tensorgrad 0.0.11 → 0.0.12

package/src/runtime.ts

@@ -1,515 +1,523 @@
-// WebGPU runtime. Reads a BufferPlan + KernelSpec[] (produced by codegen),
-// allocates real GPU buffers and pipelines, and provides a `step()` method
-// that uploads inputs, dispatches all kernels, and reads back outputs.
-//
-// Browser-only: this module needs `navigator.gpu` at runtime.
-
-import type { BufferPlan } from './buffers.js'
-import type { KernelSpec } from './codegen.js'
-
-// TS lib.dom defines WebGPU types but not the GPUMapMode runtime constant.
-// Provided by the browser per WebGPU spec; declare just what we use.
-declare const GPUMapMode: { readonly READ: number; readonly WRITE: number }
-
-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 class Captures {
-  constructor(
-    private readonly shapes: Record<string, readonly number[]>,
-    private readonly data: Map<string, Float32Array>,
-  ) {}
-  get(name: string): Float32Array {
-    const d = this.data.get(name)
-    if (!d) {
-      const known = [...this.data.keys()].sort().join(', ')
-      const detail = known ? `Known this call: ${known}` : `(call run/step with { withCaptures: true } to populate)`
-      throw new Error(`Captures.get: '${name}' not present. ${detail}`)
-    }
-    return d
-  }
-  shapeOf(name: string): readonly number[] {
-    const s = this.shapes[name]
-    if (!s) {
-      const known = Object.keys(this.shapes).sort().join(', ') || '(none registered)'
-      throw new Error(`Captures.shapeOf: '${name}' not registered. Known: ${known}`)
-    }
-    return s
-  }
-  has(name: string): boolean { return this.data.has(name) }
-  names(): string[] { return [...this.data.keys()].sort() }
-}
-
-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. Captures are always
- *  returned; their data is empty unless `{ withCaptures: true }` is passed. */
-export type RunFn = (
-  inputs: Record<string, Int32Array | Float32Array>,
-  opts?: RunOptions,
-) => Promise<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>
-}
-
-// Inlined numeric values (per WebGPU spec) so this module is importable in Node
-// for codegen-only usage. The browser provides GPUBufferUsage as a global, but
-// referencing it at module scope would crash before any browser code runs.
-const STORAGE_RW = 0x80 /*STORAGE*/ | 0x8 /*COPY_DST*/ | 0x4 /*COPY_SRC*/
-const READBACK = 0x1 /*MAP_READ*/ | 0x8 /*COPY_DST*/
-
-export async function createRuntime(
-  plan: BufferPlan,
-  kernels: KernelSpec[],
-  lossBufferId: number,
-  opts: RuntimeOpts = {},
-): Promise<CompiledRuntime> {
-  const device = opts.device ?? await acquireDevice()
-  const queue = device.queue
-
-  // ---- Allocate one GPUBuffer per BufferSpec --------------------------------
-  // State buffers also get filled with their initValue at allocation time.
-  // Param buffers may be supplied externally via opts.sharedParams; in that
-  // case we reuse the provided GPUBuffer instead of allocating, and the
-  // sibling compile that owns it is responsible for upload + lifetime.
-  const buffers = new Map<number, GPUBuffer>()
-  const sharedParams = opts.sharedParams
-  for (const spec of plan.buffers) {
-    if (spec.kind === 'param' && sharedParams?.has(spec.name!)) {
-      const shared = sharedParams.get(spec.name!)!
-      if (shared.size !== spec.byteSize) {
-        throw new Error(
-          `sharedParams: size mismatch for '${spec.name}' — supplied ${shared.size} bytes, ` +
-          `compiled graph expects ${spec.byteSize}.`,
-        )
-      }
-      buffers.set(spec.id, shared)
-      continue
-    }
-    const buf = device.createBuffer({
-      size: spec.byteSize,
-      usage: STORAGE_RW,
-      label: spec.name ?? `t${spec.id}-${spec.kind}`,
-    })
-    buffers.set(spec.id, buf)
-    if (spec.kind === 'state') fillStateBuffer(spec, buf)
-  }
-  // Track which params are externally owned — those are skipped on destroy().
-  const ownedBufferIds = new Set<number>()
-  for (const spec of plan.buffers) {
-    if (!(spec.kind === 'param' && sharedParams?.has(spec.name!))) {
-      ownedBufferIds.add(spec.id)
-    }
-  }
-
-  // ---- Compile pipelines per kernel; cache by WGSL source -------------------
-  // Push an error scope around each shader+pipeline creation so we can surface
-  // the actual compile error rather than the cryptic "previous error" that
-  // comes from using an invalid pipeline at dispatch time.
