@simulatte/doppler 0.1.8 → 0.2.0

package/src/gpu/kernels/check-stop.d.ts

@@ -1,31 +0,0 @@
-/**
- * check-stop.ts - GPU Stop Condition Kernel
- *
- * Checks if generation should stop based on EOS token or max length.
- * Used in GPU-only decode loop to eliminate CPU roundtrips.
- */
-
-import type { CommandRecorder } from '../command-recorder.js';
-
-export interface CheckStopParams {
-  sampledTokenBuffer: GPUBuffer;  // u32 storage buffer with sampled tokens
-  tokenIndex?: number;
-  shouldStopBuffer?: GPUBuffer;
-  eosTokenId: number;
-  maxTokens: number;
-  currentPos: number;
-}
-
-/**
- * Record a stop condition check into the command buffer.
- * Returns the shouldStop buffer (1 if should stop, 0 otherwise).
- */
-export declare function recordCheckStop(
-  recorder: CommandRecorder,
-  params: CheckStopParams
-): GPUBuffer;
-
-/**
- * Standalone check stop (submits immediately, for testing).
- */
-export declare function checkStop(params: CheckStopParams): Promise<boolean>;

package/src/gpu/kernels/check-stop.js

@@ -1,170 +0,0 @@
-
-
-import { getDevice, getDeviceEpoch } from '../device.js';
-import { acquireBuffer, readBufferSlice } from '../../memory/buffer-pool.js';
-import { recordDispatch } from './dispatch.js';
-import { createUniformBufferFromData, getOrCreateBindGroupLayout, getOrCreatePipelineLayout } from './utils.js';
-import { allowReadback } from '../perf-guards.js';
-
-
-let checkStopPipeline = null;
-let checkStopPipelineEpoch = -1;
-const U32_BYTES = Uint32Array.BYTES_PER_ELEMENT;
-
-const SHADER = /* wgsl */ `
-override WORKGROUP_SIZE: u32 = 1u;
-
-struct StopUniforms {
-    eosTokenId: u32,
-    maxTokens: u32,
-    currentPos: u32,
-    tokenIndex: u32,
-}
-
-@group(0) @binding(0) var<uniform> uniforms: StopUniforms;
-@group(0) @binding(1) var<storage, read> sampledToken: array<u32>;
-@group(0) @binding(2) var<storage, read_write> shouldStop: array<u32>;
-
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
-fn main() {
-    let token = sampledToken[uniforms.tokenIndex];
-    let isEOS = (token == uniforms.eosTokenId);
-    let reachedMax = (uniforms.currentPos >= uniforms.maxTokens);
-
-    shouldStop[uniforms.tokenIndex] = select(0u, 1u, isEOS || reachedMax);
-}
-`;
-
-
-function getCheckStopBindGroupLayout(device) {
-  return getOrCreateBindGroupLayout(
-    'check_stop_bind_group_layout',
-    [
-      { binding: 0, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'uniform' } },
-      { binding: 1, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'read-only-storage' } },
-      { binding: 2, visibility: GPUShaderStage.COMPUTE, buffer: { type: 'storage' } },
-    ],
-    device
-  );
-}
-
-
-function getCheckStopPipeline() {
-  const epoch = getDeviceEpoch();
-  if (checkStopPipeline && checkStopPipelineEpoch === epoch) return checkStopPipeline;
-  const device = getDevice();
-  const shaderModule = device.createShaderModule({ code: SHADER });
-  const bindGroupLayout = getCheckStopBindGroupLayout(device);
-
-  checkStopPipeline = device.createComputePipeline({
-    layout: getOrCreatePipelineLayout('check_stop_pipeline_layout', [bindGroupLayout], device),
-    compute: {
-      module: shaderModule,
-      entryPoint: 'main',
-      constants: { WORKGROUP_SIZE: 1 },
-    },
-  });
-  checkStopPipelineEpoch = epoch;
-
-  return checkStopPipeline;
-}
-
-
-export function recordCheckStop(
-  recorder,
-  params
-) {
-  const device = getDevice();
-  const pipeline = getCheckStopPipeline();
