 					const inputsNice = step.inputs.map((id, i) => `${i}: %${id}`).join(", ");
 					const outputsNice = step.outputs.map((id) => `%${id}`).join(", ");
-					return PPrint.pp(`execute (${inputsNice}) -> ${outputsNice}, kernel`).concat(step.kernel.pprint().indent(2));
+					const executeText = `execute (${inputsNice}) -> ${outputsNice}`;
+					if (step.source instanceof Kernel) return PPrint.pp(`${executeText}, kernel`).concat(step.source.pprint().indent(2));
+					else if (step.source instanceof Routine) return PPrint.pp(`${executeText}, routine ${step.source.name}`);
+					else {
+						step.source;
+						return PPrint.pp(executeText);
+					}
 				}
-				case "const": return PPrint.pp(`%${step.output} = const <Slot ${step.slot}>`);
 				case "malloc": return PPrint.pp(`%${step.output} = malloc <${step.size} bytes>`);
 				case "incref": return PPrint.pp(`incref ${step.input}`);
 				case "free": return PPrint.pp(`free ${step.input}`);
@@ -1631,12 +1751,9 @@ var JitProgram = class {
 				const inputs$1 = step.inputs.map((id) => scope.get(id));
 				const outputs = step.outputs.map((id) => scope.get(id));
 				if (inputs$1.some((s) => s === void 0) || outputs.some((s) => s === void 0)) throw new Error(`internal: JitProgram scope undefined`);
-				pending.push(new PendingExecute(this.backend, step.kernel, inputs$1, outputs));
+				pending.push(new PendingExecute(this.backend, step.source, inputs$1, outputs));
 				break;
 			}
-			case "const":
-				scope.set(step.output, step.slot);
-				break;
 			case "malloc": {
 				const slot = this.backend.malloc(step.size);
 				scope.set(step.output, slot);
@@ -1670,34 +1787,37 @@ var JitProgramBuilder = class {
 		this.#nextId = nargs;
 		this.steps = [];
 	}
-	pushConst(slot) {
-		const id = this.#nextId++;
-		this.steps.push({
-			type: "const",
-			slot,
-			output: id
-		});
-		return id;
-	}
 	pushLit(lit) {
-		const kernel = new Kernel(0, prod(lit.aval.shape), AluExp.const(lit.dtype, lit.value));
+		const kernel = new Kernel(0, lit.aval.size, AluExp.const(lit.dtype, lit.value));
 		return this.pushKernel(kernel, []);
 	}
-	pushKernel(kernel, inputs) {
+	pushBuffer(size$1) {
 		const id = this.#nextId++;
 		this.steps.push({
 			type: "malloc",
-			size: kernel.bytes,
+			size: size$1,
 			output: id
 		});
+		return id;
+	}
+	pushKernel(kernel, inputs) {
+		const id = this.pushBuffer(kernel.bytes);
 		this.steps.push({
 			type: "execute",
-			kernel,
+			source: kernel,
 			inputs,
 			outputs: [id]
 		});
 		return id;
 	}
+	pushRoutine(routine, inputs, outputs) {
+		this.steps.push({
+			type: "execute",
+			source: routine,
+			inputs,
+			outputs
+		});
+	}
 	pushIncref(id) {
 		this.steps.push({
 			type: "incref",
@@ -1723,28 +1843,18 @@ var JitProgramBuilder = class {
 	}
 };
 const jitCompileCache = /* @__PURE__ */ new Map();
-function jitCompile(backend, jaxpr, consts) {
-	if (jaxpr.inBinders.length < consts.length) throw new TypeError(`Jaxpr has ${jaxpr.inBinders.length} inputs, but ${consts.length} consts were provided`);
-	for (let i = 0; i < consts.length; i++) if (consts[i].device !== backend.type) throw new TypeError(`Const ${i} has device ${consts[i].device}, but expected ${backend.type}`);
-	const cacheKey = backend.type + FpHash.hash(jaxpr, ...consts.map((c) => c.id));
+function jitCompile(backend, jaxpr) {
+	const cacheKey = backend.type + "," + FpHash.hash(jaxpr);
 	const cached = jitCompileCache.get(cacheKey);
 	if (cached) return cached;
 	if (DEBUG >= 1) console.info("=========== JIT Compile ===========\n" + jaxpr.toString());
 	jaxpr = jaxpr.flatten().simplify();
-	const nargs = jaxpr.inBinders.length - consts.length;
+	const nargs = jaxpr.inBinders.length;
 	const builder = new JitProgramBuilder(backend, nargs);
 	const blackNodes = splitGraphDataflow(backend, jaxpr);
 	const ctx = /* @__PURE__ */ new Map();
-	for (let i = 0; i < consts.length; i++) {
-		const v = jaxpr.inBinders[i];
-		const slot = consts[i]._realizeSource();
-		ctx.set(v, {
-			type: "imm",
-			arg: builder.pushConst(slot)
-		});
-	}
 	for (let i = 0; i < nargs; i++) {
-		const v = jaxpr.inBinders[consts.length + i];
+		const v = jaxpr.inBinders[i];
 		ctx.set(v, {
 			type: "imm",
 			arg: i
@@ -1752,51 +1862,101 @@ function jitCompile(backend, jaxpr, consts) {
 	}
 	for (let i = 0; i < jaxpr.eqns.length; i++) {
 		const eqn = jaxpr.eqns[i];
+		if (routinePrimitives.has(eqn.primitive)) {
+			const routine = new Routine(routinePrimitives.get(eqn.primitive), {
+				inputShapes: eqn.inputs.map((x) => x.aval.shape),
+				inputDtypes: eqn.inputs.map((x) => x.aval.dtype),
+				outputShapes: eqn.outBinders.map((x) => x.aval.shape),
+				outputDtypes: eqn.outBinders.map((x) => x.aval.dtype)
+			}, eqn.params);
+			const inputs = [];
+			for (const input of eqn.inputs) if (input instanceof Var) {
+				const jv = ctx.get(input);
+				if (jv.type !== "imm") throw new Error(`jit: routine primitive ${eqn.primitive} input is not imm`);
+				inputs.push(jv.arg);
+			} else if (input instanceof Lit) inputs.push(builder.pushLit(input));
+			const outputs = [];
+			for (const outVar$1 of eqn.outBinders) {
+				const outId = builder.pushBuffer(outVar$1.aval.size * byteWidth(outVar$1.aval.dtype));
+				outputs.push(outId);
+				ctx.set(outVar$1, {
+					type: "imm",
+					arg: outId
+				});
+			}
+			builder.pushRoutine(routine, inputs, outputs);
+			continue;
+		}
 		const inputExps = [];
 		const inputAvals = [];
 		const inputArgs = [];
-		for (const input of eqn.inputs) if (input instanceof Var) {
-			const jitValue = ctx.get(input);
-			if (jitValue.type === "exp") {
-				const gidMap = /* @__PURE__ */ new Map();
-				for (const [gid, jitId] of jitValue.args.entries()) {
-					let newGid = inputArgs.indexOf(jitId);
-					if (newGid === -1) {
-						newGid = inputArgs.length;
-						inputArgs.push(jitId);
-					}
-					gidMap.set(gid, newGid);
-				}
-				inputExps.push(jitValue.exp.reindexGids(gidMap));
-			} else if (jitValue.type === "imm") {
-				let gid = inputArgs.indexOf(jitValue.arg);
-				if (gid === -1) {
-					gid = inputArgs.length;
-					inputArgs.push(jitValue.arg);
+		let inputReduction = null;
+		const addArgs = (args) => {
+			const newGids = [];
+			for (const jitId of args) {
+				let newGid = inputArgs.indexOf(jitId);
+				if (newGid === -1) {
+					newGid = inputArgs.length;
+					inputArgs.push(jitId);
 				}
+				newGids.push(newGid);
+			}
+			return newGids;
+		};
+		for (const input of eqn.inputs) if (input instanceof Var) {
+			const jv = ctx.get(input);
+			if (jv.type === "exp") {
+				const newGids = addArgs(jv.args);
+				inputExps.push(jv.exp.reindexGids(newGids));
+			} else if (jv.type === "imm") {
+				const [gid] = addArgs([jv.arg]);
 				const st = ShapeTracker.fromShape(input.aval.shape);
 				const indices = unravelAlu(st.shape, AluVar.gidx);
 				inputExps.push(AluExp.globalView(input.aval.dtype, gid, st, indices));
+			} else if (jv.type === "red") {
+				if (inputReduction) throw new Error("jit: unexpected, multiple red inputs");
+				const newGids = addArgs(jv.args);
+				inputExps.push(jv.reduction.epilogue.reindexGids(newGids));
+				inputReduction = jv;
 			}
 			inputAvals.push(input.aval);
 		} else if (input instanceof Lit) {
 			inputExps.push(AluExp.const(input.dtype, input.value));
 			inputAvals.push(input.aval);
 		} else throw new TypeError(`Unexpected input in Jaxpr: ${input}`);
-		const nargs$1 = inputArgs.length;
 		const rule = jitRules[eqn.primitive];
 		if (!rule) throw new TypeError(`JIT not implemented for primitive ${eqn.primitive}`);
-		const kernel = rule(nargs$1, inputExps, inputAvals, eqn.params);
+		let exp$2;
+		let reduction;
+		if (inputReduction) {
+			const jv = inputReduction;
+			const newEpilogue = rule(inputExps, inputAvals, eqn.params).exp;
+			exp$2 = jv.exp.reindexGids(addArgs(jv.args));
+			reduction = new Reduction(jv.reduction.dtype, jv.reduction.op, jv.reduction.size, newEpilogue);
+		} else {
+			const ruleOutput = rule(inputExps, inputAvals, eqn.params);
+			exp$2 = ruleOutput.exp;
+			reduction = ruleOutput.reduction;
+		}
 		const outVar = eqn.outBinders[0];
-		if (kernel.reduction || blackNodes.has(outVar)) {
+		if (blackNodes.has(outVar)) {
+			const nargs$1 = inputArgs.length;
+			const size$1 = outVar.aval.size;
+			const kernel = new Kernel(nargs$1, size$1, exp$2, reduction);
 			const outId = builder.pushKernel(kernel, inputArgs);
 			ctx.set(outVar, {
 				type: "imm",
 				arg: outId
 			});
-		} else ctx.set(outVar, {
+		} else if (reduction) ctx.set(outVar, {
+			type: "red",
+			exp: exp$2,
+			reduction,
+			args: inputArgs
+		});
+		else ctx.set(outVar, {
 			type: "exp",
-			exp: kernel.exp,
+			exp: exp$2,
 			args: inputArgs
 		});
 	}
@@ -1806,7 +1966,7 @@ function jitCompile(backend, jaxpr, consts) {
 		if (jitValue.type !== "imm") throw new Error("internal: Expected imm, since outs are black nodes");
 		outputIds.push(jitValue.arg);
 	} else if (out instanceof Lit) outputIds.push(builder.pushLit(out));
-	const outputNeedsRef = new Set([...range(nargs), ...builder.steps.filter((s) => s.type === "const").map((s) => s.output)]);
+	const outputNeedsRef = new Set(range(nargs));
 	for (const outputId of outputIds) if (outputNeedsRef.has(outputId)) builder.pushIncref(outputId);
 	else outputNeedsRef.add(outputId);
 	builder.insertFreeSteps(outputIds);
@@ -1828,31 +1988,33 @@ function reshapeViews(exp$2, mapping, reduceAxis = false) {
 	});
 }
 function broadcastedJit(fn, opts) {
-	return (nargs, exps, avals, params) => {
+	return (exps, avals, params) => {
 		let { shape: newShape, dtype: newDtype } = avals.reduce(promoteAvals);
 		const skipCastIdx = opts?.skipCastIdx ?? [];
 		if (skipCastIdx.length) newDtype = avals.filter((_, i) => !skipCastIdx.includes(i)).reduce(promoteAvals).dtype;
-		exps = exps.map((exp$3, i) => {
-			exp$3 = reshapeViews(exp$3, (st) => {
+		exps = exps.map((exp$2, i) => {
+			exp$2 = reshapeViews(exp$2, (st) => {
 				if (!deepEqual(st.shape, newShape)) return st.broadcast(newShape, range(newShape.length - st.shape.length));
 			});
-			if (exp$3.dtype !== newDtype && !skipCastIdx.includes(i)) exp$3 = AluExp.cast(newDtype, exp$3);
-			return exp$3;
+			if (exp$2.dtype !== newDtype && !skipCastIdx.includes(i)) exp$2 = AluExp.cast(newDtype, exp$2);
+			return exp$2;
 		});
-		const exp$2 = fn(exps, params);
-		return new Kernel(nargs, prod(newShape), exp$2);
+		return { exp: fn(exps, params) };
 	};
 }
 function unopJit(fn) {
-	return (nargs, [a], [as], params) => {
-		return new Kernel(nargs, prod(as.shape), fn(a, params));
+	return ([a], [_as], params) => {
+		return { exp: fn(a, params) };
 	};
 }
 function reshapeJit(fn) {
-	return (nargs, [a], [as], params) => {
-		a = reshapeViews(a, (st) => fn(st, params));
-		const newShape = fn(ShapeTracker.fromShape(as.shape), params).shape;
-		return new Kernel(nargs, prod(newShape), a);
+	return ([a], [_as], params) => {
+		return { exp: reshapeViews(a, (st) => fn(st, params)) };
+	};
+}
+function routineNoJit() {
+	return () => {
+		throw new Error("jit: rule is not implemented for routines");
 	};
 }
 const jitRules = {
@@ -1860,6 +2022,8 @@ const jitRules = {
 	[Primitive.Mul]: broadcastedJit(([a, b]) => AluExp.mul(a, b)),
 	[Primitive.Idiv]: broadcastedJit(([a, b]) => AluExp.idiv(a, b)),
 	[Primitive.Mod]: broadcastedJit(([a, b]) => AluExp.mod(a, b)),
+	[Primitive.Min]: broadcastedJit(([a, b]) => AluExp.min(a, b)),
+	[Primitive.Max]: broadcastedJit(([a, b]) => AluExp.max(a, b)),
 	[Primitive.Neg]: unopJit((a) => AluExp.sub(AluExp.const(a.dtype, 0), a)),
 	[Primitive.Reciprocal]: unopJit(AluExp.reciprocal),
 	[Primitive.Floor]: unopJit(AluExp.floor),
@@ -1867,17 +2031,6 @@ const jitRules = {
 	[Primitive.StopGradient]: unopJit((a) => a),
 	[Primitive.Cast]: unopJit((a, { dtype }) => AluExp.cast(dtype, a)),
 	[Primitive.Bitcast]: unopJit((a, { dtype }) => AluExp.bitcast(dtype, a)),
-	[Primitive.RandomBits]: (nargs, keys, keyShapes, { shape: shape$1, mode }) => {
-		const mapping = (st) => {
-			if (!deepEqual(st.shape, shape$1)) return st.broadcast(shape$1, range(shape$1.length - st.shape.length));
-		};
-		const k0 = reshapeViews(keys[0], mapping);
-		const k1 = reshapeViews(keys[1], mapping);
-		const c0 = AluExp.u32(0);
-		const c1 = AluExp.cast(DType.Uint32, AluVar.gidx);
-		const exp$2 = AluExp.threefry2x32(k0, k1, c0, c1, mode);
-		return new Kernel(nargs, prod(shape$1), exp$2);
-	},
 	[Primitive.Sin]: unopJit(AluExp.sin),
 	[Primitive.Cos]: unopJit(AluExp.cos),
 	[Primitive.Asin]: unopJit(AluExp.asin),
@@ -1887,9 +2040,7 @@ const jitRules = {
 	[Primitive.Erf]: unopJit(AluExp.erf),
 	[Primitive.Erfc]: unopJit(AluExp.erfc),
 	[Primitive.Sqrt]: unopJit(AluExp.sqrt),
-	[Primitive.Min]: broadcastedJit(([a, b]) => AluExp.min(a, b)),
-	[Primitive.Max]: broadcastedJit(([a, b]) => AluExp.max(a, b)),
-	[Primitive.Reduce](nargs, [a], [as], { op, axis }) {
+	[Primitive.Reduce]([a], [as], { op, axis }) {
 		const keptAxes = [];
 		const shiftedAxes = [];
 		const newShape = [];
@@ -1898,53 +2049,58 @@ const jitRules = {
 			keptAxes.push(i);
 			newShape.push(as.shape[i]);
 		}
-		const size$1 = prod(newShape);
 		const reductionSize = prod(shiftedAxes.map((ax) => as.shape[ax]));
 		newShape.push(reductionSize);
 		const perm = keptAxes.concat(shiftedAxes);
 		a = reshapeViews(a, (st) => st.permute(perm).reshape(newShape), true);
 		const reduction = new Reduction(a.dtype, op, reductionSize);
-		return new Kernel(nargs, size$1, a, reduction);
+		return {
+			exp: a,
+			reduction
+		};
 	},
 	[Primitive.Pool]: reshapeJit((st, { window, strides }) => pool(st, window, strides)),
-	[Primitive.PoolTranspose](nargs, [a], [as], { inShape, window, strides }) {
+	[Primitive.PoolTranspose]([a], [as], { inShape, window, strides }) {
 		let stX = poolTranspose(ShapeTracker.fromShape(as.shape), inShape, window, strides);
-		const size$1 = prod(inShape);
 		stX = stX.reshape([...inShape, prod(stX.shape.slice(inShape.length))]);
 		a = reshapeViews(a, (st) => st.compose(stX), true);
 		const reduction = new Reduction(a.dtype, AluOp.Add, stX.shape[stX.shape.length - 1]);
-		return new Kernel(nargs, size$1, a, reduction);
+		return {
+			exp: a,
+			reduction
+		};
 	},
-	[Primitive.Dot](nargs, [a, b], [as, bs]) {
-		const k1 = jitRules[Primitive.Mul](nargs, [a, b], [as, bs], {});
+	[Primitive.Dot]([a, b], [as, bs]) {
+		const k1 = jitRules[Primitive.Mul]([a, b], [as, bs], {});
 		const c = k1.exp;
 		const cs = promoteAvals(as, bs);
-		return jitRules[Primitive.Reduce](nargs, [c], [cs], {
+		return jitRules[Primitive.Reduce]([c], [cs], {
 			op: AluOp.Add,
 			axis: [cs.ndim - 1]
 		});
 	},
-	[Primitive.Conv](nargs, [a, b], [as, bs], params) {
+	[Primitive.Conv]([a, b], [as, bs], params) {
 		const [stX, stY] = prepareConv(ShapeTracker.fromShape(as.shape), ShapeTracker.fromShape(bs.shape), params);
 		a = reshapeViews(a, (st) => st.compose(stX));
 		b = reshapeViews(b, (st) => st.compose(stY));
 		as = new ShapedArray(stX.shape, as.dtype, as.weakType);
 		bs = new ShapedArray(stY.shape, bs.dtype, bs.weakType);
-		return jitRules[Primitive.Dot](nargs, [a, b], [as, bs], {});
+		return jitRules[Primitive.Dot]([a, b], [as, bs], {});
 	},
 	[Primitive.Compare]: broadcastedJit(([a, b], { op }) => aluCompare(a, b, op)),
 	[Primitive.Where]: broadcastedJit(([cond, a, b]) => AluExp.where(cond, a, b), { skipCastIdx: [0] }),
-	[Primitive.Transpose]: reshapeJit((st, { perm }) => st.permute(perm)),
-	[Primitive.Broadcast]: reshapeJit((st, { shape: shape$1, axis }) => st.broadcast(shape$1, axis)),
-	[Primitive.Reshape]: reshapeJit((st, { shape: shape$1 }) => st.reshape(shape$1)),
-	[Primitive.Flip]: reshapeJit((st, { axis }) => {
-		const arg = rep(st.shape.length, false);
-		for (const ax of axis) arg[ax] = true;
-		return st.flip(arg);
-	}),
-	[Primitive.Shrink]: reshapeJit((st, { slice }) => st.shrink(slice)),
-	[Primitive.Pad]: reshapeJit((st, { width }) => st.pad(width)),
-	[Primitive.Gather](nargs, [x, ...indices], [xs, ...indicesShapes], { axis, outDim }) {
+	[Primitive.RandomBits]: (keys, keyShapes, { shape: shape$1, mode }) => {
+		const mapping = (st) => {
+			if (!deepEqual(st.shape, shape$1)) return st.broadcast(shape$1, range(shape$1.length - st.shape.length));
+		};
+		const k0 = reshapeViews(keys[0], mapping);
+		const k1 = reshapeViews(keys[1], mapping);
+		const c0 = AluExp.u32(0);
+		const c1 = AluExp.cast(DType.Uint32, AluVar.gidx);
+		const exp$2 = AluExp.threefry2x32(k0, k1, c0, c1, mode);
+		return { exp: exp$2 };
+	},
+	[Primitive.Gather]([x, ...indices], [xs, ...indicesShapes], { axis, outDim }) {