-  const moduleCache = new Map<string, GPUShaderModule>()
-  const pipelines: (GPUComputePipeline | null)[] = []
-  type ErrorProbe = Promise<{ k: KernelSpec; module: GPUShaderModule; err: GPUError } | null>
-  const probes: ErrorProbe[] = []
-  for (const k of kernels) {
-    if (!k.wgsl) { pipelines.push(null); continue }
-    let module = moduleCache.get(k.wgsl)
-    if (!module) {
-      module = device.createShaderModule({ code: k.wgsl, label: k.opKind })
-      moduleCache.set(k.wgsl, module)
-    }
-    device.pushErrorScope('validation')
-    const pipeline = device.createComputePipeline({
-      layout: 'auto',
-      compute: { module, entryPoint: 'main' },
-      label: k.opKind,
-    })
-    pipelines.push(pipeline)
-    probes.push(device.popErrorScope().then(err => err ? { k, module: module!, err } : null))
-  }
-  const probeResults = await Promise.all(probes)
-  const failures = probeResults.filter((p): p is { k: KernelSpec; module: GPUShaderModule; err: GPUError } => p != null)
-  if (failures.length > 0) {
-    const reports: string[] = []
-    for (const { k, module, err } of failures) {
-      const info = await module.getCompilationInfo()
-      const messages = info.messages
-        .map(m => `  L${m.lineNum}:${m.linePos} [${m.type}] ${m.message}`)
-        .join('\n')
-      reports.push(
-        `[shader compile error] ${k.opKind} (op #${k.opIndex}): ${err.message}\n` +
-        (messages || '  (no compilation messages)') +
-        `\n--- WGSL ---\n${k.wgsl}\n-----------`,
-      )
-    }
-    // eslint-disable-next-line no-console
-    console.error(reports.join('\n\n'))
-    throw new Error(`tensorgrad: ${failures.length} shader(s) failed to compile (see console).`)
-  }
-
-  // ---- Pre-build bind groups (static — buffer ids don't change per step) ---
-  const bindGroups: (GPUBindGroup | null)[] = kernels.map((k, i) => {
-    const pipeline = pipelines[i]
-    if (!pipeline) return null
-    return device.createBindGroup({
-      layout: pipeline.getBindGroupLayout(0),
-      entries: k.bindings.map((bufId, idx) => ({
-        binding: idx,
-        resource: { buffer: buffers.get(bufId)! },
-      })),
-    })
-  })
-
-  // ---- Output readback staging buffer ---------------------------------------
-  // `outputBufferId` is the graph's main output (loss for training, the user's
-  // returned tensor for forward-only). step() reads back its first element;
-  // run() reads back the full Float32Array.
-  const outputSpec = plan.buffers[lossBufferId]!
-  const outputReadback = device.createBuffer({ size: outputSpec.byteSize, usage: READBACK })
-
-  // ---- Capture readback staging buffers (lazy) ------------------------------
-  // Allocated on first `step({ withCaptures: true })` call and reused across
-  // subsequent calls. When the graph has no captures registered or when the
-  // caller never opts in, no extra GPU memory is allocated.
-  let captureStagings: Map<string, GPUBuffer> | null = null
-  function ensureCaptureStagings(): Map<string, GPUBuffer> {
-    if (captureStagings) return captureStagings
-    captureStagings = new Map()
-    for (const [name, bufId] of plan.capturesByName) {
-      const spec = plan.buffers[bufId]!
-      const staging = device.createBuffer({ size: spec.byteSize, usage: READBACK, label: `cap-${name}` })
-      captureStagings.set(name, staging)
-    }
-    return captureStagings
-  }
-
-  // ---- dispatch() — shared core for step() and run() -----------------------
-  // Uploads inputs, dispatches all kernels (in order), queues writebacks, copies
-  // the output buffer into its staging, optionally copies captures into theirs,
-  // submits, and reads back. Returns the full output Float32Array; step() takes
-  // [0] for scalar loss, run() returns it whole.
-  //
-  // **Concurrent calls auto-serialize.** Two `step()`/`run()` calls on the same
-  // runtime would otherwise both try to `mapAsync` the shared output staging
-  // buffer at the same time and trip "Buffer already has an outstanding map
-  // pending." We chain each new dispatch onto the prior one's promise so they
-  // run sequentially even when fired from independent async paths (e.g., a
-  // training loop's auxiliary `refreshPrediction()` + `writeDiagnostic()`).
-  let pending: Promise<unknown> = Promise.resolve()
-  async function dispatch(
-    inputs: Record<string, Int32Array | Float32Array>,
-    wantCaptures: boolean,
-  ): Promise<{ output: Float32Array; captures: Map<string, Float32Array> }> {
-    const turn = pending.catch(() => {}).then(() => dispatchUnsynchronized(inputs, wantCaptures))
-    pending = turn
-    return turn
-  }
-  async function dispatchUnsynchronized(
-    inputs: Record<string, Int32Array | Float32Array>,
-    wantCaptures: boolean,
-  ): Promise<{ output: Float32Array; captures: Map<string, Float32Array> }> {
-    if (wantCaptures && plan.capturesByName.size === 0) {
-      throw new Error(
-        `withCaptures=true but no capture(...) calls were registered during ` +
-        `the trace. Add capture('name', tensor) inside your forward pass for ` +
-        `the intermediates you want read back.`,
-      )
-    }
-    for (const [name, bufId] of plan.inputsByName) {
-      const data = inputs[name]
-      if (!data) throw new Error(`tensorgrad: missing input '${name}'`)
-      const expectedBytes = plan.buffers[bufId]!.byteSize
-      if (data.byteLength !== expectedBytes) {
-        throw new Error(`tensorgrad: input '${name}' has ${data.byteLength} bytes, expected ${expectedBytes}`)
-      }
-      // Cast to BufferSource: typed arrays are accepted by writeBuffer at runtime
-      // but TS may infer ArrayBufferLike (vs ArrayBuffer) under strict configs.