-  const tokenIndex = params.tokenIndex ?? 0;
-
-  // Create uniform buffer
-  const uniformData = new Uint32Array([
-    params.eosTokenId,
-    params.maxTokens,
-    params.currentPos,
-    tokenIndex,
-  ]);
-  const uniformBuffer = createUniformBufferFromData('check_stop_uniforms', uniformData, recorder);
-
-  // Create output buffer
-  const requiredBytes = (tokenIndex + 1) * U32_BYTES;
-  const shouldStopBuffer = params.shouldStopBuffer ?? acquireBuffer(requiredBytes, undefined, 'check_stop_output');
-  if (shouldStopBuffer.size < requiredBytes) {
-    throw new Error('[CheckStop] shouldStopBuffer too small for tokenIndex.');
-  }
-
-  // Create bind group
-  const bindGroup = device.createBindGroup({
-    layout: getCheckStopBindGroupLayout(device),
-    entries: [
-      { binding: 0, resource: { buffer: uniformBuffer } },
-      { binding: 1, resource: { buffer: params.sampledTokenBuffer } },
-      { binding: 2, resource: { buffer: shouldStopBuffer } },
-    ],
-  });
-
-  recordDispatch(recorder, pipeline, bindGroup, 1, 'check_stop');
-
-  return shouldStopBuffer;
-}
-
-
-export async function checkStop(params) {
-  if (!allowReadback('check-stop')) {
-    throw new Error('[CheckStop] GPU readback disabled');
-  }
-
-  const device = getDevice();
-  const pipeline = getCheckStopPipeline();
-
-  const tokenIndex = params.tokenIndex ?? 0;
-  const uniformData = new Uint32Array([
-    params.eosTokenId,
-    params.maxTokens,
-    params.currentPos,
-    tokenIndex,
-  ]);
-  const uniformBuffer = createUniformBufferFromData('check_stop_uniforms', uniformData, null, device);
-
-  const requiredBytes = (tokenIndex + 1) * U32_BYTES;
-  const shouldStopBuffer = params.shouldStopBuffer ?? device.createBuffer({
-    size: requiredBytes,
-    usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC,
-  });
-  const ownsStopBuffer = !params.shouldStopBuffer;
-
-  try {
-    if (shouldStopBuffer.size < requiredBytes) {
-      throw new Error('[CheckStop] shouldStopBuffer too small for tokenIndex.');
-    }
-
-    const bindGroup = device.createBindGroup({
-      layout: getCheckStopBindGroupLayout(device),
-      entries: [
-        { binding: 0, resource: { buffer: uniformBuffer } },
-        { binding: 1, resource: { buffer: params.sampledTokenBuffer } },
-        { binding: 2, resource: { buffer: shouldStopBuffer } },
-      ],
-    });
-
-    const encoder = device.createCommandEncoder();
-    const pass = encoder.beginComputePass();
-    pass.setPipeline(pipeline);
-    pass.setBindGroup(0, bindGroup);
-    pass.dispatchWorkgroups(1, 1, 1);
-    pass.end();
-
-    device.queue.submit([encoder.finish()]);
-
-    const result = new Uint32Array(
-      await readBufferSlice(shouldStopBuffer, tokenIndex * U32_BYTES, U32_BYTES)
-    )[0];
-    return result === 1;
-  } finally {
-    uniformBuffer.destroy();
-    if (ownsStopBuffer) {
-      shouldStopBuffer.destroy();
-    }
-  }
-}

package/src/gpu/kernels/clamp.d.ts

@@ -1,22 +0,0 @@
-import type { CommandRecorder } from '../command-recorder.js';
-import type { Tensor } from '../tensor.js';
-
-export interface ClampOptions {
-  count?: number;
-}
-
-export declare function runClamp(
-  input: Tensor,
-  minValue: number,
-  maxValue: number,
-  options?: ClampOptions
-): Promise<Tensor>;
-
-export declare function recordClamp(
-  recorder: CommandRecorder,
-  input: Tensor,
-  minValue: number,
-  maxValue: number,
-  options?: ClampOptions
-): Promise<Tensor>;
-