 		const axisSet = new Set(axis);
 		const indexShape = indicesShapes.map((c) => c.shape).reduce(generalBroadcast);
 		const finalShape = xs.shape.filter((_, i) => !axisSet.has(i));
@@ -1957,24 +2113,38 @@ const jitRules = {
 		for (const [i, iexp] of indices.entries()) src[axis[i]] = AluExp.cast(DType.Int32, reshapeViews(iexp, (st) => st.broadcast(finalShape, [...range(outDim + indexShape.length - st.shape.length), ...range(outDim + indexShape.length, finalShape.length)])));
 		const [index, valid] = ShapeTracker.fromShape(xs.shape).toAluExp(src);
 		if (!valid.resolve()) throw new Error("internal: expected full validity mask in Gather");
-		return new Kernel(nargs, prod(finalShape), x.substitute({ gidx: index }));
+		return { exp: x.substitute({ gidx: index }) };
 	},
-	[Primitive.JitCall]() {
-		throw new Error("internal: JitCall should have been flattened before JIT compilation");
+	[Primitive.Transpose]: reshapeJit((st, { perm }) => st.permute(perm)),
+	[Primitive.Broadcast]: reshapeJit((st, { shape: shape$1, axis }) => st.broadcast(shape$1, axis)),
+	[Primitive.Reshape]: reshapeJit((st, { shape: shape$1 }) => st.reshape(shape$1)),
+	[Primitive.Flip]: reshapeJit((st, { axis }) => {
+		const arg = rep(st.shape.length, false);
+		for (const ax of axis) arg[ax] = true;
+		return st.flip(arg);
+	}),
+	[Primitive.Shrink]: reshapeJit((st, { slice }) => st.shrink(slice)),
+	[Primitive.Pad]: reshapeJit((st, { width }) => st.pad(width)),
+	[Primitive.Sort]: routineNoJit(),
+	[Primitive.Argsort]: routineNoJit(),
+	[Primitive.TriangularSolve]: routineNoJit(),
+	[Primitive.Cholesky]: routineNoJit(),
+	[Primitive.Jit]() {
+		throw new Error("internal: Jit should have been flattened before JIT compilation");
 	}
 };
 /** Determines how to split the Jaxpr into kernels via dataflow analysis. */
 function splitGraphDataflow(backend, jaxpr) {
-	const varToEqn = /* @__PURE__ */ new Map();
+	const varToDefn = /* @__PURE__ */ new Map();
+	const varToUsages = /* @__PURE__ */ new Map();
 	for (let i = 0; i < jaxpr.eqns.length; i++) {
 		const eqn = jaxpr.eqns[i];
-		for (const v of eqn.outBinders) if (v instanceof Var) varToEqn.set(v, i);
-	}
-	const blackNodes = /* @__PURE__ */ new Set();
-	const p1NextBlack = /* @__PURE__ */ new Map();
-	for (const v of jaxpr.outs) if (v instanceof Var) {
-		blackNodes.add(v);
-		p1NextBlack.set(v, v);
+		for (const v of eqn.outBinders) if (v instanceof Var) varToDefn.set(v, i);
+		for (const input of eqn.inputs) if (input instanceof Var) {
+			const usages = varToUsages.get(input);
+			if (usages) usages.push(i);
+			else varToUsages.set(input, [i]);
+		}
 	}
 	const reducePrimitives = [
 		Primitive.Reduce,
@@ -1982,28 +2152,94 @@ function splitGraphDataflow(backend, jaxpr) {
 		Primitive.Conv,
 		Primitive.PoolTranspose
 	];
-	const heterogeneousViewPrimitives = [Primitive.Gather, Primitive.RandomBits];
-	for (let i = jaxpr.eqns.length - 1; i >= 0; i--) {
+	const reductionEpilogueEqns = /* @__PURE__ */ new Set();
+	const reductionEndpointEqns = /* @__PURE__ */ new Set();
+	for (let i = 0; i < jaxpr.eqns.length; i++) {
 		const eqn = jaxpr.eqns[i];
-		if (reducePrimitives.includes(eqn.primitive) || heterogeneousViewPrimitives.includes(eqn.primitive) || eqn.outBinders.some((v) => blackNodes.has(v))) {
-			for (const v of eqn.outBinders) {
-				blackNodes.add(v);
-				p1NextBlack.set(v, v);
+		if (reducePrimitives.includes(eqn.primitive)) {
+			let head = i;
+			while (true) {
+				reductionEpilogueEqns.add(head);
+				const outVar = jaxpr.eqns[head].outBinders[0];
+				const usages = varToUsages.get(outVar) ?? [];
+				if (jaxpr.outs.includes(outVar) || usages.length !== 1) break;
+				if (reductionEpilogueEqns.has(usages[0])) break;
+				const nextEqn = jaxpr.eqns[usages[0]];
+				switch (nextEqn.primitive) {
+					case Primitive.Neg:
+					case Primitive.Reciprocal:
+					case Primitive.Floor:
+					case Primitive.Ceil:
+					case Primitive.StopGradient:
+					case Primitive.Cast:
+					case Primitive.Bitcast:
+					case Primitive.Sin:
+					case Primitive.Cos:
+					case Primitive.Asin:
+					case Primitive.Atan:
+					case Primitive.Exp:
+					case Primitive.Log:
+					case Primitive.Erf:
+					case Primitive.Erfc:
+					case Primitive.Sqrt:
+						head = usages[0];
+						continue;
+					case Primitive.Add:
+					case Primitive.Mul:
+					case Primitive.Idiv:
+					case Primitive.Mod:
+					case Primitive.Min:
+					case Primitive.Max: {
+						const otherInput = nextEqn.inputs.find((v) => v !== outVar);
+						if (otherInput instanceof Lit || deepEqual(generalBroadcast(otherInput.aval.shape, outVar.aval.shape), outVar.aval.shape)) {
+							head = usages[0];
+							continue;
+						}
+						break;
+					}
+				}
+				break;
 			}
-			continue;
+			reductionEndpointEqns.add(head);
 		}
-		const reach = /* @__PURE__ */ new Set();
-		for (let j = i + 1; j < jaxpr.eqns.length; j++) for (const v of jaxpr.eqns[j].inputs) if (v instanceof Var && eqn.outBinders.includes(v)) for (const o of jaxpr.eqns[j].outBinders) {
-			const u = p1NextBlack.get(o);
-			if (u) reach.add(u);
-		}
-		if (reach.size === 1) {
-			const b = reach.values().next().value;
-			for (const v of eqn.outBinders) p1NextBlack.set(v, b);
-		} else if (reach.size > 1) for (const v of eqn.outBinders) {
+	}
+	const blackNodes = /* @__PURE__ */ new Set();
+	const p1NextBlack = /* @__PURE__ */ new Map();
+	for (const v of jaxpr.outs) if (v instanceof Var) {
+		blackNodes.add(v);
+		p1NextBlack.set(v, v);
+	}
+	const heterogeneousViewPrimitives = [Primitive.RandomBits, Primitive.Gather];
+	const needsCleanShapePrimitives = [Primitive.Pad];
+	for (let i = jaxpr.eqns.length - 1; i >= 0; i--) {
+		const eqn = jaxpr.eqns[i];
+		if (reductionEndpointEqns.has(i) || heterogeneousViewPrimitives.includes(eqn.primitive) || routinePrimitives.has(eqn.primitive) || eqn.outBinders.some((v) => blackNodes.has(v))) {
+			for (const v of eqn.outBinders) {
+				blackNodes.add(v);
+				p1NextBlack.set(v, v);
+			}
+			continue;
+		}
+		const reach = /* @__PURE__ */ new Set();
+		let needsCleanOutput = false;
+		outer: for (const v of eqn.outBinders) for (const j of varToUsages.get(v) ?? []) {
+			if (needsCleanShapePrimitives.includes(jaxpr.eqns[j].primitive) || routinePrimitives.has(jaxpr.eqns[j].primitive)) {
+				needsCleanOutput = true;
+				break outer;
+			}
+			for (const o of jaxpr.eqns[j].outBinders) {
+				const u = p1NextBlack.get(o);
+				if (u) reach.add(u);
+			}
+		}
+		if (reach.size > 1 || needsCleanOutput) for (const v of eqn.outBinders) {
 			blackNodes.add(v);
 			p1NextBlack.set(v, v);
 		}
+		else if (reach.size === 1) {
+			const b = reach.values().next().value;
+			for (const v of eqn.outBinders) p1NextBlack.set(v, b);
+		}
 	}
 	const p2Deps = /* @__PURE__ */ new Map();
 	for (const v of jaxpr.inBinders) p2Deps.set(v, new Set([v]));
@@ -2011,7 +2247,6 @@ function splitGraphDataflow(backend, jaxpr) {
 	while (p2idx < jaxpr.eqns.length) {
 		const eqn = jaxpr.eqns[p2idx++];
 		const deps = [];
-		if (eqn.outBinders.some((v) => blackNodes.has(v))) continue;
 		for (const input of eqn.inputs) if (input instanceof Var) if (blackNodes.has(input)) deps.push(new Set([input]));
 		else deps.push(p2Deps.get(input));
 		else deps.push(/* @__PURE__ */ new Set());
@@ -2022,7 +2257,7 @@ function splitGraphDataflow(backend, jaxpr) {
 			let assocInput = -1;
 			for (let i = 0; i < eqn.inputs.length; i++) {
 				const input = eqn.inputs[i];
-				if (input instanceof Var && varToEqn.has(input)) {
+				if (input instanceof Var && varToDefn.has(input)) {
 					let uniqueDeps = 0;
 					for (const dep of deps[i]) if (depCounter.get(dep) === 1) uniqueDeps++;
 					if (uniqueDeps > maxUniqueDeps) {
@@ -2033,8 +2268,8 @@ function splitGraphDataflow(backend, jaxpr) {
 			}
 			if (assocInput === -1) throw new Error(`internal: maxArgs, no input found to mark as black in Jaxpr equation ${eqn}`);
 			const assocVar = eqn.inputs[assocInput];
-			p2idx = varToEqn.get(assocVar);
-			for (const out of jaxpr.eqns[p2idx].outBinders) blackNodes.add(out);
+			p2idx = varToDefn.get(assocVar);
+			for (const out of jaxpr.eqns[p2idx++].outBinders) blackNodes.add(out);
 		} else {
 			const s = new Set(depCounter.keys());
 			for (const out of eqn.outBinders) p2Deps.set(out, s);
@@ -2060,9 +2295,9 @@ var PendingExecute = class {
 	submitted = false;
 	#promise = null;
 	#rc = 1;
-	constructor(backend, kernel, inputs, outputs) {
+	constructor(backend, source, inputs, outputs) {
 		this.backend = backend;
-		this.kernel = kernel;
+		this.source = source;
 		this.inputs = inputs;
 		this.outputs = outputs;
 		for (const slot of inputs) this.backend.incRef(slot);
@@ -2083,13 +2318,15 @@ var PendingExecute = class {
 			return;
 		}
 		this.#promise = (async () => {
-			this.prepared = await this.backend.prepare(this.kernel);
+			if (this.source instanceof Kernel) this.prepared = await this.backend.prepareKernel(this.source);
+			else this.prepared = await this.backend.prepareRoutine(this.source);
 		})();
 		await this.#promise;
 	}
 	prepareSync() {
 		if (this.prepared) return;
-		this.prepared = this.backend.prepareSync(this.kernel);
+		if (this.source instanceof Kernel) this.prepared = this.backend.prepareKernelSync(this.source);
+		else this.prepared = this.backend.prepareRoutineSync(this.source);
 	}
 	submit() {
 		if (this.submitted) return;
@@ -2112,8 +2349,6 @@ var PendingExecute = class {
 * "Array" type by name.
 */
 var Array$1 = class Array$1 extends Tracer {
-	static #nextId = 1001;
-	id;
 	#dtype;
 	#weakType;
 	#source;
@@ -2130,7 +2365,6 @@ var Array$1 = class Array$1 extends Tracer {
 	*/
 	constructor(args) {
 		super(baseArrayTrace);
-		this.id = Array$1.#nextId++;
 		this.#dtype = args.dtype;
 		this.#weakType = args.weakType;
 		this.#source = args.source;
@@ -2439,6 +2673,27 @@ var Array$1 = class Array$1 extends Tracer {
 			pending
 		});
 	}
+	/** Apply an operation with custom lowering to this array. */
+	static #routine(routine, arrays, outputWeakType) {
+		const { backend, committed } = Array$1.#computeBackend(routine.name, arrays);
+		for (const ar of arrays) ar.#realize();
+		const inputs = arrays.map((ar) => ar.#source);
+		const outputs = routine.type.outputDtypes.map((dtype, i) => backend.malloc(byteWidth(dtype) * prod(routine.type.outputShapes[i])));
+		const pending = arrays.flatMap((ar) => ar.#pending);
+		for (const exe of pending) exe.updateRc(+outputs.length);
+		pending.push(new PendingExecute(backend, routine, inputs, outputs));
+		pending[pending.length - 1].updateRc(+outputs.length - 1);
+		arrays.forEach((ar) => ar.dispose());
+		return outputs.map((output, i) => new Array$1({
+			source: output,
+			st: ShapeTracker.fromShape(routine.type.outputShapes[i]),
+			dtype: routine.type.outputDtypes[i],
+			weakType: outputWeakType[i],
+			backend,
+			committed,
+			pending
+		}));
+	}
 	/**
 	* Normalizes this array into one backed by a `Slot`.
 	*
@@ -2599,6 +2854,12 @@ var Array$1 = class Array$1 extends Tracer {
 			[Primitive.Mod]([x, y]) {
 				return [x.#binary(AluOp.Mod, y)];
 			},
+			[Primitive.Min]([x, y]) {
+				return [x.#binary(AluOp.Min, y)];
+			},
+			[Primitive.Max]([x, y]) {
+				return [x.#binary(AluOp.Max, y)];
+			},
 			[Primitive.Neg]([x]) {
 				return [zerosLike$1(x.ref).#binary(AluOp.Sub, x)];
 			},
@@ -2635,25 +2896,6 @@ var Array$1 = class Array$1 extends Tracer {
 					return [y];
 				}
 			},
-			[Primitive.RandomBits]([k0, k1], { shape: shape$1, mode }) {
-				const keyShape = generalBroadcast(k0.shape, k1.shape);
-				if (!deepEqual(generalBroadcast(keyShape, shape$1), shape$1)) throw new TypeError(`Keys of shapes ${k0.shape} and ${k1.shape} cannot be broadcast to shape ${shape$1}`);
-				const c0 = zeros(shape$1, {
-					dtype: DType.Uint32,
-					device: k0.device
-				});
-				const c1 = arange(0, prod(shape$1), 1, {
-					dtype: DType.Uint32,
-					device: k0.device
-				}).reshape(shape$1);
-				const custom = ([k0$1, k1$1, c0$1, c1$1]) => AluExp.threefry2x32(k0$1, k1$1, c0$1, c1$1, mode);
-				return [Array$1.#naryCustom("random_bits", custom, [
-					k0,
-					k1,
-					c0,
-					c1
-				])];
-			},
 			[Primitive.Sin]([x]) {
 				return [x.#unary(AluOp.Sin)];
 			},
@@ -2681,12 +2923,6 @@ var Array$1 = class Array$1 extends Tracer {
 			[Primitive.Sqrt]([x]) {
 				return [x.#unary(AluOp.Sqrt)];
 			},
-			[Primitive.Min]([x, y]) {
-				return [x.#binary(AluOp.Min, y)];
-			},
-			[Primitive.Max]([x, y]) {
-				return [x.#binary(AluOp.Max, y)];
-			},
 			[Primitive.Reduce]([x], { op, axis }) {
 				if (axis.length === 0) return [x];
 				return [x.#moveAxesDown(axis).#reduce(op)];
@@ -2721,6 +2957,28 @@ var Array$1 = class Array$1 extends Tracer {
 					y
 				], { dtypeOverride: [DType.Bool] })];
 			},
+			[Primitive.RandomBits]([k0, k1], { shape: shape$1, mode }) {
+				const keyShape = generalBroadcast(k0.shape, k1.shape);
+				if (!deepEqual(generalBroadcast(keyShape, shape$1), shape$1)) throw new TypeError(`Keys of shapes ${k0.shape} and ${k1.shape} cannot be broadcast to shape ${shape$1}`);
+				const c0 = zeros(shape$1, {
+					dtype: DType.Uint32,
+					device: k0.device
+				});
+				const c1 = arange(0, prod(shape$1), 1, {
+					dtype: DType.Uint32,
+					device: k0.device
+				}).reshape(shape$1);
+				const custom = ([k0$1, k1$1, c0$1, c1$1]) => AluExp.threefry2x32(k0$1, k1$1, c0$1, c1$1, mode);
+				return [Array$1.#naryCustom("random_bits", custom, [
+					k0,
+					k1,
+					c0,
+					c1
+				])];
+			},
+			[Primitive.Gather]([x, ...indices], { axis, outDim }) {
+				return [x.#gather(indices, axis, outDim)];
+			},
 			[Primitive.Transpose]([x], { perm }) {
 				return [x.#transpose(perm)];
 			},
@@ -2741,17 +2999,48 @@ var Array$1 = class Array$1 extends Tracer {
 			[Primitive.Pad]([x], { width }) {
 				return [x.#reshape(x.#st.pad(width))];
 			},
-			[Primitive.Gather]([x, ...indices], { axis, outDim }) {
-				return [x.#gather(indices, axis, outDim)];
+			[Primitive.Sort]([x]) {
+				const routine = new Routine(Routines.Sort, {
+					inputShapes: [x.aval.shape],
+					inputDtypes: [x.aval.dtype],
+					outputShapes: [x.aval.shape],
+					outputDtypes: [x.aval.dtype]
+				});
+				return Array$1.#routine(routine, [x], [x.#weakType]);
+			},
+			[Primitive.Argsort]([x]) {
+				const routine = new Routine(Routines.Argsort, {
+					inputShapes: [x.aval.shape],
+					inputDtypes: [x.aval.dtype],
+					outputShapes: [x.aval.shape, x.aval.shape],
+					outputDtypes: [x.aval.dtype, DType.Int32]
+				});
+				return Array$1.#routine(routine, [x], [x.#weakType, false]);
+			},
+			[Primitive.TriangularSolve]([a, b], { unitDiagonal }) {
+				const routine = new Routine(Routines.TriangularSolve, {
+					inputShapes: [a.aval.shape, b.aval.shape],
+					inputDtypes: [a.aval.dtype, b.aval.dtype],
+					outputShapes: [b.aval.shape],
+					outputDtypes: [b.aval.dtype]
+				}, { unitDiagonal });
+				return Array$1.#routine(routine, [a, b], [a.#weakType && b.#weakType]);
 			},
-			[Primitive.JitCall](args, { jaxpr, numConsts }) {
-				if (jaxpr.inBinders.length !== args.length) throw new Error(`jit_call expects ${jaxpr.inBinders.length} args, got ${args.length}`);
-				const { backend, committed } = Array$1.#computeBackend("jit_call", args);
+			[Primitive.Cholesky]([a]) {
+				const routine = new Routine(Routines.Cholesky, {
+					inputShapes: [a.aval.shape],
+					inputDtypes: [a.aval.dtype],
+					outputShapes: [a.aval.shape],
+					outputDtypes: [a.aval.dtype]
+				});
+				return Array$1.#routine(routine, [a], [a.#weakType]);
+			},
+			[Primitive.Jit](args, { jaxpr }) {
+				if (jaxpr.inBinders.length !== args.length) throw new Error(`jit expects ${jaxpr.inBinders.length} args, got ${args.length}`);
+				const { backend, committed } = Array$1.#computeBackend("jit", args);
 				args = args.map((ar) => ar._putSync(backend));
-				const consts = args.slice(0, numConsts);
-				const tracers = args.slice(numConsts);
-				const jp = jitCompile(backend, jaxpr, consts);
-				const { outputs, pending } = jp.execute(tracers.map((x) => x._realizeSource()));
+				const jp = jitCompile(backend, jaxpr);
+				const { outputs, pending } = jp.execute(args.map((x) => x._realizeSource()));
 				for (const exe of pending) exe.updateRc(+outputs.length - 1);
 				const prevPending = [...new Set(args.flatMap((x) => x.#pending))];
 				for (const exe of prevPending) exe.updateRc(+outputs.length);
@@ -3050,6 +3339,43 @@ function arange(start, stop, step = 1, { dtype, device } = {}) {
 	});
 }
 /**
+* Return an array with ones on and below the diagonal and zeros elsewhere.