-      queue.writeBuffer(buffers.get(bufId)!, 0, data as unknown as BufferSource)
-    }
-
-    const encoder = device.createCommandEncoder({ label: 'tensorgrad-step' })
-    for (let i = 0; i < kernels.length; i++) {
-      const k = kernels[i]!
-      if (!k.wgsl || k.threads === 0) continue
-      const pipeline = pipelines[i]!
-      const bindGroup = bindGroups[i]!
-      const pass = encoder.beginComputePass({ label: k.opKind })
-      pass.setPipeline(pipeline)
-      pass.setBindGroup(0, bindGroup)
-      // WebGPU caps each dispatch dimension at 65535 workgroups. Split into 2D
-      // when a kernel needs more than that on the X axis. Kernels compute their
-      // global index as `gid.x + gid.y * (65535 * workgroup_size)`, matching the
-      // stride we set here. For dispatches that fit in one row, gid.y is 0.
-      const wgCount = Math.max(1, Math.ceil(k.threads / k.workgroupSize))
-      const MAX_X = 65535
-      const wgX = Math.min(wgCount, MAX_X)
-      const wgY = Math.ceil(wgCount / MAX_X)
-      pass.dispatchWorkgroups(wgX, wgY, 1)
-      pass.end()
-    }
-    // After all dispatches: writebacks (Adam state, updated params). Empty for
-    // forward-only compiles.
-    for (const wb of plan.writebacks) {
-      encoder.copyBufferToBuffer(buffers.get(wb.source)!, 0, buffers.get(wb.dest)!, 0, wb.bytes)
-    }
-    encoder.copyBufferToBuffer(buffers.get(lossBufferId)!, 0, outputReadback, 0, outputSpec.byteSize)
-    // Capture readbacks (only when opted in). Queued before submit so they
-    // observe the same kernel outputs as the main output.
-    let stagings: Map<string, GPUBuffer> | null = null
-    if (wantCaptures) {
-      stagings = ensureCaptureStagings()
-      for (const [name, bufId] of plan.capturesByName) {
-        const spec = plan.buffers[bufId]!
-        encoder.copyBufferToBuffer(buffers.get(bufId)!, 0, stagings.get(name)!, 0, spec.byteSize)
-      }
-    }
-    queue.submit([encoder.finish()])
-
-    await outputReadback.mapAsync(GPUMapMode.READ)
-    const output = new Float32Array(outputReadback.getMappedRange().slice(0))
-    outputReadback.unmap()
-
-    const captures = new Map<string, Float32Array>()
-    if (wantCaptures) {
-      for (const [name, staging] of stagings!) {
-        await staging.mapAsync(GPUMapMode.READ)
-        captures.set(name, new Float32Array(staging.getMappedRange().slice(0)))
-        staging.unmap()
-      }
-    }
-    return { output, captures }
-  }
-
-  // ---- step() — training-mode wrapper, returns scalar [0] of output ---------
-  function step(inputs: Record<string, Int32Array | Float32Array>): Promise<number>
-  function step(inputs: Record<string, Int32Array | Float32Array>, opts: { withCaptures: true }): Promise<StepResult>
-  function step(inputs: Record<string, Int32Array | Float32Array>, opts: RunOptions): Promise<number | StepResult>
-  async function step(
-    inputs: Record<string, Int32Array | Float32Array>,
-    opts?: RunOptions,
-  ): Promise<number | StepResult> {
-    const r = await dispatch(inputs, opts?.withCaptures === true)
-    if (opts?.withCaptures) return { loss: r.output[0]!, captures: new Captures(captureShapes, r.captures) }
-    return r.output[0]!