package/src/gpu/kernels/clamp.js

@@ -1,42 +0,0 @@
-import { createTensor } from '../tensor.js';
-import { dtypeBytes } from '../tensor.js';
-import { WORKGROUP_SIZES } from './constants.js';
-import { unifiedKernelWrapper } from './utils.js';
-
-function resolveCount(tensor, countOverride) {
-  if (Number.isFinite(countOverride) && countOverride > 0) {
-    return Math.floor(countOverride);
-  }
-  if (Array.isArray(tensor.shape) && tensor.shape.length > 0) {
-    return tensor.shape.reduce((acc, value) => acc * value, 1);
-  }
-  return Math.floor(tensor.buffer.size / dtypeBytes(tensor.dtype));
-}
-
-async function _clamp(target, input, minValue, maxValue, options = {}) {
-  if (input.dtype !== 'f32') {
-    throw new Error(`clamp: unsupported dtype ${input.dtype}.`);
-  }
-
-  const { count } = options;
-  const inferredCount = resolveCount(input, count);
-
-  await unifiedKernelWrapper(
-    'clamp',
-    target,
-    'default',
-    [input],
-    { size: inferredCount, min: minValue, max: maxValue },
-    Math.ceil(inferredCount / WORKGROUP_SIZES.DEFAULT)
-  );
-
-  return createTensor(input.buffer, input.dtype, [...input.shape], 'clamp_output');
-}
-
-export async function runClamp(input, minValue, maxValue, options = {}) {
-  return _clamp(null, input, minValue, maxValue, options);
-}
-
-export async function recordClamp(recorder, input, minValue, maxValue, options = {}) {
-  return _clamp(recorder, input, minValue, maxValue, options);
-}

package/src/gpu/kernels/clamp.wgsl

@@ -1,24 +0,0 @@
-// clamp.wgsl
-// Clamp an f32 buffer in-place to [min, max].
-
-override WORKGROUP_SIZE: u32 = 256u;
-
-struct Uniforms {
-    size: u32,
-    _pad0: u32,
-    min: f32,
-    max: f32,
-}
-
-@group(0) @binding(0) var<uniform> u: Uniforms;
-@group(0) @binding(1) var<storage, read_write> data: array<f32>;
-
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-    let idx = gid.x;
-    if (idx >= u.size) {
-        return;
-    }
-    data[idx] = clamp(data[idx], u.min, u.max);
-}
-