+*
+* If `k` is provided, it specifies the sub-diagonal on and below which the
+* array is filled with ones. `k=0` is the main diagonal, `k<0` is below it, and
+* `k>0` is above it.
+*/
+function tri(n, m, k = 0, { dtype, device } = {}) {
+	m ??= n;
+	dtype ??= DType.Float32;
+	if (!Number.isInteger(n) || n < 0) throw new Error(`tri: n must be a non-negative integer, got ${n}`);
+	if (!Number.isInteger(m) || m < 0) throw new Error(`tri: m must be a non-negative integer, got ${m}`);
+	if (!Number.isInteger(k)) throw new Error(`tri: k must be an integer, got ${k}`);
+	const rows = arange(k, n + k, 1, {
+		dtype: DType.Int32,
+		device
+	});
+	const cols = arange(0, m, 1, {
+		dtype: DType.Int32,
+		device
+	});
+	return rows.reshape([n, 1]).greaterEqual(cols).astype(dtype);
+}
+/** Return the lower triangle of an array. Must be of dimension >= 2. */
+function tril(a, k = 0) {
+	if (ndim$1(a) < 2) throw new Error(`tril: input array must be at least 2D, got ${ndim$1(a)}D`);
+	a = fudgeArray(a);
+	const [n, m] = a.shape.slice(-2);
+	return where$1(tri(n, m, k, { dtype: DType.Bool }), a.ref, zerosLike$1(a));
+}
+/** Return the upper triangle of an array. Must be of dimension >= 2. */
+function triu(a, k = 0) {
+	if (ndim$1(a) < 2) throw new Error(`tril: input array must be at least 2D, got ${ndim$1(a)}D`);
+	a = fudgeArray(a);
+	const [n, m] = a.shape.slice(-2);
+	return where$1(tri(n, m, k - 1, { dtype: DType.Bool }), zerosLike$1(a.ref), a);
+}
+/**
 * Return evenly spaced numbers over a specified interval.
 *
 * Returns _num_ evenly spaced samples, calculated over the interval
@@ -3096,333 +3422,106 @@ function aluCompare(a, b, op) {
 }
 
 //#endregion
-//#region src/frontend/jvp.ts
-var JVPTracer = class extends Tracer {
-	constructor(trace$1, primal, tangent) {
+//#region src/frontend/vmap.ts
+function mappedAval(batchDim, aval) {
+	const shape$1 = [...aval.shape];
+	shape$1.splice(batchDim, 1);
+	return new ShapedArray(shape$1, aval.dtype, aval.weakType);
+}
+/** Move one axis to a different index. */
+function moveaxis(x, src, dst) {
+	const t = pureArray(x);
+	src = checkAxis(src, t.ndim);
+	dst = checkAxis(dst, t.ndim);
+	if (src === dst) return t;
+	const perm = range(t.ndim);
+	perm.splice(src, 1);
+	perm.splice(dst, 0, src);
+	return transpose$1(t, perm);
+}
+function moveBatchAxis(axisSize, src, dst, x) {
+	if (src === null) {
+		const targetShape = [...x.shape];
+		targetShape.splice(dst, 0, axisSize);
+		return broadcast(x, targetShape, [dst]);
+	} else if (src === dst) return x;
+	else return moveaxis(x, src, dst);
+}
+var BatchTracer = class extends Tracer {
+	constructor(trace$1, val, batchDim) {
 		super(trace$1);
-		this.primal = primal;
-		this.tangent = tangent;
+		this.val = val;
+		this.batchDim = batchDim;
 	}
 	get aval() {
-		return this.primal.aval;
+		if (this.batchDim === null) return this.val.aval;
+		else return mappedAval(this.batchDim, this.val.aval);
 	}
 	toString() {
-		return `JVPTracer(${this.primal.toString()}, ${this.tangent.toString()})`;
+		return `BatchTracer(${this.val.toString()}, ${this.batchDim})`;
 	}
 	get ref() {
-		this.primal.ref, this.tangent.ref;
+		this.val.ref;
 		return this;
 	}
 	dispose() {
-		this.primal.dispose();
-		this.tangent.dispose();
+		this.val.dispose();
+	}
+	fullLower() {
+		if (this.batchDim === null) return this.val.fullLower();
+		else return this;
 	}
 };
-var JVPTrace = class extends Trace {
+var BatchTrace = class extends Trace {
 	pure(val) {
 		return this.lift(pureArray(val));
 	}
 	lift(val) {
-		return new JVPTracer(this, val, zerosLike$1(val.ref));
+		return new BatchTracer(this, val, null);
 	}
 	processPrimitive(primitive, tracers, params) {
-		const [primalsIn, tangentsIn] = unzip2(tracers.map((x) => [x.primal, x.tangent]));
-		const jvpRule = jvpRules[primitive];
-		if (jvpRule === void 0) throw new Error(`No JVP rule for: ${primitive}`);
-		const [primalsOut, tangentsOut] = jvpRule(primalsIn, tangentsIn, params);
-		return zip(primalsOut, tangentsOut).map(([x, t]) => new JVPTracer(this, x, t));
+		const [valsIn, bdimsIn] = unzip2(tracers.map((t) => [t.val, t.batchDim]));
+		const vmapRule = vmapRules[primitive];
+		if (vmapRule === void 0) throw new Error(`No vmap rule for: ${primitive}`);
+		if (bdimsIn.every((d) => d === null)) {
+			const valOuts$1 = bind(primitive, valsIn, params);
+			return valOuts$1.map((x) => new BatchTracer(this, x, null));
+		}
+		const [valOuts, bdimOuts] = vmapRule(this.axisSize, valsIn, bdimsIn, params);
+		return zip(valOuts, bdimOuts).map(([x, bd]) => new BatchTracer(this, x, bd));
+	}
+	get axisSize() {
+		return this.main.globalData;
 	}
 };
-/** Rule that applies the same operation to primals and tangents. */
-function linearTangentsJvp(primitive) {
-	return (primals, tangents, params) => {
-		const ys = bind(primitive, primals, params);
-		const dys = bind(primitive, tangents, params);
-		return [ys, dys];
-	};
-}
-/** Rule for product of gradients in bilinear operations. */
-function bilinearTangentsJvp(primitive) {
-	return ([x, y], [dx, dy], params) => {
-		const primal = bind1(primitive, [x.ref, y.ref], params);
-		const tangent = bind1(primitive, [x, dy], params).add(bind1(primitive, [dx, y], params));
-		return [[primal], [tangent]];
+/**
+* Process a primitive with built-in broadcasting.
+*
+* Reference: https://github.com/jax-ml/jax/blob/jax-v0.8.1/jax/_src/interpreters/batching.py#L1029
+*/
+function broadcastBatcher(op) {
+	return (axisSize, args, dims) => {
+		if (args.length === 0) throw new Error("Empty list in broadcastBatcher");
+		const nd = Math.max(...args.map((x, i) => ndim$1(x) + (dims[i] === null ? 1 : 0)));
+		const firstIdx = dims.findIndex((d) => d !== null);
+		const firstBdim = dims[firstIdx] - args[firstIdx].ndim;
+		if (zip(args, dims).every(([x, d]) => d === null && ndim$1(x) < -firstBdim || d !== null && d - x.ndim === firstBdim)) return [[op(...args)], [nd + firstBdim]];
+		args = args.map((x, i) => {
+			if (dims[i] === null) return x;
+			x = moveBatchAxis(axisSize, dims[i], 0, x);
+			if (x.ndim < nd) x = x.reshape([
+				x.shape[0],
+				...rep(nd - x.ndim, 1),
+				...x.shape.slice(1)
+			]);
+			return x;
+		});
+		return [[op(...args)], [0]];
 	};
 }
-/** Rule that zeros out any tangents. */
-function zeroTangentsJvp(primitive) {
-	return (primals, tangents, params) => {
-		for (const t of tangents) t.dispose();
-		const ys = bind(primitive, primals, params);
-		return [ys, ys.map((y) => zerosLike$1(y.ref))];
-	};
-}
-const jvpRules = {
-	[Primitive.Add]: linearTangentsJvp(Primitive.Add),
-	[Primitive.Mul]: bilinearTangentsJvp(Primitive.Mul),
-	[Primitive.Idiv]: zeroTangentsJvp(Primitive.Idiv),
-	[Primitive.Mod]([x, y], [dx, dy]) {
-		if (!isFloatDtype(x.dtype) && !isFloatDtype(y.dtype)) {
-			dx.dispose();
-			dy.dispose();
-			return [[x.ref, y.ref], [zerosLike$1(x), zerosLike$1(y)]];
-		}
-		const q = idiv(x.ref, y.ref);
-		return [[mod(x, y)], [dx.sub(dy.mul(q))]];
-	},
-	[Primitive.Neg]: linearTangentsJvp(Primitive.Neg),
-	[Primitive.Reciprocal]([x], [dx]) {
-		const xRecip = reciprocal$1(x.ref);
-		return [[xRecip.ref], [neg(xRecip.ref.mul(xRecip)).mul(dx)]];
-	},
-	[Primitive.Floor]: zeroTangentsJvp(Primitive.Floor),
-	[Primitive.Ceil]: zeroTangentsJvp(Primitive.Ceil),
-	[Primitive.StopGradient]: zeroTangentsJvp(Primitive.StopGradient),
-	[Primitive.Cast]([x], [dx], { dtype }) {
-		if (x.dtype === dtype) return [[x], [dx]];
-		if (isFloatDtype(dtype) && isFloatDtype(x.dtype)) return [[cast(x, dtype)], [cast(dx, dtype)]];
-		else {
-			dx.dispose();
-			return [[cast(x.ref, dtype)], [zerosLike$1(x)]];
-		}
-	},
-	[Primitive.Bitcast]([x], [dx], { dtype }) {
-		if (x.dtype === dtype) return [[x], [dx]];
-		dx.dispose();
-		return [[bitcast(x.ref, dtype)], [zerosLike$1(x)]];
-	},
-	[Primitive.RandomBits]: zeroTangentsJvp(Primitive.RandomBits),
-	[Primitive.Sin]([x], [dx]) {
-		return [[sin$1(x.ref)], [cos$1(x).mul(dx)]];
-	},
-	[Primitive.Cos]([x], [dx]) {
-		return [[cos$1(x.ref)], [neg(sin$1(x)).mul(dx)]];
-	},
-	[Primitive.Asin]([x], [dx]) {
-		const denom = sqrt$1(reciprocal$1(cast(1, x.dtype).sub(x.ref.mul(x.ref))));
-		return [[asin$1(x)], [denom.mul(dx)]];
-	},
-	[Primitive.Atan]([x], [dx]) {
-		const denom = cast(1, x.dtype).add(x.ref.mul(x.ref));
-		return [[atan$1(x)], [dx.div(denom)]];
-	},
-	[Primitive.Exp]([x], [dx]) {
-		const z = exp$1(x);
-		return [[z.ref], [z.mul(dx)]];
-	},
-	[Primitive.Log]([x], [dx]) {
-		return [[log$1(x.ref)], [reciprocal$1(x).mul(dx)]];
-	},
-	[Primitive.Erf]([x], [dx]) {
-		const coeff = 2 / Math.sqrt(Math.PI);
-		const expTerm = exp$1(neg(x.ref.mul(x.ref)));
-		return [[erf$1(x)], [expTerm.mul(coeff).mul(dx)]];
-	},
-	[Primitive.Erfc]([x], [dx]) {
-		const coeff = -2 / Math.sqrt(Math.PI);
-		const expTerm = exp$1(neg(x.ref.mul(x.ref)));
-		return [[erfc$1(x)], [expTerm.mul(coeff).mul(dx)]];
-	},
-	[Primitive.Sqrt]([x], [dx]) {
-		const z = sqrt$1(x);
-		return [[z.ref], [reciprocal$1(z.mul(2)).mul(dx)]];
-	},
-	[Primitive.Min]([x, y], [dx, dy]) {
-		return [[min$1(x.ref, y.ref)], [where$1(less$1(y, x), dy, dx)]];
-	},
-	[Primitive.Max]([x, y], [dx, dy]) {
-		return [[max$1(x.ref, y.ref)], [where$1(less$1(x, y), dy, dx)]];
-	},
-	[Primitive.Reduce]([x], [dx], { op, axis }) {
-		if (op === AluOp.Add) return [[reduce(x, op, axis)], [reduce(dx, op, axis)]];
-		else if (op === AluOp.Mul) {
-			const primal = reduce(x.ref, op, axis);
-			const tangent = broadcast(primal.ref, x.shape, axis).mul(reciprocal$1(x)).mul(dx).sum(axis);
-			return [[primal], [tangent]];
-		} else if (op === AluOp.Min || op === AluOp.Max) {
-			const primal = reduce(x.ref, op, axis);
-			const notMin = notEqual$1(x, broadcast(primal.ref, x.shape, axis));
-			const minCount = where$1(notMin.ref, 0, 1).sum(axis);
-			const tangent = where$1(notMin, 0, dx).sum(axis).div(minCount);
-			return [[primal], [tangent]];
-		} else throw new Error(`JVP rule not implemented for reduce op: ${op}`);
-	},
-	[Primitive.Pool]: linearTangentsJvp(Primitive.Pool),
-	[Primitive.PoolTranspose]: linearTangentsJvp(Primitive.PoolTranspose),
-	[Primitive.Dot]: bilinearTangentsJvp(Primitive.Dot),
-	[Primitive.Conv]: bilinearTangentsJvp(Primitive.Conv),
-	[Primitive.Compare]: zeroTangentsJvp(Primitive.Compare),
-	[Primitive.Where]([cond, x, y], [dcond, dx, dy]) {
-		dcond.dispose();
-		return [[where$1(cond.ref, x, y)], [where$1(cond, dx, dy)]];
-	},
-	[Primitive.Transpose]: linearTangentsJvp(Primitive.Transpose),
-	[Primitive.Broadcast]: linearTangentsJvp(Primitive.Broadcast),
-	[Primitive.Reshape]: linearTangentsJvp(Primitive.Reshape),
-	[Primitive.Flip]: linearTangentsJvp(Primitive.Flip),
-	[Primitive.Shrink]: linearTangentsJvp(Primitive.Shrink),
-	[Primitive.Pad]: linearTangentsJvp(Primitive.Pad),
-	[Primitive.Gather]([x, ...indices], [dx, ..._], { axis, outDim }) {
-		const indicesRef = indices.map((t) => t.ref);
-		return [[gather(x, indices, axis, outDim)], [gather(dx, indicesRef, axis, outDim)]];
-	},
-	[Primitive.JitCall](primals, tangents, { name, jaxpr }) {
-		const { newJaxpr, newConsts } = jvpJaxpr(jaxpr);
-		const outs = bind(Primitive.JitCall, [
-			...newConsts.map((c) => c.ref),
-			...primals,
-			...tangents
-		], {
-			name: `${name}_jvp`,
-			jaxpr: newJaxpr,
-			numConsts: newConsts.length
-		});
-		const n = outs.length / 2;
-		if (!Number.isInteger(n)) throw new Error("internal: JVP Jaxpr output length is not even");
-		const [primalsOut, tangentsOut] = [outs.slice(0, n), outs.slice(n)];
-		return [primalsOut, tangentsOut];
-	}
-};
-const jvpJaxprCache = /* @__PURE__ */ new Map();
-function jvpJaxpr(jaxpr) {
-	if (jvpJaxprCache.has(jaxpr)) return jvpJaxprCache.get(jaxpr);
-	const inAvals = jaxpr.inBinders.map((v) => v.aval);
-	const { jaxpr: newJaxpr, consts: newConsts } = makeJaxpr$1((primals, tangents) => jvpFlat(jaxprAsFun(jaxpr), primals, tangents))(inAvals, inAvals);
-	const result = {
-		newJaxpr,
-		newConsts
-	};
-	jvpJaxprCache.set(jaxpr, result);
-	return result;
-}
-function jvpFlat(f, primals, tangents) {
-	try {
-		var _usingCtx$1 = _usingCtx();
-		const main = _usingCtx$1.u(newMain(JVPTrace));
-		const trace$1 = new JVPTrace(main);
-		const tracersIn = zip(primals, tangents).map(([x, t]) => new JVPTracer(trace$1, pureArray(x), pureArray(t)));
-		const outs = f(...tracersIn);
-		const tracersOut = outs.map((out) => fullRaise(trace$1, out));
-		return unzip2(tracersOut.map((t) => [t.primal, t.tangent]));
-	} catch (_) {
-		_usingCtx$1.e = _;
-	} finally {
-		_usingCtx$1.d();
-	}
-}
-function jvp$1(f, primals, tangents) {
-	const [primalsFlat, inTree] = flatten(primals);
-	const [tangentsFlat, inTree2] = flatten(tangents);
-	if (!inTree.equals(inTree2)) throw new TreeMismatchError("jvp", inTree, inTree2);
-	const [flatFun, outTree] = flattenFun(f, inTree);
-	const [primalsOutFlat, tangentsOutFlat] = jvpFlat(flatFun, primalsFlat, tangentsFlat);
-	if (outTree.value === void 0) throw new Error("outTree was not set in jvp");
-	const primalsOut = unflatten(outTree.value, primalsOutFlat);
-	const tangentsOut = unflatten(outTree.value, tangentsOutFlat);
-	return [primalsOut, tangentsOut];
-}
-
-//#endregion
-//#region src/frontend/vmap.ts
-function mappedAval(batchDim, aval) {
-	const shape$1 = [...aval.shape];
-	shape$1.splice(batchDim, 1);
-	return new ShapedArray(shape$1, aval.dtype, aval.weakType);
-}
-/** Move one axis to a different index. */
-function moveaxis(x, src, dst) {
-	const t = pureArray(x);
-	src = checkAxis(src, t.ndim);
-	dst = checkAxis(dst, t.ndim);
-	if (src === dst) return t;
-	const perm = range(t.ndim);
-	perm.splice(src, 1);
-	perm.splice(dst, 0, src);
-	return transpose$1(t, perm);
-}
-function moveBatchAxis(axisSize, src, dst, x) {
-	if (src === null) {
-		const targetShape = [...x.shape];
-		targetShape.splice(dst, 0, axisSize);
-		return broadcast(x, targetShape, [dst]);
-	} else if (src === dst) return x;
-	else return moveaxis(x, src, dst);
-}
-var BatchTracer = class extends Tracer {
-	constructor(trace$1, val, batchDim) {
-		super(trace$1);
-		this.val = val;
-		this.batchDim = batchDim;
-	}
-	get aval() {
-		if (this.batchDim === null) return this.val.aval;
-		else return mappedAval(this.batchDim, this.val.aval);
-	}
-	toString() {
-		return `BatchTracer(${this.val.toString()}, ${this.batchDim})`;
-	}
-	get ref() {
-		this.val.ref;
-		return this;
-	}
-	dispose() {
-		this.val.dispose();
-	}
-	fullLower() {
-		if (this.batchDim === null) return this.val.fullLower();
-		else return this;
-	}
-};
-var BatchTrace = class extends Trace {
-	pure(val) {
-		return this.lift(pureArray(val));
-	}
-	lift(val) {
-		return new BatchTracer(this, val, null);
-	}
-	processPrimitive(primitive, tracers, params) {
-		const [valsIn, bdimsIn] = unzip2(tracers.map((t) => [t.val, t.batchDim]));
-		const vmapRule = vmapRules[primitive];
-		if (vmapRule === void 0) throw new Error(`No vmap rule for: ${primitive}`);
-		if (bdimsIn.every((d) => d === null)) {
-			const valOuts$1 = bind(primitive, valsIn, params);
-			return valOuts$1.map((x) => new BatchTracer(this, x, null));
-		}
-		const [valOuts, bdimOuts] = vmapRule(this.axisSize, valsIn, bdimsIn, params);
-		return zip(valOuts, bdimOuts).map(([x, bd]) => new BatchTracer(this, x, bd));
-	}
-	get axisSize() {
-		return this.main.globalData;
-	}
-};
-/**
-* Process a primitive with built-in broadcasting.