-  }
-
-  // ---- run() — forward-mode wrapper, returns { output, captures } ----------
-  async function run(
-    inputs: Record<string, Int32Array | Float32Array>,
-    opts?: RunOptions,
-  ): Promise<RunResult> {
-    const r = await dispatch(inputs, opts?.withCaptures === true)
-    return { output: r.output, captures: new Captures(captureShapes, r.captures) }
-  }
-
-  // ---- uploadParams ---------------------------------------------------------
-  function uploadParams(params: Record<string, Float32Array>, opts?: UploadParamsOptions) {
-    const partial = opts?.partial ?? false
-    for (const name of Object.keys(params)) {
-      if (!plan.paramsByName.has(name)) {
-        throw new Error(
-          `uploadParams: unknown param '${name}'. ` +
-          `Known: ${[...plan.paramsByName.keys()].sort().join(', ')}`,
-        )
-      }
-    }
-    if (!partial) {
-      for (const name of plan.paramsByName.keys()) {
-        if (!(name in params)) {
-          throw new Error(
-            `uploadParams: missing param '${name}'. ` +
-            `Pass { partial: true } if you mean to update only some params.`,
-          )
-        }
-      }
-    }
-    for (const [name, bufId] of plan.paramsByName) {
-      const data = params[name]
-      if (!data) continue
-      const expected = plan.buffers[bufId]!.byteSize / 4
-      if (data.length !== expected) {
-        throw new Error(`uploadParams: '${name}' has ${data.length} elements, expected ${expected}`)
-      }
-      queue.writeBuffer(buffers.get(bufId)!, 0, data as unknown as BufferSource)
-    }
-  }
-
-  // ---- download helpers -----------------------------------------------------
-  async function downloadFromMap(map: Map<string, number>): Promise<Record<string, Float32Array>> {
-    const stagings: { name: string; buf: GPUBuffer; bytes: number }[] = []
-    const encoder = device.createCommandEncoder({ label: 'tensorgrad-download' })
-    for (const [name, bufId] of map) {
-      const spec = plan.buffers[bufId]!
-      const staging = device.createBuffer({ size: spec.byteSize, usage: READBACK })
-      encoder.copyBufferToBuffer(buffers.get(bufId)!, 0, staging, 0, spec.byteSize)
-      stagings.push({ name, buf: staging, bytes: spec.byteSize })
-    }
-    queue.submit([encoder.finish()])
-    const out: Record<string, Float32Array> = {}
-    for (const s of stagings) {
-      await s.buf.mapAsync(GPUMapMode.READ)
-      out[s.name] = new Float32Array(s.buf.getMappedRange().slice(0))
-      s.buf.unmap()
-      s.buf.destroy()
-    }
-    return out
-  }
-
-  // Fill a state buffer with its declared initValue (typically 0). Float and
-  // int both serialize to 4 bytes per element. Used at allocation time and on
-  // resetOptimizerState() — same logic, two callers.
-  function fillStateBuffer(spec: { byteSize: number; dtype: 'f32' | 'i32' | 'bool'; initValue?: number }, target: GPUBuffer): void {
-    const elements = spec.byteSize / 4
-    const init = spec.dtype === 'f32'
-      ? new Float32Array(elements).fill(spec.initValue ?? 0)
-      : new Int32Array(elements).fill(Math.trunc(spec.initValue ?? 0))
-    queue.writeBuffer(target, 0, init as unknown as BufferSource)
-  }
-
-  function resetOptimizerState() {
-    for (const spec of plan.buffers) {
-      if (spec.kind === 'state') fillStateBuffer(spec, buffers.get(spec.id)!)
-    }
-  }
-
-  // Build the params map AFTER buffer allocation so it points at the actual
-  // GPUBuffers (shared or freshly allocated).
-  const params = new Map<string, GPUBuffer>()
-  for (const [name, bufId] of plan.paramsByName) {
-    params.set(name, buffers.get(bufId)!)
-  }
-  // Static-after-compile shape metadata so users don't have to recompute
-  // strides to interpret a flat capture readback.
-  const captureShapes: Record<string, number[]> = {}
-  for (const [name, bufId] of plan.capturesByName) {
-    captureShapes[name] = [...plan.buffers[bufId]!.shape]
-  }
-  const outputShape = [...plan.buffers[lossBufferId]!.shape]
-
-  const destroy = () => {
-    for (const [id, b] of buffers) {
-      if (ownedBufferIds.has(id)) b.destroy()
-    }
-    outputReadback.destroy()
-    if (captureStagings) for (const b of captureStagings.values()) b.destroy()
-  }
-
-  return {
-    device,
-    params,
-    outputShape,
-    uploadParams,
-    downloadParams: () => downloadFromMap(plan.paramsByName),
-    downloadParamGrads: () => downloadFromMap(plan.paramGradsByName),
-    step,
-    run,
-    resetOptimizerState,
-    destroy,
-  }
-}
-
-/** 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 async function createForwardRuntime(
-  plan: BufferPlan,
-  kernels: KernelSpec[],
-  outputBufferId: number,
-  opts: RuntimeOpts = {},
-): Promise<CompiledForward> {
-  return await createRuntime(plan, kernels, outputBufferId, opts)
-}
-
-async function acquireDevice(): Promise<GPUDevice> {
-  if (typeof navigator === 'undefined' || !navigator.gpu) {
-    throw new Error('tensorgrad: WebGPU not available in this environment')
-  }
-  const adapter = await navigator.gpu.requestAdapter()
-  if (!adapter) throw new Error('tensorgrad: no WebGPU adapter')
-  return await adapter.requestDevice()
-}
+// WebGPU runtime. Reads a BufferPlan + KernelSpec[] (produced by codegen),
+// allocates real GPU buffers and pipelines, and provides a `step()` method
+// that uploads inputs, dispatches all kernels, and reads back outputs.
