tensorgrad 0.0.12 → 0.0.13

package/dist/module.d.ts

@@ -1,55 +0,0 @@
-import type { Tensor, Shape, Dtype } from './ir.js';
-/** How a parameter's initial values are produced.
- *  - `'randn'` — Gaussian, with `scale` (default 0.02). The common case for
- *    weight matrices and embeddings.
- *  - `'zeros'` — fill with 0. Common for biases and LayerNorm beta.
- *  - `'ones'`  — fill with 1. Common for LayerNorm gain.
- *  - Custom function — receives total element count and shape, returns the
- *    Float32Array. Use for fan-in scaling or any non-standard scheme.
- */
-export type InitSpec = 'randn' | 'zeros' | 'ones' | ((size: number, shape: readonly number[]) => Float32Array);
-export interface ParamOptions {
-    dtype?: Dtype;
-    /** Init kind. Default: `'randn'`. */
-    init?: InitSpec;
-    /** Std dev for `'randn'`. Default 0.02. Ignored for non-randn init. */
-    scale?: number;
-    /** Whether AdamW (when `weightDecay > 0`) should apply decoupled weight
-     *  decay to this param. Default: `true` for `'randn'` init (weight matrices,
-     *  embeddings), `false` for `'zeros'` / `'ones'` (biases, LN gains). Override
-     *  to force or skip. Replaces `adam.decayFilter` for the common case. */
-    decay?: boolean;
-}
-type InitFn = (size: number, shape: readonly number[]) => Float32Array;
-export declare abstract class Module {
-    /**
-     * Declare a learnable parameter at this module. Must be called from inside
-     * the constructor (typically as a field assignment). Returns a placeholder
-     * that gets replaced with a real Tensor at compile time.
-     *
-     * The parameter's name is auto-derived from its property path in the model
-     * tree (e.g. `layers.0.attn.W_q`). Init metadata travels with the param;
-     * call `compiled.uploadInitialParams()` to apply it after compile.
-     */
-    protected param(shape: Shape, opts?: ParamOptions): Tensor;
-}
-export interface MaterializedParams {
-    /** Map from auto-derived path (e.g. `layers.0.attn.W_q`) to its Tensor. */
-    tensors: Record<string, Tensor>;
-    /** Init function per param path. Used by `uploadInitialParams`. */
-    initFns: Record<string, InitFn>;
-    /** Whether this param should receive AdamW weight decay. Resolved at
-     *  `param()` time from `ParamOptions.decay` (with init-based default). */
-    decayFlags: Record<string, boolean>;
-}
-/**
- * Walk the module tree and replace every ParamSentinel with a real Tensor
- * created via `paramInput(autoName, ...)`. Must be called inside an active
- * trace context (paramInput appends to the current graph).
- *
- * Returns the param tensors keyed by path, plus init functions for use by
- * `uploadInitialParams`.
- */
-export declare function materializeParams(root: Module): MaterializedParams;
-export {};
-//# sourceMappingURL=module.d.ts.map

package/dist/module.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"module.d.ts","sourceRoot":"","sources":["../src/module.ts"],"names":[],"mappings":"AA2BA,OAAO,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,SAAS,CAAA;AAOnD;;;;;;;GAOG;AACH,MAAM,MAAM,QAAQ,GAChB,OAAO,GACP,OAAO,GACP,MAAM,GACN,CAAC,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,MAAM,EAAE,KAAK,YAAY,CAAC,CAAA;AAE9D,MAAM,WAAW,YAAY;IAC3B,KAAK,CAAC,EAAE,KAAK,CAAA;IACb,qCAAqC;IACrC,IAAI,CAAC,EAAE,QAAQ,CAAA;IACf,uEAAuE;IACvE,KAAK,CAAC,EAAE,MAAM,CAAA;IACd;;;6EAGyE;IACzE,KAAK,CAAC,EAAE,OAAO,CAAA;CAChB;AAED,KAAK,MAAM,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,MAAM,EAAE,KAAK,YAAY,CAAA;AAuDtE,8BAAsB,MAAM;IAC1B;;;;;;;;OAQG;IACH,SAAS,CAAC,KAAK,CAAC,KAAK,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,YAAY,GAAG,MAAM;CAK3D;AAMD,MAAM,WAAW,kBAAkB;IACjC,2EAA2E;IAC3E,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC/B,mEAAmE;IACnE,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAC/B;8EAC0E;IAC1E,UAAU,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CACpC;AAED;;;;;;;GAOG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,GAAG,kBAAkB,CAclE"}