package/src/gpu/kernels/constants.d.ts

@@ -1,168 +0,0 @@
-/**
- * Kernel Constants - Shared constants for GPU kernels
- *
- * Centralized constants to eliminate magic numbers and improve
- * maintainability across kernel implementations.
- */
-
-/**
- * Workgroup sizes for different kernel types
- */
-export declare const WORKGROUP_SIZES: {
-  /** Default workgroup size for most kernels */
-  readonly DEFAULT: 256;
-
-  /** Vec4 workgroup thread count (64 threads × 4 elements = 256 elements) */
-  readonly VEC4_THREADS: 64;
-
-  /** Attention kernels (large blocks) */
-  readonly ATTENTION_LARGE_BLOCK: 32;
-
-  /** Attention kernels (small blocks) */
-  readonly ATTENTION_SMALL_BLOCK: 32;
-
-  /** Subgroup size (typical for modern GPUs) */
-  readonly SUBGROUP: 32;
-
-  /** RMSNorm workgroup size */
-  readonly RMSNORM: 256;
-
-  /** Softmax workgroup size */
-  readonly SOFTMAX: 256;
-
-  /** Matmul tile sizes */
-  readonly MATMUL_TILE_M: 16;
-  readonly MATMUL_TILE_N: 16;
-  readonly MATMUL_TILE_K: 16;
-
-  /** MoE workgroup size */
-  readonly MOE: 256;
-};
-
-/** Derived: Vec4 elements per workgroup (VEC4_THREADS × 4) */
-export declare const VEC4_ELEMENTS_PER_WG: number;
-
-/**
- * WebGPU limits (spec-level defaults)
- */
-export declare const GPU_LIMITS: {
-  /** Max workgroups per dimension (WebGPU minimum) */
-  readonly MAX_WORKGROUPS: 65535;
-};
-
-/**
- * Tile sizes for different operations
- */
-export declare const TILE_SIZES: {
-  /** Attention tile sizes (large) */
-  readonly ATTENTION_LARGE_BLOCK_SIZE: 32;
-  readonly ATTENTION_LARGE_HEAD_TILE: 64;
-
-  /** Attention tile sizes (small) */
-  readonly ATTENTION_SMALL_BLOCK_SIZE: 32;
-  readonly ATTENTION_SMALL_HEAD_TILE: 32;
-
-  /** Matmul tile sizes */
-  readonly MATMUL_M: 16;
-  readonly MATMUL_N: 16;
-  readonly MATMUL_K: 16;
-
-  /** Q4K dequant tile sizes */
-  readonly Q4K_BLOCK_SIZE: 32;
-  readonly Q4K_SUPER_BLOCK_SIZE: 256;
-};
-
-/**
- * Quantization constants
- */
-export declare const QUANTIZATION: {
-  /** Q4K_M bits per weight */
-  readonly Q4K_BITS: 4.5;
-  /** Q4K block bytes per 256-element super-block */
-  readonly Q4K_BLOCK_BYTES: 144;
-
-  /** Q8_0 bits per weight */
-  readonly Q8_BITS: 8.5;
-
-  /** F16 bits per weight */
-  readonly F16_BITS: 16;
-
-  /** BF16 bits per weight */
-  readonly BF16_BITS: 16;
-
-  /** F32 bits per weight */
-  readonly F32_BITS: 32;
-
-  /** MXFP4 bits per weight (including shared exponent) */
-  readonly MXFP4_BITS: 4;
-};
-
-/**
- * Buffer alignment requirements
- */
-export declare const ALIGNMENT: {
-  /** WebGPU buffer alignment */
-  readonly BUFFER: 256;
-
-  /** Uniform buffer alignment */
-  readonly UNIFORM: 256;
-
-  /** Storage buffer alignment */
-  readonly STORAGE: 256;
-
-  /** Vertex buffer alignment */
-  readonly VERTEX: 4;
-};
-
-/**
- * Performance tuning constants
- */
-export declare const PERFORMANCE: {
-  /** Number of warmup runs for benchmarks */
-  readonly WARMUP_RUNS: 5;
-
-  /** Number of timed runs for benchmarks */
-  readonly TIMED_RUNS: 20;
-
-  /** Default timeout for operations (ms) */
-  readonly DEFAULT_TIMEOUT: 120000;
-
-  /** Max buffer pool size per bucket */
-  readonly MAX_POOL_SIZE_PER_BUCKET: 8;
-
-  /** Max total pooled buffers */
-  readonly MAX_TOTAL_POOLED_BUFFERS: 64;
-};
-
-/**
- * Dtype size mappings (in bytes)
- */
-export declare const DTYPE_SIZES: {
-  readonly u8: 1;
-  readonly i8: 1;
-  readonly u16: 2;
-  readonly i16: 2;
-  readonly f16: 2;
-  readonly bf16: 2;
-  readonly u32: 4;
-  readonly i32: 4;
-  readonly f32: 4;
-  readonly f64: 8;
-};
-
-export type DType = keyof typeof DTYPE_SIZES;
-
-/**
- * Get dtype size in bytes
- */
-export declare function getDtypeSize(dtype: DType): number;
-
-/**
- * Calculate buffer size for tensor
- */
-export declare function calculateBufferSize(shape: number[], dtype: DType): number;
-
-/**
- * Round size up to alignment boundary
- */
-export declare function alignSize(size: number, alignment?: number): number;