+//
+// Browser-only: this module needs `navigator.gpu` at runtime.
+
+import type { BufferPlan } from './buffers.js'
+import type { KernelSpec } from './codegen.js'
+
+// TS lib.dom defines WebGPU types but not the GPUMapMode runtime constant.
+// Provided by the browser per WebGPU spec; declare just what we use.
+declare const GPUMapMode: { readonly READ: number; readonly WRITE: number }
+
+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 class Captures {
+  constructor(
+    private readonly shapes: Record<string, readonly number[]>,
+    private readonly data: Map<string, Float32Array>,
+  ) {}
+  get(name: string): Float32Array {
+    const d = this.data.get(name)
+    if (!d) {
+      const known = [...this.data.keys()].sort().join(', ')
+      const detail = known ? `Known this call: ${known}` : `(call run/step with { withCaptures: true } to populate)`
+      throw new Error(`Captures.get: '${name}' not present. ${detail}`)
+    }
+    return d
+  }
+  shapeOf(name: string): readonly number[] {
+    const s = this.shapes[name]
+    if (!s) {
+      const known = Object.keys(this.shapes).sort().join(', ') || '(none registered)'
+      throw new Error(`Captures.shapeOf: '${name}' not registered. Known: ${known}`)
+    }
+    return s
+  }
+  has(name: string): boolean { return this.data.has(name) }
+  names(): string[] { return [...this.data.keys()].sort() }
+}
+
+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>
+}
+
+// Inlined numeric values (per WebGPU spec) so this module is importable in Node
+// for codegen-only usage. The browser provides GPUBufferUsage as a global, but
+// referencing it at module scope would crash before any browser code runs.
+const STORAGE_RW = 0x80 /*STORAGE*/ | 0x8 /*COPY_DST*/ | 0x4 /*COPY_SRC*/
+const READBACK = 0x1 /*MAP_READ*/ | 0x8 /*COPY_DST*/
+
+export async function createRuntime(
+  plan: BufferPlan,
+  kernels: KernelSpec[],
+  lossBufferId: number,
+  opts: RuntimeOpts = {},
+): Promise<CompiledRuntime> {
+  const device = opts.device ?? await acquireDevice()
+  const queue = device.queue
+
+  // ---- Allocate one GPUBuffer per BufferSpec --------------------------------
+  // State buffers also get filled with their initValue at allocation time.
+  // Param buffers may be supplied externally via opts.sharedParams; in that
+  // case we reuse the provided GPUBuffer instead of allocating, and the
+  // sibling compile that owns it is responsible for upload + lifetime.
+  const buffers = new Map<number, GPUBuffer>()
+  const sharedParams = opts.sharedParams
+  for (const spec of plan.buffers) {
+    if (spec.kind === 'param' && sharedParams?.has(spec.name!)) {
+      const shared = sharedParams.get(spec.name!)!
+      if (shared.size !== spec.byteSize) {
+        throw new Error(
+          `sharedParams: size mismatch for '${spec.name}' — supplied ${shared.size} bytes, ` +
+          `compiled graph expects ${spec.byteSize}.`,
+        )
+      }
+      buffers.set(spec.id, shared)
+      continue
+    }
+    const buf = device.createBuffer({
+      size: spec.byteSize,
+      usage: STORAGE_RW,
+      label: spec.name ?? `t${spec.id}-${spec.kind}`,
+    })
+    buffers.set(spec.id, buf)
+    if (spec.kind === 'state') fillStateBuffer(spec, buf)
+  }
+  // Track which params are externally owned — those are skipped on destroy().
+  const ownedBufferIds = new Set<number>()
+  for (const spec of plan.buffers) {
+    if (!(spec.kind === 'param' && sharedParams?.has(spec.name!))) {
+      ownedBufferIds.add(spec.id)
+    }
+  }
+
+  // ---- Compile pipelines per kernel; cache by WGSL source -------------------
+  // Push an error scope around each shader+pipeline creation so we can surface
+  // the actual compile error rather than the cryptic "previous error" that
+  // comes from using an invalid pipeline at dispatch time.