package/dist/nn.d.ts

@@ -1,42 +0,0 @@
-import { Module } from './module.js';
-import type { Tensor } from './ir.js';
-import type { Captures } from './runtime.js';
-export interface LinearOptions {
-    /** Include a bias term (default true). */
-    bias?: boolean;
-}
-export declare class Linear extends Module {
-    readonly inDim: number;
-    readonly outDim: number;
-    W: Tensor;
-    b: Tensor | null;
-    constructor(inDim: number, outDim: number, opts?: LinearOptions);
-    fwd(x: Tensor): Tensor;
-}
-export declare class LayerNorm extends Module {
-    readonly d: number;
-    readonly eps: number;
-    g: Tensor;
-    b: Tensor;
-    constructor(d: number, eps?: number);
-    fwd(x: Tensor): Tensor;
-}
-/** [..., T, D] → [..., H, T, D/H]. Folds the standard
- *  `transpose(reshape(x, [..., T, H, d]), [..., H, T, d])` pattern into one
- *  call. Last dim of `x` must divide evenly by `nHeads`. */
-export declare function splitHeads(x: Tensor, nHeads: number): Tensor;
-/** Inverse of `splitHeads`: [..., H, T, d] → [..., T, H*d]. */
-export declare function mergeHeads(x: Tensor): Tensor;
-/** Slice a captured tensor named `name` into one Float32Array per head, using
- *  the static shape registered at compile time. The leading axis is treated as
- *  heads (matching `splitHeads` layout at B=1); a leading singleton batch is
- *  stripped if present so callers can pass capture names directly. Throws if
- *  the capture isn't registered or wasn't read back this call. */
-export declare function unsplitHeads(captures: Captures, name: string): Float32Array[];
-/** Per-position cross-entropy along the last (vocab) axis: returns
- *  `-log p(target)` at each position. `logits` is `[..., V]`; `targets` is
- *  `[...]` of i32; result is `[...]` (one rank less than logits). The user
- *  applies their own masking + reduction downstream — useful when only some
- *  positions contribute (e.g. result-digit masking) or for label smoothing. */
-export declare function crossEntropyLast(logits: Tensor, targets: Tensor): Tensor;
-//# sourceMappingURL=nn.d.ts.map

package/dist/nn.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"nn.d.ts","sourceRoot":"","sources":["../src/nn.ts"],"names":[],"mappings":"AAaA,OAAO,EAAE,MAAM,EAAE,MAAM,aAAa,CAAA;AACpC,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,CAAA;AAIrC,OAAO,KAAK,EAAE,QAAQ,EAAE,MAAM,cAAc,CAAA;AAM5C,MAAM,WAAW,aAAa;IAC5B,0CAA0C;IAC1C,IAAI,CAAC,EAAE,OAAO,CAAA;CACf;AAED,qBAAa,MAAO,SAAQ,MAAM;aAGJ,KAAK,EAAE,MAAM;aAAkB,MAAM,EAAE,MAAM;IAFzE,CAAC,EAAE,MAAM,CAAA;IACT,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;gBACY,KAAK,EAAE,MAAM,EAAkB,MAAM,EAAE,MAAM,EAAE,IAAI,GAAE,aAAkB;IAKnG,GAAG,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM;CAIvB;AAMD,qBAAa,SAAU,SAAQ,MAAM;aAGP,CAAC,EAAE,MAAM;aAAkB,GAAG,EAAE,MAAM;IAFlE,CAAC,EAAE,MAAM,CAAA;IACT,CAAC,EAAE,MAAM,CAAA;gBACmB,CAAC,EAAE,MAAM,EAAkB,GAAG,GAAE,MAAa;IAKzE,GAAG,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM;CAOvB;AAOD;;4DAE4D;AAC5D,wBAAgB,UAAU,CAAC,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,CAa5D;AAED,+DAA+D;AAC/D,wBAAgB,UAAU,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAW5C;AAED;;;;kEAIkE;AAClE,wBAAgB,YAAY,CAAC,QAAQ,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,GAAG,YAAY,EAAE,CAiB7E;AAMD;;;;+EAI+E;AAC/E,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CASxE"}