-*
-* Reference: https://github.com/jax-ml/jax/blob/jax-v0.8.1/jax/_src/interpreters/batching.py#L1029
-*/
-function broadcastBatcher(op) {
-	return (axisSize, args, dims) => {
-		if (args.length === 0) throw new Error("Empty list in broadcastBatcher");
-		const nd = Math.max(...args.map((x, i) => ndim$1(x) + (dims[i] === null ? 1 : 0)));
-		const firstIdx = dims.findIndex((d) => d !== null);
-		const firstBdim = dims[firstIdx] - args[firstIdx].ndim;
-		if (zip(args, dims).every(([x, d]) => d === null && ndim$1(x) < -firstBdim || d !== null && d - x.ndim === firstBdim)) return [[op(...args)], [nd + firstBdim]];
-		args = args.map((x, i) => {
-			if (dims[i] === null) return x;
-			x = moveBatchAxis(axisSize, dims[i], 0, x);
-			if (x.ndim < nd) x = x.reshape([
-				x.shape[0],
-				...rep(nd - x.ndim, 1),
-				...x.shape.slice(1)
-			]);
-			return x;
-		});
-		return [[op(...args)], [0]];
-	};
-}
-function unopBatcher(op) {
-	return (axisSize, [x], [xBdim], params) => {
-		return [[op(x, params)], [xBdim]];
+function unopBatcher(op) {
+	return (axisSize, [x], [xBdim], params) => {
+		return [[op(x, params)], [xBdim]];
 	};
 }
 const vmapRules = {
@@ -3430,6 +3529,8 @@ const vmapRules = {
 	[Primitive.Mul]: broadcastBatcher(mul),
 	[Primitive.Idiv]: broadcastBatcher(idiv),
 	[Primitive.Mod]: broadcastBatcher(mod),
+	[Primitive.Min]: broadcastBatcher(min$1),
+	[Primitive.Max]: broadcastBatcher(max$1),
 	[Primitive.Neg]: unopBatcher(neg),
 	[Primitive.Reciprocal]: unopBatcher(reciprocal$1),
 	[Primitive.Floor]: unopBatcher(floor$1),
@@ -3446,8 +3547,6 @@ const vmapRules = {
 	[Primitive.Erf]: unopBatcher(erf$1),
 	[Primitive.Erfc]: unopBatcher(erfc$1),
 	[Primitive.Sqrt]: unopBatcher(sqrt$1),
-	[Primitive.Min]: broadcastBatcher(min$1),
-	[Primitive.Max]: broadcastBatcher(max$1),
 	[Primitive.Reduce](axisSize, [x], [xBdim], { op, axis }) {
 		assertNonNull(xBdim);
 		const newAxis = axis.map((ax) => ax + (xBdim <= ax ? 1 : 0));
@@ -3460,10 +3559,49 @@ const vmapRules = {
 		const z = dot$2(x, y);
 		return [[z], [z.ndim - 1]];
 	},
+	[Primitive.Conv](axisSize, [x, y], [xBdim, yBdim], params) {
+		x = moveBatchAxis(axisSize, xBdim, 0, x);
+		y = moveBatchAxis(axisSize, yBdim, 0, y);
+		const z = conv$1(x, y, {
+			...params,
+			vmapDims: params.vmapDims + 1
+		});
+		return [[z], [0]];
+	},
 	[Primitive.Compare](axisSize, args, dims, { op }) {
 		return broadcastBatcher((x, y) => compare(x, y, op))(axisSize, args, dims, {});
 	},
-	[Primitive.Where]: broadcastBatcher(where$1),
+	[Primitive.Where]: broadcastBatcher(where$1),
+	[Primitive.Gather](axisSize, [x, ...indices], [xBdim, ...indicesBdim], { axis, outDim }) {
+		if (indicesBdim.every((d) => d === null)) {
+			assertNonNull(xBdim);
+			const newAxis = axis.map((ax) => ax + (xBdim <= ax ? 1 : 0));
+			let newBdim = xBdim - axis.filter((ax) => ax < xBdim).length;
+			let newOutDim = outDim;
+			if (newOutDim < newBdim) newBdim += axis.length;
+			else newOutDim += 1;
+			return [[gather(x, indices, newAxis, newOutDim)], [newBdim]];
+		}
+		const nd = Math.max(...indices.map((m, i) => ndim$1(m) + (indicesBdim[i] === null ? 1 : 0)));
+		indices = indices.map((m, i) => {
+			if (indicesBdim[i] === null) return m;
+			m = moveBatchAxis(axisSize, indicesBdim[i], 0, m);
+			if (m.ndim < nd) m = m.reshape([
+				m.shape[0],
+				...rep(nd - m.ndim, 1),
+				...m.shape.slice(1)
+			]);
+			return m;
+		});
+		if (xBdim === null) return [[gather(x, indices, axis, outDim)], [outDim]];
+		else {
+			x = moveBatchAxis(axisSize, xBdim, 0, x);
+			const newAxis = [0, ...axis.map((ax) => ax + 1)];
+			const extraBatchIndex = arange(axisSize).reshape([-1, ...rep(nd - 1, 1)]);
+			indices.splice(0, 0, extraBatchIndex);
+			return [[gather(x, indices, newAxis, outDim)], [outDim]];
+		}
+	},
 	[Primitive.Transpose](axisSize, [x], [xBdim], { perm }) {
 		assertNonNull(xBdim);
 		const newPerm = perm.map((p) => p + (xBdim <= p ? 1 : 0));
@@ -3495,42 +3633,53 @@ const vmapRules = {
 		const newWidth = width.toSpliced(xBdim, 0, [0, 0]);
 		return [[pad$1(x, newWidth)], [xBdim]];
 	},
-	[Primitive.Gather](axisSize, [x, ...indices], [xBdim, ...indicesBdim], { axis, outDim }) {
-		if (indicesBdim.every((d) => d === null)) {
-			assertNonNull(xBdim);
-			const newAxis = axis.map((ax) => ax + (xBdim <= ax ? 1 : 0));
-			let newBdim = xBdim - axis.filter((ax) => ax < xBdim).length;
-			let newOutDim = outDim;
-			if (newOutDim < newBdim) newBdim += axis.length;
-			else newOutDim += 1;
-			return [[gather(x, indices, newAxis, newOutDim)], [newBdim]];
-		}
-		const nd = Math.max(...indices.map((m, i) => ndim$1(m) + (indicesBdim[i] === null ? 1 : 0)));
-		indices = indices.map((m, i) => {
-			if (indicesBdim[i] === null) return m;
-			m = moveBatchAxis(axisSize, indicesBdim[i], 0, m);
-			if (m.ndim < nd) m = m.reshape([
-				m.shape[0],
-				...rep(nd - m.ndim, 1),
-				...m.shape.slice(1)
+	[Primitive.Sort](axisSize, [x], [xBdim]) {
+		assertNonNull(xBdim);
+		if (xBdim !== x.ndim - 1) return [[sort$1(x)], [xBdim]];
+		x = moveBatchAxis(axisSize, xBdim, 0, x);
+		return [[sort$1(x)], [0]];
+	},
+	[Primitive.Argsort](axisSize, [x], [xBdim]) {
+		assertNonNull(xBdim);
+		if (xBdim !== x.ndim - 1) return [argsort$1(x), [xBdim, xBdim]];
+		x = moveBatchAxis(axisSize, xBdim, 0, x);
+		return [argsort$1(x), [0, 0]];
+	},
+	[Primitive.TriangularSolve](axisSize, [a, b], [aBdim, bBdim], { unitDiagonal }) {
+		if (aBdim === null) {
+			b = moveBatchAxis(axisSize, bBdim, -3, b);
+			const [s, m, n] = b.shape.slice(-3);
+			b = b.reshape([
+				...b.shape.slice(0, -3),
+				s * m,
+				n
 			]);
-			return m;
-		});
-		if (xBdim === null) return [[gather(x, indices, axis, outDim)], [outDim]];
-		else {
-			x = moveBatchAxis(axisSize, xBdim, 0, x);
-			const newAxis = [0, ...axis.map((ax) => ax + 1)];
-			const extraBatchIndex = arange(axisSize).reshape([-1, ...rep(nd - 1, 1)]);
-			indices.splice(0, 0, extraBatchIndex);
-			return [[gather(x, indices, newAxis, outDim)], [outDim]];
+			let x$1 = bind1(Primitive.TriangularSolve, [a, b], { unitDiagonal });
+			x$1 = x$1.reshape([
+				...b.shape.slice(0, -2),
+				s,
+				m,
+				n
+			]);
+			return [[x$1], [x$1.ndim - 3]];
 		}
+		a = moveBatchAxis(axisSize, aBdim, 0, a);
+		b = moveBatchAxis(axisSize, bBdim, 0, b);
+		const x = bind1(Primitive.TriangularSolve, [a, b], { unitDiagonal });
+		return [[x], [0]];
+	},
+	[Primitive.Cholesky](axisSize, [x], [xBdim]) {
+		assertNonNull(xBdim);
+		if (xBdim < x.ndim - 2) return [[cholesky$2(x)], [xBdim]];
+		x = moveBatchAxis(axisSize, xBdim, 0, x);
+		return [[cholesky$2(x)], [0]];
 	},
-	[Primitive.JitCall](axisSize, args, dims, { name, jaxpr }) {
-		const { newJaxpr, newConsts } = vmapJaxpr(jaxpr, axisSize, dims);
-		const outs = bind(Primitive.JitCall, [...newConsts.map((c) => c.ref), ...args], {
+	[Primitive.Jit](axisSize, args, dims, { name, jaxpr }) {
+		const newJaxpr = vmapJaxpr(jaxpr, axisSize, dims);
+		const outs = bind(Primitive.Jit, [...newJaxpr.consts.map((c) => c.ref), ...args], {
 			name: `${name}_vmap`,
-			jaxpr: newJaxpr,
-			numConsts: newConsts.length
+			jaxpr: newJaxpr.jaxpr,
+			numConsts: newJaxpr.consts.length
 		});
 		return [outs, rep(outs.length, 0)];
 	}
@@ -3546,14 +3695,10 @@ function vmapJaxpr(jaxpr, axisSize, dims) {
 		shape$1.splice(dims[i], 0, axisSize);
 		return new ShapedArray(shape$1, v.aval.dtype, v.aval.weakType);
 	});
-	const { jaxpr: newJaxpr, consts: newConsts } = makeJaxpr$1((args) => vmapFlat(jaxprAsFun(jaxpr), dims, args))(inAvals);
-	const result = {
-		newJaxpr,
-		newConsts
-	};
+	const { jaxpr: newJaxpr } = makeJaxpr$1((args) => vmapFlat(jaxprAsFun(jaxpr), dims, args))(inAvals);
 	if (!vmapJaxprCache.has(jaxpr)) vmapJaxprCache.set(jaxpr, /* @__PURE__ */ new Map());
-	vmapJaxprCache.get(jaxpr).set(cacheKey, result);
-	return result;
+	vmapJaxprCache.get(jaxpr).set(cacheKey, newJaxpr);
+	return newJaxpr;
 }
 function vmapFlat(f, inAxes, args) {
 	let axisSize = void 0;
@@ -3608,6 +3753,260 @@ function jacfwd$1(f) {
 	};
 }
 
+//#endregion
+//#region src/frontend/jvp.ts
+var JVPTracer = class extends Tracer {
+	constructor(trace$1, primal, tangent) {
+		super(trace$1);
+		this.primal = primal;
+		this.tangent = tangent;
+	}
+	get aval() {
+		return this.primal.aval;
+	}
+	toString() {
+		return `JVPTracer(${this.primal.toString()}, ${this.tangent.toString()})`;
+	}
+	get ref() {
+		this.primal.ref, this.tangent.ref;
+		return this;
+	}
+	dispose() {
+		this.primal.dispose();
+		this.tangent.dispose();
+	}
+};
+var JVPTrace = class extends Trace {
+	pure(val) {
+		return this.lift(pureArray(val));
+	}
+	lift(val) {
+		return new JVPTracer(this, val, zerosLike$1(val.ref));
+	}
+	processPrimitive(primitive, tracers, params) {
+		const [primalsIn, tangentsIn] = unzip2(tracers.map((x) => [x.primal, x.tangent]));
+		const jvpRule = jvpRules[primitive];
+		if (jvpRule === void 0) throw new Error(`No JVP rule for: ${primitive}`);
+		const [primalsOut, tangentsOut] = jvpRule(primalsIn, tangentsIn, params);
+		return zip(primalsOut, tangentsOut).map(([x, t]) => new JVPTracer(this, x, t));
+	}
+};
+/** Rule that applies the same operation to primals and tangents. */
+function linearTangentsJvp(primitive) {
+	return (primals, tangents, params) => {
+		const ys = bind(primitive, primals, params);
+		const dys = bind(primitive, tangents, params);
+		return [ys, dys];
+	};
+}
+/** Rule for product of gradients in bilinear operations. */
+function bilinearTangentsJvp(primitive) {
+	return ([x, y], [dx, dy], params) => {
+		const primal = bind1(primitive, [x.ref, y.ref], params);
+		const tangent = bind1(primitive, [x, dy], params).add(bind1(primitive, [dx, y], params));
+		return [[primal], [tangent]];
+	};
+}
+/** Rule that zeros out any tangents. */
+function zeroTangentsJvp(primitive) {
+	return (primals, tangents, params) => {
+		for (const t of tangents) t.dispose();
+		const ys = bind(primitive, primals, params);
+		return [ys, ys.map((y) => zerosLike$1(y.ref))];
+	};
+}
+/** Compute `a @ b.T`, batched to last two axes. */
+function batchMatmulT(a, b) {
+	return dot$2(a.reshape(a.shape.toSpliced(-1, 0, 1)), b.reshape(b.shape.toSpliced(-2, 0, 1)));
+}
+/** Batch matrix transpose. */
+function mT(a) {
+	return moveaxis(a, -2, -1);
+}
+const jvpRules = {
+	[Primitive.Add]: linearTangentsJvp(Primitive.Add),
+	[Primitive.Mul]: bilinearTangentsJvp(Primitive.Mul),
+	[Primitive.Idiv]: zeroTangentsJvp(Primitive.Idiv),
+	[Primitive.Mod]([x, y], [dx, dy]) {
+		if (!isFloatDtype(x.dtype) && !isFloatDtype(y.dtype)) {
+			dx.dispose();
+			dy.dispose();
+			return [[x.ref, y.ref], [zerosLike$1(x), zerosLike$1(y)]];
+		}
+		const q = idiv(x.ref, y.ref);
+		return [[mod(x, y)], [dx.sub(dy.mul(q))]];
+	},
+	[Primitive.Min]([x, y], [dx, dy]) {
+		return [[min$1(x.ref, y.ref)], [where$1(less$1(y, x), dy, dx)]];
+	},
+	[Primitive.Max]([x, y], [dx, dy]) {
+		return [[max$1(x.ref, y.ref)], [where$1(less$1(x, y), dy, dx)]];
+	},
+	[Primitive.Neg]: linearTangentsJvp(Primitive.Neg),
+	[Primitive.Reciprocal]([x], [dx]) {
+		const xRecip = reciprocal$1(x.ref);
+		return [[xRecip.ref], [neg(xRecip.ref.mul(xRecip)).mul(dx)]];
+	},
+	[Primitive.Floor]: zeroTangentsJvp(Primitive.Floor),
+	[Primitive.Ceil]: zeroTangentsJvp(Primitive.Ceil),
+	[Primitive.StopGradient]: zeroTangentsJvp(Primitive.StopGradient),
+	[Primitive.Cast]([x], [dx], { dtype }) {
+		if (x.dtype === dtype) return [[x], [dx]];
+		if (isFloatDtype(dtype) && isFloatDtype(x.dtype)) return [[cast(x, dtype)], [cast(dx, dtype)]];
+		else {
+			dx.dispose();
+			return [[cast(x.ref, dtype)], [zerosLike$1(x)]];
+		}
+	},
+	[Primitive.Bitcast]([x], [dx], { dtype }) {
+		if (x.dtype === dtype) return [[x], [dx]];
+		dx.dispose();
+		return [[bitcast(x.ref, dtype)], [zerosLike$1(x)]];
+	},
+	[Primitive.Sin]([x], [dx]) {
+		return [[sin$1(x.ref)], [cos$1(x).mul(dx)]];
+	},
+	[Primitive.Cos]([x], [dx]) {
+		return [[cos$1(x.ref)], [neg(sin$1(x)).mul(dx)]];
+	},
+	[Primitive.Asin]([x], [dx]) {
+		const denom = sqrt$1(reciprocal$1(cast(1, x.dtype).sub(x.ref.mul(x.ref))));
+		return [[asin$1(x)], [denom.mul(dx)]];
+	},
+	[Primitive.Atan]([x], [dx]) {
+		const denom = cast(1, x.dtype).add(x.ref.mul(x.ref));
+		return [[atan$1(x)], [dx.div(denom)]];
+	},
+	[Primitive.Exp]([x], [dx]) {
+		const z = exp$1(x);
+		return [[z.ref], [z.mul(dx)]];
+	},
+	[Primitive.Log]([x], [dx]) {
+		return [[log$1(x.ref)], [reciprocal$1(x).mul(dx)]];
+	},
+	[Primitive.Erf]([x], [dx]) {
+		const coeff = 2 / Math.sqrt(Math.PI);
+		const expTerm = exp$1(neg(x.ref.mul(x.ref)));
+		return [[erf$1(x)], [expTerm.mul(coeff).mul(dx)]];
+	},
+	[Primitive.Erfc]([x], [dx]) {
+		const coeff = -2 / Math.sqrt(Math.PI);
+		const expTerm = exp$1(neg(x.ref.mul(x.ref)));
+		return [[erfc$1(x)], [expTerm.mul(coeff).mul(dx)]];
+	},
+	[Primitive.Sqrt]([x], [dx]) {
+		const z = sqrt$1(x);
+		return [[z.ref], [reciprocal$1(z.mul(2)).mul(dx)]];
+	},
+	[Primitive.Reduce]([x], [dx], { op, axis }) {
+		if (op === AluOp.Add) return [[reduce(x, op, axis)], [reduce(dx, op, axis)]];
+		else if (op === AluOp.Mul) {
+			const primal = reduce(x.ref, op, axis);
+			const tangent = broadcast(primal.ref, x.shape, axis).mul(reciprocal$1(x)).mul(dx).sum(axis);
+			return [[primal], [tangent]];
+		} else if (op === AluOp.Min || op === AluOp.Max) {
+			const primal = reduce(x.ref, op, axis);
+			const notMin = notEqual$1(x, broadcast(primal.ref, x.shape, axis));
+			const minCount = where$1(notMin.ref, 0, 1).sum(axis);
+			const tangent = where$1(notMin, 0, dx).sum(axis).div(minCount);
+			return [[primal], [tangent]];
+		} else throw new Error(`JVP rule not implemented for reduce op: ${op}`);
+	},
+	[Primitive.Pool]: linearTangentsJvp(Primitive.Pool),
+	[Primitive.PoolTranspose]: linearTangentsJvp(Primitive.PoolTranspose),
+	[Primitive.Dot]: bilinearTangentsJvp(Primitive.Dot),
+	[Primitive.Conv]: bilinearTangentsJvp(Primitive.Conv),
+	[Primitive.Compare]: zeroTangentsJvp(Primitive.Compare),
+	[Primitive.Where]([cond, x, y], [dcond, dx, dy]) {
+		dcond.dispose();
+		return [[where$1(cond.ref, x, y)], [where$1(cond, dx, dy)]];
+	},
+	[Primitive.RandomBits]: zeroTangentsJvp(Primitive.RandomBits),
+	[Primitive.Gather]([x, ...indices], [dx, ..._], { axis, outDim }) {
+		const indicesRef = indices.map((t) => t.ref);
+		return [[gather(x, indices, axis, outDim)], [gather(dx, indicesRef, axis, outDim)]];
+	},
+	[Primitive.Transpose]: linearTangentsJvp(Primitive.Transpose),
+	[Primitive.Broadcast]: linearTangentsJvp(Primitive.Broadcast),