+  const moduleCache = new Map<string, GPUShaderModule>()
+  const pipelines: (GPUComputePipeline | null)[] = []
+  type ErrorProbe = Promise<{ k: KernelSpec; module: GPUShaderModule; err: GPUError } | null>
+  const probes: ErrorProbe[] = []
+  for (const k of kernels) {
+    if (!k.wgsl) { pipelines.push(null); continue }
+    let module = moduleCache.get(k.wgsl)
+    if (!module) {
+      module = device.createShaderModule({ code: k.wgsl, label: k.opKind })
+      moduleCache.set(k.wgsl, module)
+    }
+    device.pushErrorScope('validation')
+    const pipeline = device.createComputePipeline({
+      layout: 'auto',
+      compute: { module, entryPoint: 'main' },
+      label: k.opKind,
+    })
+    pipelines.push(pipeline)
+    probes.push(device.popErrorScope().then(err => err ? { k, module: module!, err } : null))
+  }
+  const probeResults = await Promise.all(probes)
+  const failures = probeResults.filter((p): p is { k: KernelSpec; module: GPUShaderModule; err: GPUError } => p != null)
+  if (failures.length > 0) {
+    const reports: string[] = []
+    for (const { k, module, err } of failures) {
+      const info = await module.getCompilationInfo()
+      const messages = info.messages
+        .map(m => `  L${m.lineNum}:${m.linePos} [${m.type}] ${m.message}`)
+        .join('\n')
+      reports.push(
+        `[shader compile error] ${k.opKind} (op #${k.opIndex}): ${err.message}\n` +
+        (messages || '  (no compilation messages)') +
+        `\n--- WGSL ---\n${k.wgsl}\n-----------`,
+      )
+    }
+    // eslint-disable-next-line no-console
+    console.error(reports.join('\n\n'))
+    throw new Error(`tensorgrad: ${failures.length} shader(s) failed to compile (see console).`)
+  }
+
+  // ---- Pre-build bind groups (static — buffer ids don't change per step) ---
+  const bindGroups: (GPUBindGroup | null)[] = kernels.map((k, i) => {
+    const pipeline = pipelines[i]
+    if (!pipeline) return null
+    return device.createBindGroup({
+      layout: pipeline.getBindGroupLayout(0),
+      entries: k.bindings.map((bufId, idx) => ({
+        binding: idx,
+        resource: { buffer: buffers.get(bufId)! },
+      })),
+    })
+  })
+
+  // ---- Output readback staging buffer ---------------------------------------
+  // `outputBufferId` is the graph's main output (loss for training, the user's
+  // returned tensor for forward-only). step() reads back its first element;
+  // run() reads back the full Float32Array.
+  const outputSpec = plan.buffers[lossBufferId]!
+  const outputReadback = device.createBuffer({ size: outputSpec.byteSize, usage: READBACK })
+
+  // ---- Capture readback staging buffers (lazy) ------------------------------
+  // Allocated on first `step({ withCaptures: true })` call and reused across
+  // subsequent calls. When the graph has no captures registered or when the
+  // caller never opts in, no extra GPU memory is allocated.
+  let captureStagings: Map<string, GPUBuffer> | null = null
+  function ensureCaptureStagings(): Map<string, GPUBuffer> {
+    if (captureStagings) return captureStagings
+    captureStagings = new Map()
+    for (const [name, bufId] of plan.capturesByName) {
+      const spec = plan.buffers[bufId]!
+      const staging = device.createBuffer({ size: spec.byteSize, usage: READBACK, label: `cap-${name}` })
+      captureStagings.set(name, staging)
+    }
+    return captureStagings
+  }
+
+  // ---- dispatch() — shared core for step() and run() -----------------------
+  // Uploads inputs, dispatches all kernels (in order), queues writebacks, copies
+  // the output buffer into its staging, optionally copies captures into theirs,
+  // submits, and reads back. Returns the full output Float32Array; step() takes
+  // [0] for scalar loss, run() returns it whole.
+  //
+  // **Concurrent calls auto-serialize.** Two `step()`/`run()` calls on the same
+  // runtime would otherwise both try to `mapAsync` the shared output staging
+  // buffer at the same time and trip "Buffer already has an outstanding map
+  // pending." We chain each new dispatch onto the prior one's promise so they
+  // run sequentially even when fired from independent async paths (e.g., a
+  // training loop's auxiliary `refreshPrediction()` + `writeDiagnostic()`).
+  let pending: Promise<unknown> = Promise.resolve()
+  async function dispatch(
+    inputs: Record<string, Int32Array | Float32Array>,
+    wantCaptures: boolean,
+  ): Promise<{ output: Float32Array; captures: Map<string, Float32Array> }> {
+    const turn = pending.catch(() => {}).then(() => dispatchUnsynchronized(inputs, wantCaptures))
+    pending = turn
+    return turn
+  }
+  async function dispatchUnsynchronized(
+    inputs: Record<string, Int32Array | Float32Array>,
+    wantCaptures: boolean,
+  ): Promise<{ output: Float32Array; captures: Map<string, Float32Array> }> {
+    if (wantCaptures && plan.capturesByName.size === 0) {
+      throw new Error(
+        `withCaptures=true but no capture(...) calls were registered during ` +
+        `the trace. Add capture('name', tensor) inside your forward pass for ` +
+        `the intermediates you want read back.`,
+      )
+    }
+    for (const [name, bufId] of plan.inputsByName) {
+      const data = inputs[name]
+      if (!data) throw new Error(`tensorgrad: missing input '${name}'`)
+      const expectedBytes = plan.buffers[bufId]!.byteSize
+      if (data.byteLength !== expectedBytes) {
+        throw new Error(`tensorgrad: input '${name}' has ${data.byteLength} bytes, expected ${expectedBytes}`)
+      }
+      // Cast to BufferSource: typed arrays are accepted by writeBuffer at runtime
+      // but TS may infer ArrayBufferLike (vs ArrayBuffer) under strict configs.