package/dist/ops.d.ts

@@ -1,48 +0,0 @@
-import type { Tensor, Shape, Dtype } from './ir.js';
-export declare function add(a: Tensor, b: Tensor | number): Tensor;
-export declare function sub(a: Tensor, b: Tensor | number): Tensor;
-export declare function mul(a: Tensor, b: Tensor | number): Tensor;
-export declare function div(a: Tensor, b: Tensor | number): Tensor;
-export declare function mulScalar(a: Tensor, scalar: number): Tensor;
-export declare function addScalar(a: Tensor, scalar: number): Tensor;
-export declare const sqrt: (a: Tensor) => Tensor;
-export declare const rsqrt: (a: Tensor) => Tensor;
-export declare const log: (a: Tensor) => Tensor;
-export declare const exp: (a: Tensor) => Tensor;
-export declare const relu: (a: Tensor) => Tensor;
-export declare function meanLast(a: Tensor): Tensor;
-export declare function sumLast(a: Tensor): Tensor;
-/** Reduce all elements to a 0-d scalar. Composes `reshape` + `sumLast`. */
-export declare function sumAll(a: Tensor): Tensor;
-export declare function reshape(a: Tensor, newShape: Shape): Tensor;
-export declare function transpose(a: Tensor, perm: readonly number[]): Tensor;
-/** Swap two axes of a tensor. Negative indices count from the end (so
- *  `swapAxes(x, -1, -2)` swaps the last two — the common attention pattern).
- *  All other axes keep their position. Implemented as `transpose` with the
- *  permutation `[0, 1, ..., axis2, ..., axis1, ..., n-1]`. */
-export declare function swapAxes(a: Tensor, axis1: number, axis2: number): Tensor;
-export declare function matmul(a: Tensor, b: Tensor): Tensor;
-export declare function matmulBatched(a: Tensor, b: Tensor): Tensor;
-export declare function oneHot(indices: Tensor, depth: number, dtype?: Dtype): Tensor;
-/** Embedding lookup: pull rows from `table` indexed by `indices`. Decomposes
- *  to `oneHot(indices, vocab) @ table` so autograd works without a dedicated
- *  scatter-with-atomic-add backward — the matmul transpose rule handles it.
- *  `table` is `[vocab, dim]`; `indices` is any shape `[...]` of i32; result
- *  is `[..., dim]`. The vocab size is taken from `table.shape[0]`. */
-export declare function embedding(table: Tensor, indices: Tensor): Tensor;
-export declare function arange(n: number, dtype?: Dtype): Tensor;
-export declare function softmaxCausalLast(a: Tensor): Tensor;
-export declare function logSoftmaxLast(a: Tensor): Tensor;
-export declare function whereCausal(a: Tensor, fillValue: number): Tensor;
-export declare function sliceLastRange(a: Tensor, start: number, end: number): Tensor;
-export declare function broadcastTo(a: Tensor, targetShape: Shape): Tensor;
-export declare function sumToShape(a: Tensor, targetShape: Shape): Tensor;
-export declare function constScalar(value: number, dtype?: Dtype): Tensor;
-export declare const less: (a: Tensor, b: Tensor) => Tensor;
-export declare const greater: (a: Tensor, b: Tensor) => Tensor;
-export declare function where(cond: Tensor, a: Tensor, b: Tensor): Tensor;
-export declare function reluGrad(x: Tensor, dy: Tensor): Tensor;
-export declare function adamUpdateM(m: Tensor, g: Tensor, b1: number): Tensor;
-export declare function adamUpdateV(v: Tensor, g: Tensor, b2: number): Tensor;
-export declare function adamUpdateP(p: Tensor, mNew: Tensor, vNew: Tensor, lrt: Tensor, eps: number, decayShrink?: number | Tensor): Tensor;
-//# sourceMappingURL=ops.d.ts.map