+	[Primitive.Reshape]: linearTangentsJvp(Primitive.Reshape),
+	[Primitive.Flip]: linearTangentsJvp(Primitive.Flip),
+	[Primitive.Shrink]: linearTangentsJvp(Primitive.Shrink),
+	[Primitive.Pad]: linearTangentsJvp(Primitive.Pad),
+	[Primitive.Sort]([x], [dx]) {
+		const [y, idx] = argsort$1(x);
+		return [[y], [gather(dx, [idx], [-1], -1)]];
+	},
+	[Primitive.Argsort]([x], [dx]) {
+		const [y, idx] = argsort$1(x);
+		return [[y, idx.ref], [gather(dx, [idx.ref], [-1], -1), zerosLike$1(idx)]];
+	},
+	[Primitive.TriangularSolve]([a, b], [da, db], { unitDiagonal }) {
+		const x = triangularSolve$1(a.ref, b, { unitDiagonal });
+		const dax = batchMatmulT(da, x.ref);
+		const rhsT = db.sub(mT(dax));
+		const dx = triangularSolve$1(a, rhsT, { unitDiagonal });
+		return [[x], [dx]];
+	},
+	[Primitive.Cholesky]([a], [da]) {
+		const L = cholesky$2(a.ref);
+		da = da.ref.add(mT(da)).mul(.5);
+		const W = triangularSolve$1(L.ref, da, { lower: true });
+		const ST = triangularSolve$1(L.ref, mT(W), { lower: true });
+		const dL = batchMatmulT(L.ref, triu(ST.ref, 1).add(triu(ST)).mul(.5));
+		return [[L], [dL]];
+	},
+	[Primitive.Jit](primals, tangents, { name, jaxpr }) {
+		const newJaxpr = jvpJaxpr(jaxpr);
+		const outs = bind(Primitive.Jit, [
+			...newJaxpr.consts.map((c) => c.ref),
+			...primals,
+			...tangents
+		], {
+			name: `${name}_jvp`,
+			jaxpr: newJaxpr.jaxpr,
+			numConsts: newJaxpr.consts.length
+		});
+		const n = outs.length / 2;
+		if (!Number.isInteger(n)) throw new Error("internal: JVP Jaxpr output length is not even");
+		const [primalsOut, tangentsOut] = [outs.slice(0, n), outs.slice(n)];
+		return [primalsOut, tangentsOut];
+	}
+};
+const jvpJaxprCache = /* @__PURE__ */ new Map();
+function jvpJaxpr(jaxpr) {
+	if (jvpJaxprCache.has(jaxpr)) return jvpJaxprCache.get(jaxpr);
+	const inAvals = jaxpr.inBinders.map((v) => v.aval);
+	const { jaxpr: newJaxpr } = makeJaxpr$1((primals, tangents) => jvpFlat(jaxprAsFun(jaxpr), primals, tangents))(inAvals, inAvals);
+	jvpJaxprCache.set(jaxpr, newJaxpr);
+	return newJaxpr;
+}
+function jvpFlat(f, primals, tangents) {
+	try {
+		var _usingCtx$1 = _usingCtx();
+		const main = _usingCtx$1.u(newMain(JVPTrace));
+		const trace$1 = new JVPTrace(main);
+		const tracersIn = zip(primals, tangents).map(([x, t]) => new JVPTracer(trace$1, pureArray(x), pureArray(t)));
+		const outs = f(...tracersIn);
+		const tracersOut = outs.map((out) => fullRaise(trace$1, out));
+		return unzip2(tracersOut.map((t) => [t.primal, t.tangent]));
+	} catch (_) {
+		_usingCtx$1.e = _;
+	} finally {
+		_usingCtx$1.d();
+	}
+}
+function jvp$1(f, primals, tangents) {
+	const [primalsFlat, inTree] = flatten(primals);
+	const [tangentsFlat, inTree2] = flatten(tangents);
+	if (!inTree.equals(inTree2)) throw new TreeMismatchError("jvp", inTree, inTree2);
+	const [flatFun, outTree] = flattenFun(f, inTree);
+	const [primalsOutFlat, tangentsOutFlat] = jvpFlat(flatFun, primalsFlat, tangentsFlat);
+	if (outTree.value === void 0) throw new Error("outTree was not set in jvp");
+	const primalsOut = unflatten(outTree.value, primalsOutFlat);
+	const tangentsOut = unflatten(outTree.value, tangentsOutFlat);
+	return [primalsOut, tangentsOut];
+}
+
 //#endregion
 //#region src/frontend/linearize.ts
 /** Array value that can either be known or unknown. */
@@ -3638,11 +4037,10 @@ function partialEvalFlat(f, pvalsIn) {
 	const tracersOut = outs.map((out) => fullRaise(trace$1, out));
 	const pvalsOut = tracersOut.map((t) => t.pval);
 	const unknownTracersOut = tracersOut.filter((t) => !t.pval.isKnown);
-	const { jaxpr, consts } = partialEvalGraphToJaxpr(unknownTracersIn, unknownTracersOut);
+	const jaxpr = partialEvalGraphToJaxpr(unknownTracersIn, unknownTracersOut);
 	return {
 		jaxpr,
-		pvalsOut,
-		consts
+		pvalsOut
 	};
 }
 /**
@@ -3659,22 +4057,19 @@ function linearizeFlatUtil(f, primalsIn) {
 		const [primalsOut$1, tangentsOut] = jvp$1(f, x.slice(0, k), x.slice(k, 2 * k));
 		return [...primalsOut$1, ...tangentsOut];
 	};
-	const { jaxpr, pvalsOut, consts } = partialEvalFlat(fJvp, pvalsIn);
+	const { jaxpr, pvalsOut } = partialEvalFlat(fJvp, pvalsIn);
 	const primalPvals = pvalsOut.slice(0, pvalsOut.length / 2);
 	if (!primalPvals.every((pval) => pval.isKnown)) throw new Error("Not all primal values are known after partial evaluation");
 	const primalsOut = primalPvals.map((pval) => pval.val);
 	return {
 		primalsOut,
-		jaxpr,
-		consts
+		jaxpr
 	};
 }
 function linearizeFlat(f, primalsIn) {
-	const { primalsOut, jaxpr, consts } = linearizeFlatUtil(f, primalsIn);
-	const fLin = (...tangents) => evalJaxpr(jaxpr, [...consts.map((c) => c.ref), ...tangents]);
-	const dispose$1 = () => {
-		for (const c of consts) c.dispose();
-	};
+	const { primalsOut, jaxpr } = linearizeFlatUtil(f, primalsIn);
+	const fLin = (...tangents) => evalJaxpr(jaxpr.jaxpr, [...jaxpr.consts.map((c) => c.ref), ...tangents]);
+	const dispose$1 = () => jaxpr.dispose();
 	return [
 		primalsOut,
 		fLin,
@@ -3758,7 +4153,7 @@ var PartialEvalTrace = class extends Trace {
 	}
 	processPrimitive(primitive, tracers, params) {
 		if (tracers.every((t) => t.pval.isKnown)) return bind(primitive, tracers.map((t) => t.fullLower()), params);
-		if (primitive === Primitive.JitCall) {
+		if (primitive === Primitive.Jit) {
 			const { name, jaxpr, numConsts } = params;
 			return this.#partialEvalJaxpr(name, jaxpr, numConsts, tracers);
 		}
@@ -3784,14 +4179,14 @@ var PartialEvalTrace = class extends Trace {
 	* Evaluate a Jaxpr on a set of PartialEvalTracers, computing as many known
 	* values as possible (with JIT) and forwarding the unknown ones.
 	*
-	* Used when encountering a JitCall rule during the trace.
+	* Used when encountering a Jit rule during the trace.
 	*/
 	#partialEvalJaxpr(name, jaxpr, numConsts, tracers) {
 		jaxpr = jaxpr.flatten();
 		const inUnknowns = tracers.map((t) => !t.pval.isKnown);
 		const { jaxpr1, jaxpr2, outUnknowns, numRes } = partialEvalJaxpr(jaxpr, inUnknowns);
 		const [knownTracers, unknownTracers] = partitionList(inUnknowns, tracers);
-		const outs1Res = bind(Primitive.JitCall, knownTracers.map((t) => t.ref.fullLower()), {
+		const outs1Res = bind(Primitive.Jit, knownTracers.map((t) => t.ref.fullLower()), {
 			name: `${name}_peval`,
 			jaxpr: jaxpr1,
 			numConsts: 0
@@ -3801,7 +4196,7 @@ var PartialEvalTrace = class extends Trace {
 		const resTracers = res.map((x) => this.instantiateConst(fullRaise(this, x)));
 		const recipe = {
 			type: "JaxprEqn",
-			prim: Primitive.JitCall,
+			prim: Primitive.Jit,
 			tracersIn: resTracers.concat(unknownTracers),
 			params: {
 				name: `${name}_resid`,
@@ -3830,7 +4225,7 @@ function partialEvalJaxpr(jaxpr, inUnknowns, instantiate) {
 	const eqns1 = [];
 	const eqns2 = [];
 	for (const eqn of jaxpr.eqns) {
-		if (eqn.primitive === Primitive.JitCall) throw new TypeError("partialEvalJaxpr requires flattened Jaxpr");
+		if (eqn.primitive === Primitive.Jit) throw new TypeError("partialEvalJaxpr requires flattened Jaxpr");
 		const hasUnknowns = eqn.inputs.some((x) => x instanceof Var && !knownVars.has(x));
 		if (hasUnknowns) {
 			for (const x of eqn.inputs) if (x instanceof Var && knownVars.has(x)) residuals.add(x);
@@ -3904,11 +4299,8 @@ function partialEvalGraphToJaxpr(tracersIn, tracersOut) {
 	for (const t of tracersIn) t.dispose();
 	for (const t of tracersOut) t.dispose();
 	jaxpr = jaxpr.simplify();
-	if (DEBUG >= 5) console.log("jaxpr from partial evaluation:\n" + jaxpr.toString());
-	return {
-		jaxpr,
-		consts
-	};
+	if (DEBUG >= 5) console.info("jaxpr from partial evaluation:\n" + jaxpr.toString());
+	return new ClosedJaxpr(jaxpr, consts);
 }
 /** Marker type for pullback, used by transpose rules. */
 var UndefPrimal = class {
@@ -4038,22 +4430,25 @@ const transposeRules = {
 	},
 	[Primitive.Conv]([ct], [lhs, rhs], params) {
 		if (lhs instanceof UndefPrimal === rhs instanceof UndefPrimal) throw new NonlinearError(Primitive.Conv);
+		const v = params.vmapDims;
 		const rev01 = [
-			1,
-			0,
-			...range(2, ct.ndim)
+			...range(v),
+			v + 1,
+			v,
+			...range(v + 2, ct.ndim)
 		];
 		if (lhs instanceof UndefPrimal) {
 			let kernel = rhs;
 			kernel = transpose$1(kernel, rev01);
-			kernel = flip$1(kernel, range(2, kernel.ndim));
+			kernel = flip$1(kernel, range(v + 2, kernel.ndim));
 			const result = conv$1(ct, kernel, {
+				vmapDims: v,
 				strides: params.lhsDilation,
 				padding: params.padding.map(([pl, _pr], i) => {
-					const dilatedKernel = (kernel.shape[i + 2] - 1) * params.rhsDilation[i] + 1;
-					const dilatedCt = (ct.shape[i + 2] - 1) * params.strides[i] + 1;
+					const dilatedKernel = (kernel.shape[i + v + 2] - 1) * params.rhsDilation[i] + 1;
+					const dilatedCt = (ct.shape[i + v + 2] - 1) * params.strides[i] + 1;
 					const padBefore = dilatedKernel - 1 - pl;
-					const dilatedLhs = (lhs.aval.shape[i + 2] - 1) * params.lhsDilation[i] + 1;
+					const dilatedLhs = (lhs.aval.shape[i + v + 2] - 1) * params.lhsDilation[i] + 1;
 					const padAfter = dilatedLhs + dilatedKernel - 1 - dilatedCt - padBefore;
 					return [padBefore, padAfter];
 				}),
@@ -4065,11 +4460,12 @@ const transposeRules = {
 			const newLhs = transpose$1(lhs, rev01);
 			const newRhs = transpose$1(ct, rev01);
 			let result = conv$1(newLhs, newRhs, {
+				vmapDims: v,
 				strides: params.rhsDilation,
 				padding: params.padding.map(([pl, _pr], i) => {
-					const dilatedLhs = (lhs.aval.shape[i + 2] - 1) * params.lhsDilation[i] + 1;
-					const dilatedKernel = (rhs.aval.shape[i + 2] - 1) * params.rhsDilation[i] + 1;
-					const dilatedCt = (ct.shape[i + 2] - 1) * params.strides[i] + 1;
+					const dilatedLhs = (lhs.aval.shape[i + v + 2] - 1) * params.lhsDilation[i] + 1;
+					const dilatedKernel = (rhs.aval.shape[i + v + 2] - 1) * params.rhsDilation[i] + 1;
+					const dilatedCt = (ct.shape[i + v + 2] - 1) * params.strides[i] + 1;
 					const padFromLhs = dilatedCt - dilatedLhs;
 					const padFromRhs = dilatedKernel - pl - 1;
 					return [pl, padFromLhs + padFromRhs];
@@ -4096,6 +4492,11 @@ const transposeRules = {
 		cond.dispose();
 		return cts;
 	},
+	[Primitive.Gather]([ct], [x, ...indices], { axis, outDim }) {
+		if (!(x instanceof UndefPrimal)) throw new NonlinearError(Primitive.Gather);
+		if (indices.some((i) => i instanceof UndefPrimal)) throw new NonlinearError(Primitive.Gather);
+		throw new Error("Gather transpose rule is not yet implemented, requires complex Scatter sum operation");
+	},
 	[Primitive.Transpose]([ct], [x], { perm }) {
 		if (!(x instanceof UndefPrimal)) throw new NonlinearError(Primitive.Transpose);
 		return [transpose$1(ct, invertPermutation(perm))];
@@ -4122,23 +4523,26 @@ const transposeRules = {
 		const slice = width.map(([s, _e], i) => [s, s + x.aval.shape[i]]);
 		return [shrink(ct, slice)];
 	},
-	[Primitive.Gather]([ct], [x, ...indices], { axis, outDim }) {
-		if (!(x instanceof UndefPrimal)) throw new NonlinearError(Primitive.Gather);
-		if (indices.some((i) => i instanceof UndefPrimal)) throw new NonlinearError(Primitive.Gather);
-		throw new Error("Gather transpose rule is not yet implemented, requires complex Scatter sum operation");
+	[Primitive.TriangularSolve]([ct], [a, b], { unitDiagonal }) {
+		if (a instanceof UndefPrimal || !(b instanceof UndefPrimal)) throw new NonlinearError(Primitive.TriangularSolve);
+		const ctB = triangularSolve$1(moveaxis(a, -2, -1), ct, {
+			lower: true,
+			unitDiagonal
+		});
+		return [null, ctB];
 	},
-	[Primitive.JitCall](cts, args, { name, jaxpr }) {
+	[Primitive.Jit](cts, args, { name, jaxpr }) {
 		const undefPrimals = args.map((x) => x instanceof UndefPrimal);
-		const { newJaxpr, newConsts } = transposeJaxpr(jaxpr, undefPrimals);
+		const newJaxpr = transposeJaxpr(jaxpr, undefPrimals);
 		const residuals = args.filter((x, i$1) => !undefPrimals[i$1]);
-		const outs = bind(Primitive.JitCall, [
-			...newConsts.map((c) => c.ref),
+		const outs = bind(Primitive.Jit, [
+			...newJaxpr.consts.map((c) => c.ref),
 			...residuals,
 			...cts
 		], {
 			name: `${name}_t`,
-			jaxpr: newJaxpr,
-			numConsts: newConsts.length
+			jaxpr: newJaxpr.jaxpr,
+			numConsts: newJaxpr.consts.length
 		});
 		let i = 0;
 		return undefPrimals.map((isUndef) => isUndef ? outs[i++] : null);
@@ -4151,31 +4555,25 @@ function transposeJaxpr(jaxpr, undefPrimals) {
 	if (prevResult) return prevResult;
 	const { inTypes, outTypes } = typecheckJaxpr(jaxpr);
 	const forwardInTypes = inTypes.filter((_, i) => !undefPrimals[i]);
-	const { jaxpr: newJaxpr, consts: newConsts } = makeJaxpr$1((forwardIn, cotangents) => {
+	const { jaxpr: newJaxpr } = makeJaxpr$1((forwardIn, cotangents) => {
 		const args = [];
 		let forwardInIdx = 0;
 		for (let i = 0; i < undefPrimals.length; i++) if (undefPrimals[i]) args.push(new UndefPrimal(inTypes[i]));
 		else args.push(forwardIn[forwardInIdx++]);
 		return evalJaxprTransposed(jaxpr, args, cotangents);
 	})(forwardInTypes, outTypes);
-	typecheckJaxpr(newJaxpr);
-	const result = {
-		newJaxpr,
-		newConsts
-	};
+	typecheckJaxpr(newJaxpr.jaxpr);
 	if (!transposeJaxprCache.has(jaxpr)) transposeJaxprCache.set(jaxpr, /* @__PURE__ */ new Map());
-	transposeJaxprCache.get(jaxpr).set(cacheKey, result);
-	return result;
+	transposeJaxprCache.get(jaxpr).set(cacheKey, newJaxpr);
+	return newJaxpr;
 }
 function vjpFlat(f, primalsIn) {
-	const { primalsOut, jaxpr, consts } = linearizeFlatUtil(f, primalsIn);
+	const { primalsOut, jaxpr } = linearizeFlatUtil(f, primalsIn);
 	const fVjp = (...cotangents) => {
-		const transposeInputs = [...consts.map((c) => c.ref), ...primalsIn.map((t) => new UndefPrimal(t.aval))];
-		return evalJaxprTransposed(jaxpr, transposeInputs, cotangents);
-	};
-	const dispose$1 = () => {
-		for (const c of consts) c.dispose();
+		const transposeInputs = [...jaxpr.consts.map((c) => c.ref), ...primalsIn.map((t) => new UndefPrimal(t.aval))];
+		return evalJaxprTransposed(jaxpr.jaxpr, transposeInputs, cotangents);
 	};
+	const dispose$1 = () => jaxpr.dispose();
 	return [
 		primalsOut,
 		fVjp,
@@ -4232,150 +4630,6 @@ function jacrev$1(f) {
 	};
 }
 
-//#endregion
-//#region src/library/lax.ts
-var lax_exports = {};
-__export(lax_exports, {
-	conv: () => conv,
-	convGeneralDilated: () => convGeneralDilated,
-	convWithGeneralPadding: () => convWithGeneralPadding,
-	dot: () => dot$1,
-	erf: () => erf,
-	erfc: () => erfc,
-	reduceWindow: () => reduceWindow,
-	stopGradient: () => stopGradient$1
-});
-/**
-* General dot product/contraction operator.