+      queue.writeBuffer(buffers.get(bufId)!, 0, data as unknown as BufferSource)
+    }
+
+    const encoder = device.createCommandEncoder({ label: 'tensorgrad-step' })
+    for (let i = 0; i < kernels.length; i++) {
+      const k = kernels[i]!
+      if (!k.wgsl || k.threads === 0) continue
+      const pipeline = pipelines[i]!
+      const bindGroup = bindGroups[i]!
+      const pass = encoder.beginComputePass({ label: k.opKind })
+      pass.setPipeline(pipeline)
+      pass.setBindGroup(0, bindGroup)
+      // WebGPU caps each dispatch dimension at 65535 workgroups. Split into 2D
+      // when a kernel needs more than that on the X axis. Kernels compute their
+      // global index as `gid.x + gid.y * (65535 * workgroup_size)`, matching the
+      // stride we set here. For dispatches that fit in one row, gid.y is 0.
+      const wgCount = Math.max(1, Math.ceil(k.threads / k.workgroupSize))
+      const MAX_X = 65535
+      const wgX = Math.min(wgCount, MAX_X)
+      const wgY = Math.ceil(wgCount / MAX_X)
+      pass.dispatchWorkgroups(wgX, wgY, 1)
+      pass.end()
+    }
+    // After all dispatches: writebacks (Adam state, updated params). Empty for
+    // forward-only compiles.
+    for (const wb of plan.writebacks) {
+      encoder.copyBufferToBuffer(buffers.get(wb.source)!, 0, buffers.get(wb.dest)!, 0, wb.bytes)
+    }
+    encoder.copyBufferToBuffer(buffers.get(lossBufferId)!, 0, outputReadback, 0, outputSpec.byteSize)
+    // Capture readbacks (only when opted in). Queued before submit so they
+    // observe the same kernel outputs as the main output.
+    let stagings: Map<string, GPUBuffer> | null = null
+    if (wantCaptures) {
+      stagings = ensureCaptureStagings()
+      for (const [name, bufId] of plan.capturesByName) {
+        const spec = plan.buffers[bufId]!
+        encoder.copyBufferToBuffer(buffers.get(bufId)!, 0, stagings.get(name)!, 0, spec.byteSize)
+      }
+    }
+    queue.submit([encoder.finish()])
+
+    await outputReadback.mapAsync(GPUMapMode.READ)
+    const output = new Float32Array(outputReadback.getMappedRange().slice(0))
+    outputReadback.unmap()
+
+    const captures = new Map<string, Float32Array>()
+    if (wantCaptures) {
+      for (const [name, staging] of stagings!) {
+        await staging.mapAsync(GPUMapMode.READ)
+        captures.set(name, new Float32Array(staging.getMappedRange().slice(0)))
+        staging.unmap()
+      }
+    }
+    return { output, captures }
+  }
+
+  // ---- step() — training-mode wrapper, returns scalar [0] of output ---------
+  function step(inputs: Record<string, Int32Array | Float32Array>): Promise<number>
+  function step(inputs: Record<string, Int32Array | Float32Array>, opts: { withCaptures: true }): Promise<StepResult>
+  function step(inputs: Record<string, Int32Array | Float32Array>, opts: RunOptions): Promise<number | StepResult>
+  async function step(
+    inputs: Record<string, Int32Array | Float32Array>,
+    opts?: RunOptions,
+  ): Promise<number | StepResult> {
+    const r = await dispatch(inputs, opts?.withCaptures === true)
+    if (opts?.withCaptures) return { loss: r.output[0]!, captures: new Captures(captureShapes, r.captures) }
+    return r.output[0]!
+  }
+
+  // ---- run() — forward-mode wrapper, returns Float32Array by default -------
+  // Same overloaded shape as step(): scalar-shaped result (here Float32Array,
+  // there a JS number) is the default; { ..., captures } is the opt-in form.