package/dist/ops.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"ops.d.ts","sourceRoot":"","sources":["../src/ops.ts"],"names":[],"mappings":"AAWA,OAAO,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAU,MAAM,SAAS,CAAA;AAmC3D,wBAAgB,GAAG,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAEzD;AACD,wBAAgB,GAAG,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAEzD;AACD,wBAAgB,GAAG,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAEzD;AACD,wBAAgB,GAAG,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,MAAM,CAMzD;AAQD,wBAAgB,SAAS,CAAC,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,CAG3D;AAED,wBAAgB,SAAS,CAAC,CAAC,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,MAAM,CAG3D;AAYD,eAAO,MAAM,IAAI,GAAK,GAAG,MAAM,KAAG,MAA2B,CAAA;AAC7D,eAAO,MAAM,KAAK,GAAI,GAAG,MAAM,KAAG,MAA2B,CAAA;AAC7D,eAAO,MAAM,GAAG,GAAM,GAAG,MAAM,KAAG,MAA2B,CAAA;AAC7D,eAAO,MAAM,GAAG,GAAM,GAAG,MAAM,KAAG,MAA2B,CAAA;AAC7D,eAAO,MAAM,IAAI,GAAK,GAAG,MAAM,KAAG,MAA2B,CAAA;AAO7D,wBAAgB,QAAQ,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAK1C;AAED,wBAAgB,OAAO,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAKzC;AAED,2EAA2E;AAC3E,wBAAgB,MAAM,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAExC;AAMD,wBAAgB,OAAO,CAAC,CAAC,EAAE,MAAM,EAAE,QAAQ,EAAE,KAAK,GAAG,MAAM,CAI1D;AAED,wBAAgB,SAAS,CAAC,CAAC,EAAE,MAAM,EAAE,IAAI,EAAE,SAAS,MAAM,EAAE,GAAG,MAAM,CAIpE;AAED;;;8DAG8D;AAC9D,wBAAgB,QAAQ,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,GAAG,MAAM,CAcxE;AAMD,wBAAgB,MAAM,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAOnD;AAED,wBAAgB,aAAa,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAO1D;AAMD,wBAAgB,MAAM,CAAC,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,GAAE,KAAa,GAAG,MAAM,CAOnF;AAED;;;;sEAIsE;AACtE,wBAAgB,SAAS,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,MAAM,CAShE;AAGD,wBAAgB,MAAM,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,GAAE,KAAa,GAAG,MAAM,CAM9D;AASD,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAKnD;AAGD,wBAAgB,cAAc,CAAC,CAAC,EAAE,MAAM,GAAG,MAAM,CAIhD;AAMD,wBAAgB,WAAW,CAAC,CAAC,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,GAAG,MAAM,CAKhE;AAQD,wBAAgB,cAAc,CAAC,CAAC,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,GAAG,MAAM,CAI5E;AAOD,wBAAgB,WAAW,CAAC,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,KAAK,GAAG,MAAM,CAIjE;AAED,wBAAgB,UAAU,CAAC,CAAC,EAAE,MAAM,EAAE,WAAW,EAAE,KAAK,GAAG,MAAM,CAIhE;AAOD,wBAAgB,WAAW,CAAC,KAAK,EAAE,MAAM,EAAE,KAAK,GAAE,KAAa,GAAG,MAAM,CAGvE;AAWD,eAAO,MAAM,IAAI,GAAO,GAAG,MAAM,EAAE,GAAG,MAAM,KAAG,MAAqD,CAAA;AACpG,eAAO,MAAM,OAAO,GAAI,GAAG,MAAM,EAAE,GAAG,MAAM,KAAG,MAAqD,CAAA;AAGpG,wBAAgB,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAMhE;AAID,wBAAgB,QAAQ,CAAC,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,GAAG,MAAM,CAOtD;AAMD,wBAAgB,WAAW,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,GAAG,MAAM,CAOpE;AAED,wBAAgB,WAAW,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,EAAE,EAAE,EAAE,MAAM,GAAG,MAAM,CAOpE;AAED,wBAAgB,WAAW,CACzB,CAAC,EAAE,MAAM,EACT,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,MAAM,EACZ,GAAG,EAAE,MAAM,EACX,GAAG,EAAE,MAAM,EACX,WAAW,GAAE,MAAM,GAAG,MAAU,GAC/B,MAAM,CA2BR"}