-*
-* Prefer higher-level functions like `jax.numpy.dot()`, `jax.numpy.matmul()`,
-* `jax.numpy.tensordot(), and `jax.numpy.einsum()` where possible.
-*/
-function dot$1(lhs, rhs, { lhsContractingDims: lc = [], rhsContractingDims: rc = [], lhsBatchDims: lb = [], rhsBatchDims: rb = [] } = {}) {
-	if (lc.length !== rc.length) throw new Error(`dot: contracting dims lengths mismatch, got ${JSON.stringify(lc)} and ${JSON.stringify(rc)}`);
-	else if (lb.length !== rb.length) throw new Error(`dot: batch dims lengths mismatch, got ${JSON.stringify(lb)} and ${JSON.stringify(rb)}`);
-	lc = lc.map((a) => checkAxis(a, lhs.ndim));
-	rc = rc.map((a) => checkAxis(a, rhs.ndim));
-	lb = lb.map((a) => checkAxis(a, lhs.ndim));
-	rb = rb.map((a) => checkAxis(a, rhs.ndim));
-	if (lc.some((a) => lb.includes(a))) throw new Error(`dot: lhs contracting dims ${JSON.stringify(lc)} overlap with batch dims ${JSON.stringify(lb)}`);
-	else if (rc.some((a) => rb.includes(a))) throw new Error(`dot: rhs contracting dims ${JSON.stringify(rc)} overlap with batch dims ${JSON.stringify(rb)}`);
-	const lf = range(lhs.ndim).filter((a) => !lc.includes(a) && !lb.includes(a));
-	const rf = range(rhs.ndim).filter((a) => !rc.includes(a) && !rb.includes(a));
-	const lhs2 = lhs.transpose([
-		...lb,
-		...lf,
-		...lc
-	]);
-	const rhs2 = rhs.transpose([
-		...rb,
-		...rf,
-		...rc
-	]);
-	if (lc.length === 0) return mul(lhs2.reshape([
-		...lb.map((a) => lhs.shape[a]),
-		...lf.map((a) => lhs.shape[a]),
-		...rep(rf.length, 1)
-	]), rhs2.reshape([
-		...rb.map((a) => rhs.shape[a]),
-		...rep(lf.length, 1),
-		...rf.map((a) => rhs.shape[a])
-	]));
-	const dotShapeX = lc.map((a) => lhs.shape[a]);
-	const dotShapeY = rc.map((a) => rhs.shape[a]);
-	if (!deepEqual(dotShapeX, dotShapeY)) throw new Error(`dot: shapes not aligned along contracting dims: ${JSON.stringify(dotShapeX)} != ${JSON.stringify(dotShapeY)}`);
-	return dot$2(lhs2.reshape([
-		...lb.map((a) => lhs.shape[a]),
-		...lf.map((a) => lhs.shape[a]),
-		...rep(rf.length, 1),
-		prod(dotShapeX)
-	]), rhs2.reshape([
-		...rb.map((a) => rhs.shape[a]),
-		...rep(lf.length, 1),
-		...rf.map((a) => rhs.shape[a]),
-		prod(dotShapeY)
-	]));
-}
-function padtypeToPads(inShape, filterShape, strides, dilation, padding) {
-	const padType = padding.toUpperCase();
-	switch (padType) {
-		case "VALID": return rep(inShape.length, [0, 0]);
-		case "SAME":
-		case "SAME_LOWER": {
-			const outShape = inShape.map((size$1, i) => Math.ceil(size$1 / strides[i]));
-			const padSizes = zipn(outShape, strides, filterShape, dilation, inShape).map(([o, s, k, d, i]) => Math.max(0, (o - 1) * s + 1 + (k - 1) * d - i));
-			if (padType === "SAME") return padSizes.map((size$1) => [size$1 >> 1, size$1 - (size$1 >> 1)]);
-			else return padSizes.map((size$1) => [size$1 - (size$1 >> 1), size$1 >> 1]);
-		}
-		default: throw new Error(`Unknown padding type: ${padType}`);
-	}
-}
-/**
-* General n-dimensional convolution operator, with optional dilation.
-*
-* The semantics of this operation mimic the `jax.lax.conv_general_dilated`
-* function in JAX, which wraps XLA's general convolution operator.
-*
-* Grouped convolutions are not supported right now.
-*/
-function convGeneralDilated(lhs, rhs, windowStrides, padding, { lhsDilation, rhsDilation } = {}) {
-	if (lhs.ndim < 2) throw new Error("lhs must have at least 2 dimensions");
-	if (rhs.ndim < 2) throw new Error("rhs must have at least 2 dimensions");
-	if (typeof padding === "string") {
-		if (lhsDilation?.some((d) => d !== 1)) throw new Error("String padding is not supported for transposed convolutions");
-		padding = padtypeToPads(lhs.shape.slice(2), rhs.shape.slice(2), windowStrides, rhsDilation ?? rep(rhs.ndim - 2, 1), padding);
-	}
-	return conv$1(lhs, rhs, {
-		strides: windowStrides,
-		padding,
-		lhsDilation,
-		rhsDilation
-	});
-}
-/** Convenience wrapper around `convGeneralDilated`. */
-function convWithGeneralPadding(lhs, rhs, windowStrides, padding, lhsDilation, rhsDilation) {
-	return convGeneralDilated(lhs, rhs, windowStrides, padding, {
-		lhsDilation,
-		rhsDilation
-	});
-}
-/** Convenience wrapper around `convGeneralDilated`. */
-function conv(lhs, rhs, windowStrides, padding) {
-	return convGeneralDilated(lhs, rhs, windowStrides, padding);
-}
-/** Reduce a computation over padded windows. */
-function reduceWindow(operand, computation, windowDimensions, windowStrides) {
-	if (operand.ndim < windowDimensions.length) throw new Error(`Operand dimensions ${operand.ndim} < window ${windowDimensions.length}`);
-	if (!windowStrides) windowStrides = rep(windowDimensions.length, 1);
-	for (let i = 0; i < operand.ndim; i++) computation = vmap$1(computation, 0);
-	return computation(bind1(Primitive.Pool, [operand], {
-		window: windowDimensions,
-		strides: windowStrides
-	}));
-}
-/** The error function: `erf(x) = 2/sqrt(pi) * int[0..x] exp(-t^2) dt`. */
-function erf(x) {
-	return erf$1(x);
-}
-/**
-* The complementary error function: `erfc(x) = 1 - erf(x)`.
-*
-* This function is more accurate than `1 - erf(x)` for large values of `x`,
-* where `erf(x)` is very close to 1.
-*/
-function erfc(x) {
-	return erfc$1(x);
-}
-/**
-* Stops gradient computation.
-*
-* Behaves as the identity function but prevents the flow of gradients during
-* forward or reverse-mode automatic differentiation.
-*/
-function stopGradient$1(x) {
-	return stopGradient(x);
-}
-
 //#endregion
 //#region src/library/numpy/einsum.ts
 const bprod = (...xs) => xs.reduce((acc, x) => acc * BigInt(x), 1n);
@@ -4571,34 +4825,207 @@ function* allPaths(tensors, next) {
 	}
 }
 
+//#endregion
+//#region src/library/numpy-fft.ts
+var numpy_fft_exports = {};
+__export(numpy_fft_exports, {
+	fft: () => fft,
+	ifft: () => ifft
+});
+function checkPairInput(name, a) {
+	const fullName = `jax.numpy.fft.${name}`;
+	if (!deepEqual(a.real.shape, a.imag.shape)) throw new Error(`${fullName}: real and imaginary parts must have the same shape, got ${JSON.stringify(a.real.shape)} and ${JSON.stringify(a.imag.shape)}`);
+	if (a.real.dtype !== a.imag.dtype) throw new Error(`${fullName}: real and imaginary parts must have the same dtype, got ${a.real.dtype} and ${a.imag.dtype}`);
+	if (!isFloatDtype(a.real.dtype)) throw new Error(`${fullName}: input must have a float dtype, got ${a.real.dtype}`);
+}
+function checkPowerOfTwo(name, n) {
+	if ((n & n - 1) !== 0) throw new Error(`jax.numpy.fft.${name}: size must be a power of two, got ${n}`);
+}
+const fftUpdate = jit$1(function fftUpdate$1(i, { real, imag }) {
+	const half = 2 ** i;
+	real = real.reshape([-1, 2 * half]);
+	imag = imag.reshape([-1, 2 * half]);
+	const k = arange(0, half, 1, { dtype: real.dtype });
+	const theta = k.mul(-Math.PI / half);
+	const wr = cos(theta.ref);
+	const wi = sin(theta);
+	const ur = real.ref.slice([], [0, half]);
+	const ui = imag.ref.slice([], [0, half]);
+	const vr = real.slice([], [half, 2 * half]);
+	const vi = imag.slice([], [half, 2 * half]);
+	const tr = vr.ref.mul(wr.ref).sub(vi.ref.mul(wi.ref));
+	const ti = vr.mul(wi).add(vi.mul(wr));
+	return {
+		real: concatenate([ur.ref.add(tr.ref), ur.sub(tr)], -1),
+		imag: concatenate([ui.ref.add(ti.ref), ui.sub(ti)], -1)
+	};
+}, { staticArgnums: [0] });
+/**
+* Compute a one-dimensional discrete Fourier transform.
+*
+* Currently, the size of the axis must be a power of two.
+*/
+function fft(a, axis = -1) {
+	checkPairInput("fft", a);
+	let { real, imag } = a;
+	axis = checkAxis(axis, real.ndim);
+	const n = real.shape[axis];
+	checkPowerOfTwo("fft", n);
+	const logN = Math.log2(n);
+	let perm = null;
+	if (axis !== real.ndim - 1) {
+		perm = range(real.ndim);
+		perm.splice(axis, 1);
+		perm.push(axis);
+		real = real.transpose(perm);
+		imag = imag.transpose(perm);
+	}
+	const originalShape = real.shape;
+	real = real.reshape([-1, ...rep(logN, 2)]).transpose([0, ...range(1, logN + 1).reverse()]).flatten();
+	imag = imag.reshape([-1, ...rep(logN, 2)]).transpose([0, ...range(1, logN + 1).reverse()]).flatten();
+	for (let i = 0; i < logN; i++) ({real, imag} = fftUpdate(i, {
+		real,
+		imag
+	}));
+	real = real.reshape(originalShape);
+	imag = imag.reshape(originalShape);
+	if (perm !== null) {
+		real = real.transpose(invertPermutation(perm));
+		imag = imag.transpose(invertPermutation(perm));
+	}
+	return {
+		real,
+		imag
+	};
+}
+/**
+* Compute a one-dimensional inverse discrete Fourier transform.
+*
+* Currently, the size of the axis must be a power of two.
+*/
+function ifft(a, axis = -1) {
+	checkPairInput("ifft", a);
+	let { real, imag } = a;
+	axis = checkAxis(axis, real.ndim);
+	const n = real.shape[axis];
+	checkPowerOfTwo("ifft", n);
+	imag = imag.mul(-1);
+	const result = fft({
+		real,
+		imag
+	}, axis);
+	return {
+		real: result.real.div(n),
+		imag: result.imag.mul(-1).div(n)
+	};
+}
+
+//#endregion
+//#region src/library/numpy-linalg.ts
+var numpy_linalg_exports = {};
+__export(numpy_linalg_exports, {
+	cholesky: () => cholesky$1,
+	diagonal: () => diagonal,
+	lstsq: () => lstsq,
+	matmul: () => matmul,
+	matrixTranspose: () => matrixTranspose,
+	outer: () => outer,
+	tensordot: () => tensordot,
+	trace: () => trace,
+	vecdot: () => vecdot
+});
+/**
+* Compute the Cholesky decomposition of a (batched) positive-definite matrix.
+*
+* This is like `jax.lax.linalg.cholesky()`, except with an option to symmetrize
+* the input matrix, which is on by default.
+*/
+function cholesky$1(a, { upper = false, symmetrizeInput = true } = {}) {
+	a = fudgeArray(a);
+	if (a.ndim < 2 || a.shape[a.ndim - 1] !== a.shape[a.ndim - 2]) throw new Error(`cholesky: input must be at least 2D square matrix, got ${a.aval}`);
+	if (symmetrizeInput) a = a.ref.add(matrixTranspose(a)).mul(.5);
+	return cholesky(a, { upper });
+}
+/**
+* Return the least-squares solution to a linear equation.
+*
+* For overdetermined systems, this finds the `x` that minimizes `norm(ax - b)`.
+* For underdetermined systems, this finds the minimum-norm solution for `x`.
+*
+* This currently uses Cholesky decomposition to solve the normal equations,
+* under the hood. The method is not as robust as QR or SVD.
+*
+* @param a coefficient matrix of shape `(M, N)`
+* @param b right-hand side of shape `(M,)` or `(M, K)`
+* @return least-squares solution of shape `(N,)` or `(N, K)`
+*/
+function lstsq(a, b) {
+	a = fudgeArray(a);
+	b = fudgeArray(b);
+	if (a.ndim !== 2) throw new Error(`lstsq: 'a' must be a 2D array, got ${a.aval}`);
+	const [m, n] = a.shape;
+	if (b.shape[0] !== m) throw new Error(`lstsq: leading dimension of 'b' must match number of rows of 'a', got ${b.aval}`);
+	const at = matrixTranspose(a.ref);
+	if (m <= n) {
+		const aat = matmul(a, at.ref);
+		const l = cholesky$1(aat, { symmetrizeInput: false });
+		const lb = triangularSolve(l.ref, b, {
+			leftSide: true,
+			lower: true
+		});
+		const llb = triangularSolve(l, lb, {
+			leftSide: true,
+			transposeA: true
+		});
+		return matmul(at, llb.ref);
+	} else {
+		const ata = matmul(at.ref, a);
+		const l = cholesky$1(ata, { symmetrizeInput: false });
+		const atb = matmul(at, b);
+		const lb = triangularSolve(l.ref, atb, {
+			leftSide: true,
+			lower: true
+		});
+		const llb = triangularSolve(l, lb, {
+			leftSide: true,
+			transposeA: true
+		});
+		return llb;
+	}
+}
+
 //#endregion
 //#region src/library/numpy.ts
 var numpy_exports = {};
 __export(numpy_exports, {
 	Array: () => Array$1,
 	DType: () => DType,
-	abs: () => abs,
+	abs: () => absolute,
 	absolute: () => absolute,
 	acos: () => acos,
-	acosh: () => acosh,
+	acosh: () => arccosh,
 	add: () => add,
+	all: () => all,
 	allclose: () => allclose,
+	any: () => any,
 	arange: () => arange,
-	arccos: () => arccos,
+	arccos: () => acos,
 	arccosh: () => arccosh,
+	arcsin: () => asin,
 	arcsinh: () => arcsinh,
-	arctan: () => arctan,
-	arctan2: () => arctan2,
+	arctan: () => atan,
+	arctan2: () => atan2,
 	arctanh: () => arctanh,
 	argmax: () => argmax,
 	argmin: () => argmin,
+	argsort: () => argsort,
 	array: () => array,
 	asin: () => asin,
-	asinh: () => asinh,
+	asinh: () => arcsinh,
 	astype: () => astype,
 	atan: () => atan,
 	atan2: () => atan2,
-	atanh: () => atanh,
+	atanh: () => arctanh,
 	bool: () => bool,
 	broadcastArrays: () => broadcastArrays,
 	broadcastShapes: () => broadcastShapes,
@@ -4608,14 +5035,20 @@ __export(numpy_exports, {
 	clip: () => clip,
 	columnStack: () => columnStack,
 	concatenate: () => concatenate,
+	convolve: () => convolve,
+	corrcoef: () => corrcoef,
+	correlate: () => correlate,
 	cos: () => cos,
 	cosh: () => cosh,
+	cov: () => cov,
+	cumsum: () => cumsum,
+	cumulativeSum: () => cumsum,
 	deg2rad: () => deg2rad,
 	degrees: () => degrees,
 	diag: () => diag,
 	diagonal: () => diagonal,
-	divide: () => divide,
-	dot: () => dot,
+	divide: () => trueDivide,
+	dot: () => dot$1,
 	dstack: () => dstack,
 	e: () => e,
 	einsum: () => einsum,
@@ -4623,8 +5056,10 @@ __export(numpy_exports, {
 	eulerGamma: () => eulerGamma,
 	exp: () => exp,
 	exp2: () => exp2,
+	expandDims: () => expandDims,
 	expm1: () => expm1,
 	eye: () => eye,
+	fft: () => numpy_fft_exports,
 	flip: () => flip,
 	fliplr: () => fliplr,
 	flipud: () => flipud,
@@ -4655,12 +5090,14 @@ __export(numpy_exports, {
 	ldexp: () => ldexp,
 	less: () => less,
 	lessEqual: () => lessEqual,
+	linalg: () => numpy_linalg_exports,
 	linspace: () => linspace,
 	log: () => log,
 	log10: () => log10,
 	log1p: () => log1p,
 	log2: () => log2,
 	matmul: () => matmul,
+	matrixTranspose: () => matrixTranspose,
 	max: () => max,
 	maximum: () => maximum,
 	mean: () => mean,
@@ -4677,10 +5114,10 @@ __export(numpy_exports, {
 	onesLike: () => onesLike,
 	outer: () => outer,
 	pad: () => pad,
-	permuteDims: () => permuteDims,
+	permuteDims: () => transpose,
 	pi: () => pi,
 	positive: () => positive,
-	pow: () => pow,
+	pow: () => power,
 	power: () => power,
 	prod: () => prod$1,
 	promoteTypes: () => promoteTypes,
@@ -4697,6 +5134,7 @@ __export(numpy_exports, {
 	sin: () => sin,
 	sinh: () => sinh,
 	size: () => size,
+	sort: () => sort,
 	sqrt: () => sqrt,
 	square: () => square,
 	squeeze: () => squeeze,
@@ -4861,6 +5299,26 @@ function min(a, axis = null, opts) {
 function max(a, axis = null, opts) {
 	return reduce(a, AluOp.Max, axis, opts);
 }
+/**
+* Test whether all array elements along a given axis evaluate to True.