+  function run(inputs: Record<string, Int32Array | Float32Array>): Promise<Float32Array>
+  function run(inputs: Record<string, Int32Array | Float32Array>, opts: { withCaptures: true }): Promise<RunResult>
+  function run(inputs: Record<string, Int32Array | Float32Array>, opts: RunOptions): Promise<Float32Array | RunResult>
+  async function run(
+    inputs: Record<string, Int32Array | Float32Array>,
+    opts?: RunOptions,
+  ): Promise<Float32Array | RunResult> {
+    const r = await dispatch(inputs, opts?.withCaptures === true)
+    if (opts?.withCaptures) return { output: r.output, captures: new Captures(captureShapes, r.captures) }
+    return r.output
+  }
+
+  // ---- uploadParams ---------------------------------------------------------
+  function uploadParams(params: Record<string, Float32Array>, opts?: UploadParamsOptions) {
+    const partial = opts?.partial ?? false
+    for (const name of Object.keys(params)) {
+      if (!plan.paramsByName.has(name)) {
+        throw new Error(
+          `uploadParams: unknown param '${name}'. ` +
+          `Known: ${[...plan.paramsByName.keys()].sort().join(', ')}`,
+        )
+      }
+    }
+    if (!partial) {
+      for (const name of plan.paramsByName.keys()) {
+        if (!(name in params)) {
+          throw new Error(
+            `uploadParams: missing param '${name}'. ` +
+            `Pass { partial: true } if you mean to update only some params.`,
+          )
+        }
+      }
+    }
+    for (const [name, bufId] of plan.paramsByName) {
+      const data = params[name]
+      if (!data) continue
+      const expected = plan.buffers[bufId]!.byteSize / 4
+      if (data.length !== expected) {
+        throw new Error(`uploadParams: '${name}' has ${data.length} elements, expected ${expected}`)
+      }
+      queue.writeBuffer(buffers.get(bufId)!, 0, data as unknown as BufferSource)
+    }
+  }
+
+  // ---- download helpers -----------------------------------------------------
+  async function downloadFromMap(map: Map<string, number>): Promise<Record<string, Float32Array>> {
+    const stagings: { name: string; buf: GPUBuffer; bytes: number }[] = []
+    const encoder = device.createCommandEncoder({ label: 'tensorgrad-download' })
+    for (const [name, bufId] of map) {
+      const spec = plan.buffers[bufId]!
+      const staging = device.createBuffer({ size: spec.byteSize, usage: READBACK })
+      encoder.copyBufferToBuffer(buffers.get(bufId)!, 0, staging, 0, spec.byteSize)
+      stagings.push({ name, buf: staging, bytes: spec.byteSize })
+    }
+    queue.submit([encoder.finish()])
+    const out: Record<string, Float32Array> = {}
+    for (const s of stagings) {
+      await s.buf.mapAsync(GPUMapMode.READ)
+      out[s.name] = new Float32Array(s.buf.getMappedRange().slice(0))
+      s.buf.unmap()
+      s.buf.destroy()
+    }
+    return out
+  }
+
+  // Fill a state buffer with its declared initValue (typically 0). Float and
+  // int both serialize to 4 bytes per element. Used at allocation time and on
+  // resetOptimizerState() — same logic, two callers.
+  function fillStateBuffer(spec: { byteSize: number; dtype: 'f32' | 'i32' | 'bool'; initValue?: number }, target: GPUBuffer): void {
+    const elements = spec.byteSize / 4
+    const init = spec.dtype === 'f32'
+      ? new Float32Array(elements).fill(spec.initValue ?? 0)
+      : new Int32Array(elements).fill(Math.trunc(spec.initValue ?? 0))
+    queue.writeBuffer(target, 0, init as unknown as BufferSource)
+  }
+
+  function resetOptimizerState() {
+    for (const spec of plan.buffers) {
+      if (spec.kind === 'state') fillStateBuffer(spec, buffers.get(spec.id)!)
+    }
+  }
+
+  // Build the params map AFTER buffer allocation so it points at the actual
+  // GPUBuffers (shared or freshly allocated).
+  const params = new Map<string, GPUBuffer>()
+  for (const [name, bufId] of plan.paramsByName) {
+    params.set(name, buffers.get(bufId)!)
+  }
+  // Static-after-compile shape metadata so users don't have to recompute
+  // strides to interpret a flat capture readback.
+  const captureShapes: Record<string, number[]> = {}
+  for (const [name, bufId] of plan.capturesByName) {
+    captureShapes[name] = [...plan.buffers[bufId]!.shape]
+  }
+  const outputShape = [...plan.buffers[lossBufferId]!.shape]
+
+  const destroy = () => {
+    for (const [id, b] of buffers) {
+      if (ownedBufferIds.has(id)) b.destroy()
+    }
+    outputReadback.destroy()
+    if (captureStagings) for (const b of captureStagings.values()) b.destroy()
+  }
+
+  return {
+    device,
+    params,
+    outputShape,
+    uploadParams,
+    downloadParams: () => downloadFromMap(plan.paramsByName),
+    downloadParamGrads: () => downloadFromMap(plan.paramGradsByName),
+    step,
+    run,
+    resetOptimizerState,
+    destroy,
+  }
+}
+
+/** 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 async function createForwardRuntime(
+  plan: BufferPlan,
+  kernels: KernelSpec[],
+  outputBufferId: number,
+  opts: RuntimeOpts = {},
+): Promise<CompiledForward> {
+  return await createRuntime(plan, kernels, outputBufferId, opts)
+}
+
+async function acquireDevice(): Promise<GPUDevice> {
+  if (typeof navigator === 'undefined' || !navigator.gpu) {
+    throw new Error('tensorgrad: WebGPU not available in this environment')
+  }
+  const adapter = await navigator.gpu.requestAdapter()
+  if (!adapter) throw new Error('tensorgrad: no WebGPU adapter')
+  return await adapter.requestDevice()
+}