package/src/gpu/kernels/constants.js

@@ -1,129 +0,0 @@
-
-
-
-import { QK_K, Q4K_BLOCK_BYTES } from '../../config/schema/index.js';
-
-export const WORKGROUP_SIZES = {
-
-  DEFAULT: 256,
-
-  
-  VEC4_THREADS: 64,
-
-  
-  ATTENTION_LARGE_BLOCK: 32,
-
-  
-  ATTENTION_SMALL_BLOCK: 32,
-
-  
-  SUBGROUP: 32,
-
-  
-  RMSNORM: 256,
-
-  
-  SOFTMAX: 256,
-
-  
-  MATMUL_TILE_M: 16,
-  MATMUL_TILE_N: 16,
-  MATMUL_TILE_K: 16,
-
-  
-  MOE: 256,
-};
-
-
-export const VEC4_ELEMENTS_PER_WG = WORKGROUP_SIZES.VEC4_THREADS * 4;  // 256
-
-
-export const GPU_LIMITS = {
-  
-  MAX_WORKGROUPS: 65535,
-};
-
-export const TILE_SIZES = {
-
-  ATTENTION_LARGE_BLOCK_SIZE: 32,
-  ATTENTION_LARGE_HEAD_TILE: 64,
-
-
-  ATTENTION_SMALL_BLOCK_SIZE: 32,
-  ATTENTION_SMALL_HEAD_TILE: 32,
-
-
-  MATMUL_M: 16,
-  MATMUL_N: 16,
-  MATMUL_K: 16,
-
-  // Q4K tile size (sub-blocks within super-block)
-  Q4K_BLOCK_SIZE: 32,
-  // Q4K super-block size imported from schema (single source of truth)
-  Q4K_SUPER_BLOCK_SIZE: QK_K,
-};
-
-
-export const QUANTIZATION = {
-
-  Q4K_BITS: 4.5,
-  // Q4K block bytes imported from schema (single source of truth)
-  Q4K_BLOCK_BYTES,
-
-  
-  Q8_BITS: 8.5,
-
-  
-  F16_BITS: 16,
-
-  
-  BF16_BITS: 16,
-
-  
-  F32_BITS: 32,
-
-  
-  MXFP4_BITS: 4,
-};
-
-
-export const ALIGNMENT = {
-  
-  BUFFER: 256,
-
-  
-  UNIFORM: 256,
-
-  
-  STORAGE: 256,
-
-  
-  VERTEX: 4,
-};
-
-
-export const PERFORMANCE = {
-  
-  WARMUP_RUNS: 5,
-
-  
-  TIMED_RUNS: 20,
-
-  
-  DEFAULT_TIMEOUT: 120000,
-
-  
-  MAX_POOL_SIZE_PER_BUCKET: 8,
-
-  
-  MAX_TOTAL_POOLED_BUFFERS: 64,
-};
-
-
-// DTYPE_SIZES and getDtypeSize moved to config/schema/kernel-thresholds.schema.js
-// Import from config/schema/index.js for the canonical source
-
-
-export function alignSize(size, alignment = ALIGNMENT.BUFFER) {
-  return Math.ceil(size / alignment) * alignment;
-}

package/src/gpu/kernels/conv2d.d.ts

@@ -1,34 +0,0 @@
-import type { Tensor } from '../tensor.js';
-import type { CommandRecorder } from '../command-recorder.js';
-import type { OutputBufferOptions } from './types.js';
-import type { WeightBuffer } from '../weight-buffer.js';
-
-export interface Conv2DOptions extends OutputBufferOptions {
-  inChannels: number;
-  outChannels: number;
-  height: number;
-  width: number;
-  kernelH: number;
-  kernelW: number;
-  stride?: number;
-  pad?: number;
-}
-
-export interface Conv2DSelectionContext {
-  isF16: boolean;
-}
-
-export declare function runConv2D(
-  input: Tensor,
-  weight: GPUBuffer | WeightBuffer,
-  bias: GPUBuffer | WeightBuffer | null,
-  options: Conv2DOptions
-): Promise<Tensor>;
-
-export declare function recordConv2D(
-  recorder: CommandRecorder,
-  input: Tensor,
-  weight: GPUBuffer | WeightBuffer,
-  bias: GPUBuffer | WeightBuffer | null,
-  options: Conv2DOptions
-): Promise<Tensor>;