+*
+* Returns a boolean array with the same shape as `a` with the specified axis
+* removed. If axis is None, returns a scalar.
+*/
+function all(a, axis = null, opts) {
+	a = fudgeArray(a).astype(DType.Bool);
+	return min(a, axis, opts);
+}
+/**
+* Test whether any array element along a given axis evaluates to True.
+*
+* Returns a boolean array with the same shape as `a` with the specified axis
+* removed. If axis is None, returns a scalar.
+*/
+function any(a, axis = null, opts) {
+	a = fudgeArray(a).astype(DType.Bool);
+	return max(a, axis, opts);
+}
 /** Return the peak-to-peak range along a given axis (`max - min`). */
 function ptp(a, axis = null, opts) {
 	a = fudgeArray(a);
@@ -4918,6 +5376,23 @@ function argmax(a, axis, opts) {
 	}).reshape([shape$1[axis], ...rep(shape$1.length - axis - 1, 1)]));
 	return length.sub(max(idx, axis, opts));
 }
+/**
+* Cumulative sum of elements along an axis.
+*
+* Currently this function is `O(n^2)`, we'll improve this later on with a
+* two-phase parallel reduction algorithm.
+*/
+function cumsum(a, axis) {
+	a = fudgeArray(a);
+	if (axis === void 0) {
+		a = a.ravel();
+		axis = 0;
+	} else axis = checkAxis(axis, a.ndim);
+	const n = a.shape[axis];
+	a = moveaxis$1(a, axis, -1);
+	a = broadcast(a, a.shape.concat(n), [-2]);
+	return moveaxis$1(tril(a).sum(-1), -1, axis);
+}
 /** Reverse the elements in an array along the given axes. */
 function flip(x, axis = null) {
 	const nd = ndim(x);
@@ -5027,8 +5502,11 @@ function flipud(x) {
 function fliplr(x) {
 	return flip(x, 1);
 }
-/** @function Alternative name for `numpy.transpose()`. */
-const permuteDims = transpose;
+/** Transpose the last two dimensions of an array. */
+function matrixTranspose(a) {
+	if (ndim(a) < 2) throw new Error(`matrixTranspose: input array must be at least 2D`);
+	return moveaxis$1(a, -1, -2);
+}
 /** Return a 1-D flattened array containing the elements of the input. */
 function ravel(a) {
 	return fudgeArray(a).ravel();
@@ -5044,6 +5522,32 @@ function squeeze(a, axis = null) {
 	return reshape(a, newShape);
 }
 /**
+* Expand the shape of an array by inserting new axes of length 1.
+*
+* @param a - Input array.
+* @param axis - Position(s) in the expanded axes where the new axis (or axes)
+*   is placed. Can be a single integer or an array of integers.
+* @returns Array with the number of dimensions increased.
+*
+* @example
+* ```ts
+* const x = np.array([1, 2]);
+* np.expandDims(x, 0); // Shape [1, 2]
+* np.expandDims(x, 1); // Shape [2, 1]
+* np.expandDims(x, [0, 2]); // Shape [1, 2, 1]
+* ```
+*/
+function expandDims(a, axis) {
+	const as = shape(a);
+	axis = typeof axis === "number" ? [axis] : axis;
+	axis = normalizeAxis(axis, as.length + axis.length);
+	const newShape = [];
+	let srcIdx = 0;
+	for (let i = 0; i < as.length + axis.length; i++) if (axis.includes(i)) newShape.push(1);
+	else newShape.push(as[srcIdx++]);
+	return reshape(a, newShape);
+}
+/**
 * Repeat each element of an array after themselves.
 *
 * If no axis is provided, use the flattened input array, and return a flat
@@ -5131,7 +5635,7 @@ function diagonal(a, offset, axis1, axis2) {
 */
 function diag(v, k = 0) {
 	const a = fudgeArray(v);
-	if (!Number.isInteger(k)) throw new TypeError(`k must be an integer, got ${k}`);
+	if (!Number.isInteger(k)) throw new Error(`k must be an integer, got ${k}`);
 	if (a.ndim === 1) {
 		const n = a.shape[0];
 		const ret = where(eye(n).equal(1), a.ref, zerosLike(a));
@@ -5139,12 +5643,32 @@ function diag(v, k = 0) {
 		else if (k < 0) return pad(ret, [[-k, 0], [0, -k]]);
 		else return ret;
 	} else if (a.ndim === 2) return diagonal(a, k);
-	else throw new TypeError("numpy.diag only supports 1D and 2D arrays");
+	else throw new Error("numpy.diag only supports 1D and 2D arrays");
 }
 /** Calculate the sum of the diagonal of an array along the given axes. */
 function trace(a, offset = 0, axis1 = 0, axis2 = 1) {
 	return diagonal(a, offset, axis1, axis2).sum(-1);
 }
+/**
+* Return a sorted copy of an array.
+*
+* The array is sorted along a specified axis (the last by default). This may be
+* an unstable sort, and it dispatches to device-specific implementation.
+*/
+function sort(a, axis = -1) {
+	return fudgeArray(a).sort(axis);
+}
+/**
+* Return indices that would sort an array. This may be an unstable sorting
+* algorithm; it need not preserve order of indices in ties.
+*
+* Returns an array of `int32` indices.
+*
+* The array is sorted along a specified axis (the last by default).
+*/
+function argsort(a, axis = -1) {
+	return fudgeArray(a).argsort(axis);
+}
 /** Return if two arrays are element-wise equal within a tolerance. */
 function allclose(actual, expected, options) {
 	const { rtol = 1e-5, atol = 1e-7 } = options ?? {};
@@ -5153,16 +5677,19 @@ function allclose(actual, expected, options) {
 	if (!deepEqual(x.shape, y.shape)) return false;
 	const xData = x.dataSync();
 	const yData = y.dataSync();
-	for (let i = 0; i < xData.length; i++) if (Math.abs(xData[i] - yData[i]) > atol + rtol * Math.abs(yData[i])) return false;
+	for (let i = 0; i < xData.length; i++) {
+		if (isNaN(xData[i]) !== isNaN(yData[i])) return false;
+		if (Math.abs(xData[i] - yData[i]) > atol + rtol * Math.abs(yData[i])) return false;
+	}
 	return true;
 }
 /** Matrix product of two arrays. */
 function matmul(x, y) {
-	if (ndim(x) === 0 || ndim(y) === 0) throw new TypeError("matmul: x and y must be at least 1D");
+	if (ndim(x) === 0 || ndim(y) === 0) throw new Error("matmul: x and y must be at least 1D");
 	x = x, y = y;
 	if (y.ndim === 1) return dot$2(x, y);
 	const numBatchDims = Math.min(Math.max(x.ndim, 2), y.ndim) - 2;
-	return dot$1(x, y, {
+	return dot(x, y, {
 		lhsContractingDims: [-1],
 		rhsContractingDims: [-2],
 		lhsBatchDims: range(-2 - numBatchDims, -2),
@@ -5170,11 +5697,11 @@ function matmul(x, y) {
 	});
 }
 /** Dot product of two arrays. */
-function dot(x, y) {
+function dot$1(x, y) {
 	if (ndim(x) === 0 || ndim(y) === 0) return multiply(x, y);
 	x = x, y = y;
 	if (y.ndim === 1) return dot$2(x, y);
-	return dot$1(x, y, {
+	return dot(x, y, {
 		lhsContractingDims: [-1],
 		rhsContractingDims: [-2]
 	});
@@ -5190,7 +5717,7 @@ function tensordot(x, y, axes = 2) {
 	x = fudgeArray(x);
 	y = fudgeArray(y);
 	if (typeof axes === "number") axes = [range(-axes, 0), range(axes)];
-	return dot$1(x, y, {
+	return dot(x, y, {
 		lhsContractingDims: axes[0],
 		rhsContractingDims: axes[1]
 	});
@@ -5283,7 +5810,7 @@ function einsum(...args) {
 		const [b, bidx] = processSingleTensor(operands[j], indices[j], indices[i]);
 		indexReduced = indexReduced.filter((idx) => aidx.includes(idx));
 		const indexBatch = aidx.filter((idx) => bidx.includes(idx) && !indexReduced.includes(idx));
-		const result = dot$1(a, b, {
+		const result = dot(a, b, {
 			lhsContractingDims: indexReduced.map((idx) => aidx.indexOf(idx)),
 			rhsContractingDims: indexReduced.map((idx) => bidx.indexOf(idx)),
 			lhsBatchDims: indexBatch.map((idx) => aidx.indexOf(idx)),
@@ -5311,7 +5838,7 @@ function einsum(...args) {
 * Returned array has shape `[...x.shape[:-1], ...y.shape[:-1]]`.
 */
 function inner(x, y) {
-	return dot$1(fudgeArray(x), fudgeArray(y), {
+	return dot(fudgeArray(x), fudgeArray(y), {
 		lhsContractingDims: [-1],
 		rhsContractingDims: [-1]
 	});
@@ -5344,6 +5871,30 @@ function vecdot(x, y, { axis } = {}) {
 function vdot(x, y) {
 	return dot$2(ravel(x), ravel(y));
 }
+function _convImpl(name, x, y, mode) {
+	if (x.ndim !== 1 || y.ndim !== 1) throw new Error(`${name}: both inputs must be 1D arrays, got ${x.ndim}D and ${y.ndim}D`);
+	let flipOutput = false;
+	if (x.shape[0] < y.shape[0]) {
+		[x, y] = [y, x];
+		if (name === "correlate") flipOutput = true;
+	}
+	if (name === "convolve") y = flip(y);
+	let padding;
+	if (mode === "valid") padding = "VALID";
+	else if (mode === "same") padding = "SAME_LOWER";
+	else if (mode === "full") padding = [[y.shape[0] - 1, y.shape[0] - 1]];
+	else throw new Error(`${name}: invalid mode ${mode}, expected "full", "same", or "valid"`);
+	const z = conv(x.slice(null, null), y.slice(null, null), [1], padding).slice(0, 0);
+	return flipOutput ? flip(z) : z;
+}
+/** Convolution of two one-dimensional arrays. */
+function convolve(x, y, mode = "full") {
+	return _convImpl("convolve", x, y, mode);
+}
+/** Correlation of two one dimensional arrays. */
+function correlate(x, y, mode = "valid") {
+	return _convImpl("correlate", x, y, mode);
+}
 /**
 * Return a tuple of coordinate matrices from coordinate vectors.
 *
@@ -5352,7 +5903,7 @@ function vdot(x, y) {
 */
 function meshgrid(xs, { indexing } = {}) {
 	indexing ??= "xy";
-	for (const x of xs) if (x.ndim !== 1) throw new TypeError(`meshgrid: all inputs must be 1D arrays, got ${x.ndim}D array`);
+	for (const x of xs) if (x.ndim !== 1) throw new Error(`meshgrid: all inputs must be 1D arrays, got ${x.ndim}D array`);
 	if (xs.length <= 1) return xs;
 	if (indexing === "xy") {
 		const [a, b, ...rest] = xs;
@@ -5371,43 +5922,6 @@ function meshgrid(xs, { indexing } = {}) {
 	return xs.map((x, i) => broadcast(x, shape$1, [...range(i), ...range(i + 1, xs.length)]));
 }
 /**
-* Return an array with ones on and below the diagonal and zeros elsewhere.
-*
-* If `k` is provided, it specifies the sub-diagonal on and below which the
-* array is filled with ones. `k=0` is the main diagonal, `k<0` is below it, and
-* `k>0` is above it.
-*/
-function tri(n, m, k = 0, { dtype, device } = {}) {
-	m ??= n;
-	dtype ??= DType.Float32;
-	if (!Number.isInteger(n) || n < 0) throw new TypeError(`tri: n must be a non-negative integer, got ${n}`);
-	if (!Number.isInteger(m) || m < 0) throw new TypeError(`tri: m must be a non-negative integer, got ${m}`);
-	if (!Number.isInteger(k)) throw new TypeError(`tri: k must be an integer, got ${k}`);
-	const rows = arange(k, n + k, 1, {
-		dtype: DType.Int32,
-		device
-	});
-	const cols = arange(0, m, 1, {
-		dtype: DType.Int32,
-		device
-	});
-	return rows.reshape([n, 1]).greaterEqual(cols).astype(dtype);
-}
-/** Return the lower triangle of an array. Must be of dimension >= 2. */
-function tril(a, k = 0) {
-	if (ndim(a) < 2) throw new TypeError(`tril: input array must be at least 2D, got ${ndim(a)}D`);
-	a = fudgeArray(a);
-	const [n, m] = a.shape.slice(-2);
-	return where(tri(n, m, k, { dtype: bool }), a.ref, zerosLike(a));
-}
-/** Return the upper triangle of an array. Must be of dimension >= 2. */
-function triu(a, k = 0) {
-	if (ndim(a) < 2) throw new TypeError(`tril: input array must be at least 2D, got ${ndim(a)}D`);
-	a = fudgeArray(a);
-	const [n, m] = a.shape.slice(-2);
-	return where(tri(n, m, k - 1, { dtype: bool }), zerosLike(a.ref), a);
-}
-/**
 * Clip (limit) the values in an array.
 *
 * Given an interval, values outside the interval are clipped to the interval
@@ -5431,8 +5945,6 @@ function absolute(x) {
 	x = fudgeArray(x);
 	return where(less(x.ref, 0), x.ref.mul(-1), x);
 }
-/** @function Alias of `jax.numpy.absolute()`. */
-const abs = absolute;
 /** Return an element-wise indication of sign of the input. */
 function sign(x) {
 	x = fudgeArray(x);
@@ -5511,12 +6023,6 @@ const atan2 = jit$1(function atan2$1(y, x) {
 	const denom = where(xNeg, y, r.add(x));
 	return atan(numer.div(denom)).mul(2);
 });
-/** @function Alias of `jax.numpy.acos()`. */
-const arccos = acos;
-/** @function Alias of `jax.numpy.atan()`. */
-const arctan = atan;
-/** @function Alias of `jax.numpy.atan2()`. */
-const arctan2 = atan2;
 /** Element-wise subtraction, with broadcasting. */
 function subtract(x, y) {
 	x = fudgeArray(x);
@@ -5547,8 +6053,6 @@ const fmod = jit$1(function fmod$1(x, y) {
 const remainder = jit$1(function remainder$1(x, y) {
 	return mod(mod(x, y.ref).add(y.ref), y);
 });
-/** @function Alias of `jax.numpy.trueDivide()`. */
-const divide = trueDivide;
 /** Round input to the nearest integer towards zero. */
 function trunc(x) {
 	return idiv(x, 1);
@@ -5570,9 +6074,9 @@ function ldexp(x1, x2) {
 */
 function frexp(x) {
 	x = fudgeArray(x);
-	const absx = abs(x.ref);
+	const absx = absolute(x.ref);
 	const exponent = where(equal(x.ref, 0), 0, floor(log2(absx)).add(1).astype(DType.Int32));
-	const mantissa = divide(x, exp2(exponent.ref.astype(x.dtype)));
+	const mantissa = x.div(exp2(exponent.ref.astype(x.dtype)));
 	return [mantissa, exponent];
 }
 /** Calculate `2**p` for all p in the input array. */
@@ -5612,10 +6116,11 @@ const degrees = rad2deg;
 * Computes first array raised to power of second array, element-wise.
 */
 const power = jit$1(function power$1(x1, x2) {
-	return exp(log(x1).mul(x2));
+	const x2i = trunc(x2.ref);
+	const shouldBeNaN = multiply(x2.ref.notEqual(x2i.ref), x1.ref.less(0));
+	const resultSign = where(mod(x2i, 2).notEqual(0), where(x1.ref.less(0), -1, 1), 1);
+	return where(shouldBeNaN, nan, exp(log(absolute(x1)).mul(x2)).mul(resultSign));
 });
-/** @function Alias of `jax.numpy.power()`. */
-const pow = power;
 /** @function Calculate the element-wise cube root of the input array. */
 const cbrt = jit$1(function cbrt$1(x) {
 	const sgn = where(less(x.ref, 0), -1, 1);
@@ -5681,12 +6186,6 @@ const arccosh = jit$1(function arccosh$1(x) {
 const arctanh = jit$1(function arctanh$1(x) {
 	return log(add(1, x.ref).div(subtract(1, x))).mul(.5);
 });
-/** @function Alias of `jax.numpy.arcsinh()`. */
-const asinh = arcsinh;
-/** @function Alias of `jax.numpy.arccosh()`. */
-const acosh = arccosh;
-/** @function Alias of `jax.numpy.arctanh()`. */
-const atanh = arctanh;
 /**
 * Compute the variance of an array.
 *
@@ -5716,6 +6215,26 @@ function var_(x, axis = null, opts) {
 function std(x, axis = null, opts) {
 	return sqrt(var_(x, axis, opts));
 }
+/** Estimate the sample covariance of a set of variables. */
+function cov(x, y) {
+	x = fudgeArray(x);
+	if (x.ndim === 1) x = x.reshape([1, x.shape[0]]);
+	if (y !== void 0) {
+		y = fudgeArray(y);
+		if (y.ndim === 1) y = y.reshape([1, y.shape[0]]);
+		x = vstack([x, y]);
+	}
+	const [_M, N] = x.shape;
+	x = x.ref.sub(x.mean(1, { keepdims: true }));
+	return dot$1(x.ref, x.transpose()).div(N - 1);
+}
+/** Compute the Pearson correlation coefficients (in range `[-1, 1]`). */
+function corrcoef(x, y) {
+	const c = cov(x, y);
+	const variances = diag(c.ref);
+	const norm = sqrt(outer(variances.ref, variances));
+	return c.div(norm);
+}
 /** Test element-wise for positive or negative infinity, return bool array. */
 function isinf(x) {
 	x = fudgeArray(x);
@@ -5745,6 +6264,253 @@ const isfinite = jit$1(function isfinite$1(x) {
 	return isnan(x.ref).add(isinf(x)).notEqual(true);
 });
 
+//#endregion
+//#region src/library/lax-linalg.ts
+var lax_linalg_exports = {};
+__export(lax_linalg_exports, {
+	cholesky: () => cholesky,
+	triangularSolve: () => triangularSolve
+});
+/**
+* Compute the Cholesky decomposition of a symmetric positive-definite matrix.
+*
+* The Cholesky decomposition of a matrix `A` is:
+*
+* - A = L @ L^T  (for upper=false, default)
+* - A = U^T @ U  (for upper=true)
+*
+* where `L` is a lower-triangular matrix and `U` is an upper-triangular matrix.
+* The input matrix must be symmetric and positive-definite.
+*
+* @example
+* ```ts
+* import { lax, numpy as np } from "@jax-js/jax";
+*
+* const x = np.array([[2., 1.], [1., 2.]]);
+*
+* // Lower Cholesky factorization (default):
+* const L = lax.linalg.cholesky(x);
+* // L ≈ [[1.4142135, 0], [0.70710677, 1.2247449]]
+*
+* // Upper Cholesky factorization:
+* const U = lax.linalg.cholesky(x, { upper: true });
+* // U ≈ [[1.4142135, 0.70710677], [0, 1.2247449]]
+* ```
+*/
+function cholesky(a, { upper = false } = {}) {
+	const L = cholesky$2(a);
+	return upper ? moveaxis$1(L, -2, -1) : L;
+}
+/**
+* Solve a triangular linear system.