package/dist/runtime.d.ts

@@ -1,115 +0,0 @@
-import type { BufferPlan } from './buffers.js';
-import type { KernelSpec } from './codegen.js';
-export interface UploadParamsOptions {
-    /** Skip the "missing param" check, allowing the caller to update only some
-     *  params and leave the rest at their current GPU values. Extra (unknown)
-     *  keys are still rejected — that's always a typo. Default: false. */
-    partial?: boolean;
-}
-/**
- * Activation readbacks for one `step()`/`run()` call. Keyed by the names
- * passed to `capture(name, t)` during the trace. `get(name)` throws if the
- * name isn't registered or wasn't read back this call (i.e., the call was
- * made without `{ withCaptures: true }`); use `has(name)` if you need to
- * branch. `shapeOf(name)` returns the static-after-compile shape and works
- * regardless of whether captures were read back.
- */
-export declare class Captures {
-    private readonly shapes;
-    private readonly data;
-    constructor(shapes: Record<string, readonly number[]>, data: Map<string, Float32Array>);
-    get(name: string): Float32Array;
-    shapeOf(name: string): readonly number[];
-    has(name: string): boolean;
-    names(): string[];
-}
-export interface RunResult {
-    output: Float32Array;
-    captures: Captures;
-}
-export interface StepResult {
-    loss: number;
-    captures: Captures;
-}
-export interface RunOptions {
-    /** Read back tensors registered via `capture(name, t)` during the trace.
-     *  Default false. When false, the returned `captures` is empty (calling
-     *  `.get` throws); when true, captures are read back and accessible. */
-    withCaptures?: boolean;
-}
-/** Common surface for both training and forward-only compiled runtimes. */
-export interface CompiledBase {
-    /** The GPUDevice this runtime is bound to. Pass to sibling compiles to
-     *  share the device, or use directly for other GPU work. */
-    device: GPUDevice;
-    /** Param name -> the underlying GPUBuffer. Pass to a sibling compile via
-     *  `sharedParams` to share without copies. */
-    params: Map<string, GPUBuffer>;
-    /** Shape of the graph's output (loss scalar `[]` for training; the user's
-     *  returned tensor for forward-only compiles). */
-    outputShape: number[];
-    /** Upload parameter Float32Arrays to their GPU buffers. By default, requires
-     *  *all* params to be present; throws on any unknown or missing key. Pass
-     *  `{ partial: true }` to skip the missing-key check. */
-    uploadParams(params: Record<string, Float32Array>, opts?: UploadParamsOptions): void;
-    /** Read all parameters back as Float32Arrays — used for UI panels. */
-    downloadParams(): Promise<Record<string, Float32Array>>;
-    /** Free GPU resources. */
-    destroy(): void;
-}
-/** Run a dispatch and read back the full output tensor. Default returns the
- *  output as a `Float32Array`; with `{ withCaptures: true }` returns
- *  `{ output, captures }`. Same shape as `step()`'s overloads. */
-export interface RunFn {
-    (inputs: Record<string, Int32Array | Float32Array>): Promise<Float32Array>;
-    (inputs: Record<string, Int32Array | Float32Array>, opts: {
-        withCaptures: true;
-    }): Promise<RunResult>;
-    (inputs: Record<string, Int32Array | Float32Array>, opts: RunOptions): Promise<Float32Array | RunResult>;
-}
-export interface CompiledRuntime extends CompiledBase {
-    /** Read all parameter gradients back. Mostly for verification / debugging. */
-    downloadParamGrads(): Promise<Record<string, Float32Array>>;
-    /**
-     * One full forward+backward step.
-     *   1. Uploads `inputs` (tokens, targets, masks) to input buffers.
-     *   2. Dispatches every kernel in order.
-     *   3. Reads back the loss scalar (and any registered captures, if requested).
-     * Default returns the loss as a JS number; with `{ withCaptures: true }`
-     * returns `{ loss, captures }`.
-     */
-    step(inputs: Record<string, Int32Array | Float32Array>): Promise<number>;
-    step(inputs: Record<string, Int32Array | Float32Array>, opts: {
-        withCaptures: true;
-    }): Promise<StepResult>;
-    step(inputs: Record<string, Int32Array | Float32Array>, opts: RunOptions): Promise<number | StepResult>;
-    /** Same dispatch as step() but returns the full output Float32Array — for
-     *  training graphs the output is a scalar loss, so step() is usually more
-     *  convenient. Provided for parity with `compileForward`. */
-    run: RunFn;
-    /** Re-zero all optimizer state buffers (Adam's m/v) in place. Pair with
-     *  `uploadInitialParams()` for a full training reset without recompile. */
-    resetOptimizerState(): void;
-}
-/** Forward-only compiled runtime — produced by `compileForward`. No optimizer,
- *  no backward. Returns the output tensor (not just a scalar) per `run()` call. */
-export interface CompiledForward extends CompiledBase {
-    run: RunFn;
-}
-export interface RuntimeOpts {
-    /** Pre-acquired GPUDevice. If omitted, runtime requests its own. */
-    device?: GPUDevice;
-    /** External param buffers to bind in place of allocating fresh ones, keyed
-     *  by param name. Used to share params between a training compile and a
-     *  sibling forward-only compile (e.g., a B=1 inference graph). When a name
-     *  is in this map, the runtime reuses the provided GPUBuffer; otherwise it
-     *  allocates as usual. */
-    sharedParams?: Map<string, GPUBuffer>;
-}
-export declare function createRuntime(plan: BufferPlan, kernels: KernelSpec[], lossBufferId: number, opts?: RuntimeOpts): Promise<CompiledRuntime>;
-/** Same machinery as `createRuntime`, narrower public type: a forward-only
- *  graph exposes `run()` instead of `step()` (no optimizer state, no scalar-
- *  loss readback). The full runtime object is built once and projected by
- *  `compileForward` to the public shape. */
-export declare function createForwardRuntime(plan: BufferPlan, kernels: KernelSpec[], outputBufferId: number, opts?: RuntimeOpts): Promise<CompiledForward>;
-//# sourceMappingURL=runtime.d.ts.map