+*
+* Solves `a @ x = b` (if leftSide=true) or `x @ a = b` (if leftSide=false)
+* where `a` is a triangular matrix.
+*
+* @example
+* ```ts
+* import { lax, numpy as np } from "@jax-js/jax";
+*
+* const L = np.array([[2., 0.], [1., 3.]]);
+* const b = np.array([4., 7.]).reshape([2, 1]);
+*
+* // Solve L @ x = b
+* const x = lax.linalg.triangularSolve(L, b, { leftSide: true, lower: true });
+* // x = [[2.], [5./3.]]
+* ```
+*/
+function triangularSolve(a, b, { leftSide = false, lower = false, transposeA = false, unitDiagonal = false } = {}) {
+	a = fudgeArray(a);
+	b = fudgeArray(b);
+	if (!leftSide) transposeA = !transposeA;
+	else b = moveaxis$1(b, -2, -1);
+	if (transposeA) a = moveaxis$1(a, -2, -1);
+	let x = triangularSolve$1(a, b, {
+		lower,
+		unitDiagonal
+	});
+	if (leftSide) x = moveaxis$1(x, -2, -1);
+	return x;
+}
+
+//#endregion
+//#region src/library/lax.ts
+var lax_exports = {};
+__export(lax_exports, {
+	conv: () => conv,
+	convGeneralDilated: () => convGeneralDilated,
+	convWithGeneralPadding: () => convWithGeneralPadding,
+	dot: () => dot,
+	erf: () => erf,
+	erfc: () => erfc,
+	linalg: () => lax_linalg_exports,
+	reduceWindow: () => reduceWindow,
+	stopGradient: () => stopGradient$1
+});
+/**
+* General dot product/contraction operator.
+*
+* Prefer higher-level functions like `jax.numpy.dot()`, `jax.numpy.matmul()`,
+* `jax.numpy.tensordot(), and `jax.numpy.einsum()` where possible.
+*/
+function dot(lhs, rhs, { lhsContractingDims: lc = [], rhsContractingDims: rc = [], lhsBatchDims: lb = [], rhsBatchDims: rb = [] } = {}) {
+	if (lc.length !== rc.length) throw new Error(`dot: contracting dims lengths mismatch, got ${JSON.stringify(lc)} and ${JSON.stringify(rc)}`);
+	else if (lb.length !== rb.length) throw new Error(`dot: batch dims lengths mismatch, got ${JSON.stringify(lb)} and ${JSON.stringify(rb)}`);
+	lc = lc.map((a) => checkAxis(a, lhs.ndim));
+	rc = rc.map((a) => checkAxis(a, rhs.ndim));
+	lb = lb.map((a) => checkAxis(a, lhs.ndim));
+	rb = rb.map((a) => checkAxis(a, rhs.ndim));
+	if (lc.some((a) => lb.includes(a))) throw new Error(`dot: lhs contracting dims ${JSON.stringify(lc)} overlap with batch dims ${JSON.stringify(lb)}`);
+	else if (rc.some((a) => rb.includes(a))) throw new Error(`dot: rhs contracting dims ${JSON.stringify(rc)} overlap with batch dims ${JSON.stringify(rb)}`);
+	const lf = range(lhs.ndim).filter((a) => !lc.includes(a) && !lb.includes(a));
+	const rf = range(rhs.ndim).filter((a) => !rc.includes(a) && !rb.includes(a));
+	const lhs2 = lhs.transpose([
+		...lb,
+		...lf,
+		...lc
+	]);
+	const rhs2 = rhs.transpose([
+		...rb,
+		...rf,
+		...rc
+	]);
+	if (lc.length === 0) return mul(lhs2.reshape([
+		...lb.map((a) => lhs.shape[a]),
+		...lf.map((a) => lhs.shape[a]),
+		...rep(rf.length, 1)
+	]), rhs2.reshape([
+		...rb.map((a) => rhs.shape[a]),
+		...rep(lf.length, 1),
+		...rf.map((a) => rhs.shape[a])
+	]));
+	const dotShapeX = lc.map((a) => lhs.shape[a]);
+	const dotShapeY = rc.map((a) => rhs.shape[a]);
+	if (!deepEqual(dotShapeX, dotShapeY)) throw new Error(`dot: shapes not aligned along contracting dims: ${JSON.stringify(dotShapeX)} != ${JSON.stringify(dotShapeY)}`);
+	return dot$2(lhs2.reshape([
+		...lb.map((a) => lhs.shape[a]),
+		...lf.map((a) => lhs.shape[a]),
+		...rep(rf.length, 1),
+		prod(dotShapeX)
+	]), rhs2.reshape([
+		...rb.map((a) => rhs.shape[a]),
+		...rep(lf.length, 1),
+		...rf.map((a) => rhs.shape[a]),
+		prod(dotShapeY)
+	]));
+}
+function padtypeToPads(inShape, filterShape, strides, dilation, padding) {
+	const padType = padding.toUpperCase();
+	switch (padType) {
+		case "VALID": return rep(inShape.length, [0, 0]);
+		case "SAME":
+		case "SAME_LOWER": {
+			const outShape = inShape.map((size$1, i) => Math.ceil(size$1 / strides[i]));
+			const padSizes = zipn(outShape, strides, filterShape, dilation, inShape).map(([o, s, k, d, i]) => Math.max(0, (o - 1) * s + 1 + (k - 1) * d - i));
+			if (padType === "SAME") return padSizes.map((size$1) => [size$1 >> 1, size$1 - (size$1 >> 1)]);
+			else return padSizes.map((size$1) => [size$1 - (size$1 >> 1), size$1 >> 1]);
+		}
+		default: throw new Error(`Unknown padding type: ${padType}`);
+	}
+}
+/**
+* General n-dimensional convolution operator, with optional dilation.
+*
+* The semantics of this operation mimic the `jax.lax.conv_general_dilated`
+* function in JAX, which wraps XLA's general convolution operator.
+*
+* Grouped convolutions are not supported right now.
+*/
+function convGeneralDilated(lhs, rhs, windowStrides, padding, { lhsDilation, rhsDilation, featureGroupCount = 1 } = {}) {
+	if (lhs.ndim < 2) throw new Error("lhs must have at least 2 dimensions");
+	if (rhs.ndim < 2) throw new Error("rhs must have at least 2 dimensions");
+	if (typeof padding === "string") {
+		if (lhsDilation?.some((d) => d !== 1)) throw new Error("String padding is not supported for transposed convolutions");
+		padding = padtypeToPads(lhs.shape.slice(2), rhs.shape.slice(2), windowStrides, rhsDilation ?? rep(rhs.ndim - 2, 1), padding);
+	}
+	if (featureGroupCount !== 1) {
+		const G = featureGroupCount;
+		const [N, C_in, ...xs] = lhs.shape;
+		const [C_out, C_in_per_group, ...ks] = rhs.shape;
+		if (C_in % G !== 0) throw new Error(`featureGroupCount=${G} must divide input channels=${C_in}`);
+		if (C_out % G !== 0) throw new Error(`featureGroupCount=${G} must divide output channels=${C_out}`);
+		if (C_in / G !== C_in_per_group) throw new Error(`rhs input channels=${C_in_per_group} must equal lhs input channels / groups=${C_in / G}`);
+		const lhsGrouped = moveaxis(lhs.reshape([
+			N,
+			G,
+			C_in / G,
+			...xs
+		]), 1, 0);
+		const rhsGrouped = rhs.reshape([
+			G,
+			C_out / G,
+			C_in_per_group,
+			...ks
+		]);
+		const result = conv$1(lhsGrouped, rhsGrouped, {
+			vmapDims: 1,
+			strides: windowStrides,
+			padding,
+			lhsDilation,
+			rhsDilation
+		});
+		const ys = result.shape.slice(3);
+		return moveaxis(result, 0, 1).reshape([
+			N,
+			C_out,
+			...ys
+		]);
+	}
+	return conv$1(lhs, rhs, {
+		strides: windowStrides,
+		padding,
+		lhsDilation,
+		rhsDilation
+	});
+}
+/** Convenience wrapper around `convGeneralDilated`. */
+function convWithGeneralPadding(lhs, rhs, windowStrides, padding, lhsDilation, rhsDilation) {
+	return convGeneralDilated(lhs, rhs, windowStrides, padding, {
+		lhsDilation,
+		rhsDilation
+	});
+}
+/** Convenience wrapper around `convGeneralDilated`. */
+function conv(lhs, rhs, windowStrides, padding) {
+	return convGeneralDilated(lhs, rhs, windowStrides, padding);
+}
+/** Reduce a computation over padded windows. */
+function reduceWindow(operand, computation, windowDimensions, windowStrides) {
+	if (operand.ndim < windowDimensions.length) throw new Error(`Operand dimensions ${operand.ndim} < window ${windowDimensions.length}`);
+	if (!windowStrides) windowStrides = rep(windowDimensions.length, 1);
+	for (let i = 0; i < operand.ndim; i++) computation = vmap$1(computation, 0);
+	return computation(bind1(Primitive.Pool, [operand], {
+		window: windowDimensions,
+		strides: windowStrides
+	}));
+}
+/** The error function: `erf(x) = 2/sqrt(pi) * int[0..x] exp(-t^2) dt`. */
+function erf(x) {
+	return erf$1(x);
+}
+/**
+* The complementary error function: `erfc(x) = 1 - erf(x)`.
+*
+* This function is more accurate than `1 - erf(x)` for large values of `x`,
+* where `erf(x)` is very close to 1.
+*/
+function erfc(x) {
+	return erfc$1(x);
+}
+/**
+* Stops gradient computation.
+*
+* Behaves as the identity function but prevents the flow of gradients during
+* forward or reverse-mode automatic differentiation.
+*/
+function stopGradient$1(x) {
+	return stopGradient(x);
+}
+
 //#endregion
 //#region src/library/nn.ts
 var nn_exports = {};
@@ -5753,6 +6519,10 @@ __export(nn_exports, {
 	elu: () => elu,
 	gelu: () => gelu,
 	glu: () => glu,
+	hardSigmoid: () => hardSigmoid,
+	hardSilu: () => hardSilu,
+	hardSwish: () => hardSilu,
+	hardTanh: () => hardTanh,
 	identity: () => identity,
 	leakyRelu: () => leakyRelu,
 	logSigmoid: () => logSigmoid,
@@ -5763,14 +6533,17 @@ __export(nn_exports, {
 	oneHot: () => oneHot,
 	relu: () => relu,
 	relu6: () => relu6,
+	selu: () => selu,
 	sigmoid: () => sigmoid,
 	silu: () => silu,
 	softSign: () => softSign,
 	softmax: () => softmax,
 	softplus: () => softplus,
+	sparsePlus: () => sparsePlus,
+	sparseSigmoid: () => sparseSigmoid,
 	squareplus: () => squareplus,
 	standardize: () => standardize,
-	swish: () => swish
+	swish: () => silu
 });
 /**
 * Rectified Linear Unit (ReLU) activation function:
@@ -5805,6 +6578,28 @@ function softplus(x) {
 	return log(exp(x).add(1));
 }
 /**
+* @function
+* Sparse plus function:
+*
+* - When `x <= -1`: `0`
+* - When `-1 < x < 1`: `(x+1)**2 / 4`
+* - When `x >= 1`: `x`
+*/
+const sparsePlus = jit$1((x) => {
+	return where(x.ref.lessEqual(-1), 0, where(x.ref.less(1), square(x.ref.add(1)).mul(.25), x));
+});
+/**
+* @function
+* Sparse sigmoid activation function.
+*
+* - When `x <= -1`: `0`
+* - When `-1 < x < 1`: `(x + 1) / 2`
+* - When `x >= 1`: `1`
+*/
+const sparseSigmoid = jit$1((x) => {
+	return clip(x.add(1).mul(.5), 0, 1);
+});
+/**
 * Soft-sign activation function, computed element-wise:
 * `softsign(x) = x / (|x| + 1)`.
 */
@@ -5826,17 +6621,6 @@ const silu = jit$1(function silu$1(x) {
 	return x.ref.mul(sigmoid(x));
 });
 /**
-* @function
-* Sigmoid-weighted Linear Unit (SiLU) activation function, also known as
-* Swish, computed element-wise:
-* `silu(x) = x * sigmoid(x) = x / (1 + exp(-x))`.
-*
-* `swish()` and `silu()` are both aliases for the same function.
-*
-* Reference: https://en.wikipedia.org/wiki/Swish_function
-*/
-const swish = silu;
-/**
 * Log-sigmoid activation function, computed element-wise:
 * `log_sigmoid(x) = log(sigmoid(x)) = -log(1 + exp(-x))`.
 */
@@ -5853,6 +6637,19 @@ function leakyRelu(x, negativeSlope = .01) {
 	x = fudgeArray(x);
 	return where(less(x.ref, 0), x.ref.mul(negativeSlope), x);
 }
+/** Hard sigmoid activation function: `relu6(x+3)/6`. */
+function hardSigmoid(x) {
+	return relu6(add(x, 3)).mul(1 / 6);
+}
+/** Hard SiLU (swish) activation function: `x * hardSigmoid(x)`. */
+function hardSilu(x) {
+	x = fudgeArray(x);
+	return x.ref.mul(hardSigmoid(x));
+}
+/** Hard tanh activation function: `clip(x, -1, 1)`. */
+function hardTanh(x) {
+	return clip(x, -1, 1);
+}
 /**
 * Exponential linear unit activation function.
 *
@@ -5875,6 +6672,20 @@ function celu(x, alpha = 1) {
 }
 /**
 * @function
+* Scaled exponential linear unit activation.
+*
+* Computes the element-wise function:
+* `selu(x) = lambda * (x > 0 ? x : alpha * (exp(x) - 1))`
+*
+* Where `alpha = 1.6732632423543772` and `lambda = 1.0507009873554805`.
+*/
+const selu = jit$1(function selu$1(x) {
+	const alpha = 1.6732632423543772;
+	const lambda = 1.0507009873554805;
+	return where(x.ref.less(0), expm1(x.ref).mul(alpha), x).mul(lambda);
+});
+/**
+* @function
 * Gaussion error linear unit (GELU) activation function.
 *
 * This is computed element-wise. There are two variants depending on whether
@@ -5968,22 +6779,22 @@ function logSoftmax(x, axis = -1) {
 *
 * Reference: https://en.wikipedia.org/wiki/LogSumExp
 */
-function logsumexp(x, axis = null) {
+function logsumexp(x, axis = null, opts) {
 	x = fudgeArray(x);
 	axis = normalizeAxis(axis, x.ndim);
 	if (axis.length === 0) return x;
-	const xMax = stopGradient(max(x.ref, axis));
-	const xMaxDims = broadcast(xMax.ref, x.shape, axis);
-	const shifted = x.sub(xMaxDims);
-	return xMax.add(log(exp(shifted).sum(axis)));
+	const xMax = stopGradient(max(x.ref, axis, { keepdims: true }));
+	const shifted = x.sub(xMax.ref);
+	const result = xMax.add(log(exp(shifted).sum(axis, { keepdims: true })));
+	return opts?.keepdims ? result : squeeze(result, axis);
 }
 /** Log-mean-exp reduction, like `jax.nn.logsumexp()` but subtracts `log(n)`. */
-function logmeanexp(x, axis = null) {
+function logmeanexp(x, axis = null, opts) {
 	x = fudgeArray(x);
 	axis = normalizeAxis(axis, x.ndim);
 	if (axis.length === 0) return x;
 	const n = axis.reduce((acc, a) => acc * x.shape[a], 1);
-	return logsumexp(x, axis).sub(Math.log(n));
+	return logsumexp(x, axis, opts).sub(Math.log(n));
 }
 /**
 * Standardizes input to zero mean and unit variance.
@@ -6028,8 +6839,11 @@ var random_exports = {};
 __export(random_exports, {
 	bernoulli: () => bernoulli,
 	bits: () => bits,
+	cauchy: () => cauchy,
 	exponential: () => exponential,
+	gumbel: () => gumbel,
 	key: () => key,
+	laplace: () => laplace,
 	normal: () => normal,
 	split: () => split,
 	uniform: () => uniform
@@ -6088,6 +6902,16 @@ function bernoulli(key$1, p = .5, shape$1 = []) {
 }
 /**
 * @function
+* Sample from a Cauchy distribution with location 0 and scale 1.
+*
+* Uses inverse transform sampling: `x = tan(π * (u - 0.5))` where u ~ Uniform(0, 1).
+*/
+const cauchy = jit$1(function cauchy$1(key$1, shape$1 = []) {
+	const u = uniform(key$1, shape$1);
+	return tan(u.sub(.5).mul(Math.PI));
+}, { staticArgnums: [1] });
+/**
+* @function
 * Sample exponential random values according to `p(x) = exp(-x)`.
 */
 const exponential = jit$1(function exponential$1(key$1, shape$1 = []) {
@@ -6096,6 +6920,30 @@ const exponential = jit$1(function exponential$1(key$1, shape$1 = []) {
 }, { staticArgnums: [1] });
 /**
 * @function
+* Sample from a Gumbel distribution with location 0 and scale 1.
+*
+* Uses inverse transform sampling: `x = -log(-log(u))` where u ~ Uniform(0, 1).
+*/
+const gumbel = jit$1(function gumbel$1(key$1, shape$1 = []) {
+	const u = uniform(key$1, shape$1);
+	return negative(log(negative(log1p(negative(u)))));
+}, { staticArgnums: [1] });
+/**
+* @function
+* Sample from a Laplace distribution with location 0 and scale 1.
+*
+* Uses inverse transform sampling: the CDF is `F(x) = 0.5 + 0.5 * sign(x) * (1 - exp(-|x|))`.
+* Inverting: `x = -sign(u - 0.5) * log(1 - 2 * |u - 0.5|)`.
+*/
+const laplace = jit$1(function laplace$1(key$1, shape$1 = []) {
+	const u = uniform(key$1, shape$1);
+	const centered = u.sub(.5);
+	const s = sign(centered.ref);
+	const absVal = absolute(centered);
+	return s.mul(log1p(absVal.mul(-2)).mul(-1));
+}, { staticArgnums: [1] });
+/**
+* @function
 * Sample random values according to `p(x) = 1/sqrt(2pi) * exp(-x^2/2)`.
 *
 * Unlike JAX, this uses the Box-Muller transform. JAX uses the erf_inv primitive instead and
@@ -6204,11 +7052,6 @@ const valueAndGrad = valueAndGrad$1;
 */
 const jacrev = jacrev$1;
 /**
-* @function
-* Compute the Jacobian with reverse-mode AD. Alias for `jacrev()`.
-*/
-const jacobian = jacrev;
-/**
 * Wait until all `Array` leaves are ready by calling `Array.blockUntilReady()`.
 *
 * This can be used to wait for the results of an intermediate computation to
@@ -6243,5 +7086,4 @@ async function devicePut(x, device) {
 }
 
 //#endregion
-export { Array$1 as Array, DType, Jaxpr, blockUntilReady, defaultDevice, devicePut, devices, grad, init, jacfwd, jacobian, jacrev, jit, jvp, lax_exports as lax, linearize, makeJaxpr, nn_exports as nn, numpy_exports as numpy, random_exports as random, scipy_special_exports as scipySpecial, setDebug, tree_exports as tree, valueAndGrad, vjp, vmap };
-//# sourceMappingURL=index.js.map
+export { Array$1 as Array, ClosedJaxpr, DType, Jaxpr, blockUntilReady, defaultDevice, devicePut, devices, grad, init, jacfwd, jacrev as jacobian, jacrev, jit, jvp, lax_exports as lax, linearize, makeJaxpr, nn_exports as nn, numpy_exports as numpy, random_exports as random, scipy_special_exports as scipySpecial, setDebug, tree_exports as tree, valueAndGrad, vjp, vmap };