package/dist/runtime.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"runtime.d.ts","sourceRoot":"","sources":["../src/runtime.ts"],"names":[],"mappings":"AAMA,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,cAAc,CAAA;AAC9C,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,cAAc,CAAA;AAM9C,MAAM,WAAW,mBAAmB;IAClC;;0EAEsE;IACtE,OAAO,CAAC,EAAE,OAAO,CAAA;CAClB;AAED;;;;;;;GAOG;AACH,qBAAa,QAAQ;IAEjB,OAAO,CAAC,QAAQ,CAAC,MAAM;IACvB,OAAO,CAAC,QAAQ,CAAC,IAAI;gBADJ,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAC,EACzC,IAAI,EAAE,GAAG,CAAC,MAAM,EAAE,YAAY,CAAC;IAElD,GAAG,CAAC,IAAI,EAAE,MAAM,GAAG,YAAY;IAS/B,OAAO,CAAC,IAAI,EAAE,MAAM,GAAG,SAAS,MAAM,EAAE;IAQxC,GAAG,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO;IAC1B,KAAK,IAAI,MAAM,EAAE;CAClB;AAED,MAAM,WAAW,SAAS;IACxB,MAAM,EAAE,YAAY,CAAA;IACpB,QAAQ,EAAE,QAAQ,CAAA;CACnB;AAED,MAAM,WAAW,UAAU;IACzB,IAAI,EAAE,MAAM,CAAA;IACZ,QAAQ,EAAE,QAAQ,CAAA;CACnB;AAED,MAAM,WAAW,UAAU;IACzB;;4EAEwE;IACxE,YAAY,CAAC,EAAE,OAAO,CAAA;CACvB;AAED,2EAA2E;AAC3E,MAAM,WAAW,YAAY;IAC3B;gEAC4D;IAC5D,MAAM,EAAE,SAAS,CAAA;IACjB;kDAC8C;IAC9C,MAAM,EAAE,GAAG,CAAC,MAAM,EAAE,SAAS,CAAC,CAAA;IAC9B;sDACkD;IAClD,WAAW,EAAE,MAAM,EAAE,CAAA;IACrB;;6DAEyD;IACzD,YAAY,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,YAAY,CAAC,EAAE,IAAI,CAAC,EAAE,mBAAmB,GAAG,IAAI,CAAA;IACpF,sEAAsE;IACtE,cAAc,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,CAAA;IACvD,0BAA0B;IAC1B,OAAO,IAAI,IAAI,CAAA;CAChB;AAED;;kEAEkE;AAClE,MAAM,WAAW,KAAK;IACpB,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,GAAG,OAAO,CAAC,YAAY,CAAC,CAAA;IAC1E,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,EAAE,IAAI,EAAE;QAAE,YAAY,EAAE,IAAI,CAAA;KAAE,GAAG,OAAO,CAAC,SAAS,CAAC,CAAA;IACrG,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,EAAE,IAAI,EAAE,UAAU,GAAG,OAAO,CAAC,YAAY,GAAG,SAAS,CAAC,CAAA;CACzG;AAED,MAAM,WAAW,eAAgB,SAAQ,YAAY;IACnD,8EAA8E;IAC9E,kBAAkB,IAAI,OAAO,CAAC,MAAM,CAAC,MAAM,EAAE,YAAY,CAAC,CAAC,CAAA;IAC3D;;;;;;;OAOG;IACH,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,GAAG,OAAO,CAAC,MAAM,CAAC,CAAA;IACxE,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,EAAE,IAAI,EAAE;QAAE,YAAY,EAAE,IAAI,CAAA;KAAE,GAAG,OAAO,CAAC,UAAU,CAAC,CAAA;IAC1G,IAAI,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,YAAY,CAAC,EAAE,IAAI,EAAE,UAAU,GAAG,OAAO,CAAC,MAAM,GAAG,UAAU,CAAC,CAAA;IACvG;;iEAE6D;IAC7D,GAAG,EAAE,KAAK,CAAA;IACV;+EAC2E;IAC3E,mBAAmB,IAAI,IAAI,CAAA;CAC5B;AAED;mFACmF;AACnF,MAAM,WAAW,eAAgB,SAAQ,YAAY;IACnD,GAAG,EAAE,KAAK,CAAA;CACX;AAED,MAAM,WAAW,WAAW;IAC1B,oEAAoE;IACpE,MAAM,CAAC,EAAE,SAAS,CAAA;IAClB;;;;8BAI0B;IAC1B,YAAY,CAAC,EAAE,GAAG,CAAC,MAAM,EAAE,SAAS,CAAC,CAAA;CACtC;AAQD,wBAAsB,aAAa,CACjC,IAAI,EAAE,UAAU,EAChB,OAAO,EAAE,UAAU,EAAE,EACrB,YAAY,EAAE,MAAM,EACpB,IAAI,GAAE,WAAgB,GACrB,OAAO,CAAC,eAAe,CAAC,CA4V1B;AAED;;;4CAG4C;AAC5C,wBAAsB,oBAAoB,CACxC,IAAI,EAAE,UAAU,EAChB,OAAO,EAAE,UAAU,EAAE,EACrB,cAAc,EAAE,MAAM,EACtB,IAAI,GAAE,WAAgB,GACrB,OAAO,CAAC,eAAe,CAAC,CAE1B"}

package/dist/shape.d.ts

@@ -1,24 +0,0 @@
-import type { Shape, CallSite } from './ir.js';
-export declare class ShapeError extends Error {
-    constructor(message: string, site: CallSite | null);
-}
-export declare function shapesEqual(a: Shape, b: Shape): boolean;
-export declare function shapeSize(shape: Shape): number;
-export declare function showShape(shape: Shape): string;
-export declare function broadcastTrailing(a: Shape, b: Shape): Shape | null;
-export declare function inferElementwiseBinop(opName: string, aShape: Shape, bShape: Shape, site: CallSite | null): Shape;
-export declare function inferUnary(_opName: string, aShape: Shape, _site: CallSite | null): Shape;
-export declare function inferMeanLast(opName: string, aShape: Shape, site: CallSite | null): Shape;
-export declare function inferSumLast(opName: string, aShape: Shape, site: CallSite | null): Shape;
-export declare function inferReshape(opName: string, aShape: Shape, newShape: Shape, site: CallSite | null): Shape;
-export declare function inferTranspose(opName: string, aShape: Shape, perm: readonly number[], site: CallSite | null): Shape;
-export declare function inferMatmul(opName: string, aShape: Shape, bShape: Shape, site: CallSite | null): Shape;
-export declare function inferMatmulBatched(opName: string, aShape: Shape, bShape: Shape, site: CallSite | null): Shape;
-export declare function inferOneHot(opName: string, indicesShape: Shape, depth: number, site: CallSite | null): Shape;
-export declare function inferWhereCausal(opName: string, aShape: Shape, site: CallSite | null): Shape;
-export declare function inferSliceLastRange(opName: string, aShape: Shape, start: number, end: number, site: CallSite | null): Shape;
-export declare function inferBroadcastTo(opName: string, aShape: Shape, targetShape: Shape, site: CallSite | null): Shape;
-export declare function inferSumToShape(opName: string, aShape: Shape, targetShape: Shape, site: CallSite | null): Shape;
-export declare function inferWhere(opName: string, condShape: Shape, aShape: Shape, bShape: Shape, site: CallSite | null): Shape;
-export declare function inferReluGrad(opName: string, xShape: Shape, dyShape: Shape, site: CallSite | null): Shape;
-//# sourceMappingURL=shape.d.ts.map

package/dist/shape.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"shape.d.ts","sourceRoot":"","sources":["../src/shape.ts"],"names":[],"mappings":"AAkBA,OAAO,KAAK,EAAE,KAAK,EAAE,QAAQ,EAAE,MAAM,SAAS,CAAA;AAO9C,qBAAa,UAAW,SAAQ,KAAK;gBACvB,OAAO,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI;CAKnD;AAUD,wBAAgB,WAAW,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,GAAG,OAAO,CAIvD;AAED,wBAAgB,SAAS,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAI9C;AAED,wBAAgB,SAAS,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE9C;AAKD,wBAAgB,iBAAiB,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,GAAG,KAAK,GAAG,IAAI,CAclE;AASD,wBAAgB,qBAAqB,CACnC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAClE,KAAK,CAWP;AAED,wBAAgB,UAAU,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAExF;AAED,wBAAgB,aAAa,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAIzF;AAED,wBAAgB,YAAY,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAIxF;AAED,wBAAgB,YAAY,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,QAAQ,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CA0BzG;AAED,wBAAgB,cAAc,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,SAAS,MAAM,EAAE,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAWnH;AAGD,wBAAgB,WAAW,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAStG;AAGD,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAoB7G;AAED,wBAAgB,WAAW,CAAC,MAAM,EAAE,MAAM,EAAE,YAAY,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAG5G;AAGD,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAM5F;AAED,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,MAAM,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAO3H;AAID,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,WAAW,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAahH;AAID,wBAAgB,eAAe,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,WAAW,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAa/G;AAID,wBAAgB,UAAU,CAAC,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAMvH;AAED,wBAAgB,aAAa,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,EAAE,QAAQ,GAAG,IAAI,GAAG,KAAK,CAKzG"}

package/dist/trace.d.ts

@@ -1,9 +0,0 @@
-import type { Graph, Tensor, Shape, Dtype } from './ir.js';
-export declare function currentGraph(): Graph;
-export declare function isCaptureEnabled(): boolean;
-export declare function trace(fn: () => Tensor | Tensor[]): Graph;
-export declare function traceInto<T>(g: Graph, fn: () => T): T;
-export declare function paramInput(name: string, shape: Shape, dtype?: Dtype): Tensor;
-export declare function tensorInput(name: string, shape: Shape, dtype?: Dtype): Tensor;
-export declare function stateInput(name: string, shape: Shape, dtype?: Dtype, initValue?: number): Tensor;
-//# sourceMappingURL=trace.d.ts.map

package/dist/trace.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"trace.d.ts","sourceRoot":"","sources":["../src/trace.ts"],"names":[],"mappings":"AAgBA,OAAO,KAAK,EAAE,KAAK,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,SAAS,CAAA;AAU1D,wBAAgB,YAAY,IAAI,KAAK,CAQpC;AAED,wBAAgB,gBAAgB,IAAI,OAAO,CAE1C;AAID,wBAAgB,KAAK,CAAC,EAAE,EAAE,MAAM,MAAM,GAAG,MAAM,EAAE,GAAG,KAAK,CAkBxD;AAQD,wBAAgB,SAAS,CAAC,CAAC,EAAE,CAAC,EAAE,KAAK,EAAE,EAAE,EAAE,MAAM,CAAC,GAAG,CAAC,CAWrD;AAOD,wBAAgB,UAAU,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,GAAE,KAAa,GAAG,MAAM,CAOnF;AAED,wBAAgB,WAAW,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,GAAE,KAAa,GAAG,MAAM,CAOpF;AAID,wBAAgB,UAAU,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,KAAK,EAAE,KAAK,GAAE,KAAa,EAAE,SAAS,SAAI,GAAG,MAAM,CAOlG"}