@simulatte/doppler 0.1.9 → 0.2.0

package/src/gpu/kernels/index.js

@@ -1,315 +0,0 @@
-
-
-// Utilities
-export {
-  KERNEL_CONFIGS,
-  validateAttentionLimits,
-  loadShaderSource,
-  hasRequiredFeatures,
-  getKernelConfig,
-  compileShader,
-  getOrCreateBindGroupLayout,
-  getOrCreatePipelineLayout,
-  createPipeline,
-  clearKernelCaches,
-  clearPipelineCache,
-  getCacheStats,
-  getTunedWorkgroupSize,
-  autoTuneKernels,
-  prewarmKernels,
-} from './utils.js';
-
-// Matrix Multiplication
-export {
-  selectMatmulKernel,
-  createMatmulBindGroupLayout,
-  runMatmul,
-  recordMatmul,
-  isFusedQ4KDisabled,
-} from './matmul.js';
-
-// Dequantization
-export {
-  selectDequantKernel,
-  createDequantBindGroupLayout,
-  dequantize,
-  dequantizeRowwise,
-  dequantizeQ6K,
-  dequantizeQ8_0,
-  dequantizeMXFP4,
-  dequantizeMXFP4Expert,
-  recordDequantize,
-} from './dequant.js';
-
-// Attention
-export {
-  runAttention,
-  recordAttention,
-  runAttentionTiered,
-  recordAttentionTiered,
-  runAttentionTieredQuant,
-  recordAttentionTieredQuant,
-  runAttentionBDPA,
-  recordAttentionBDPA,
-} from './attention.js';
-
-// RMSNorm
-export {
-  selectRMSNormKernel,
-  runRMSNorm,
-  recordRMSNorm,
-} from './rmsnorm.js';
-
-// LayerNorm
-export {
-  selectLayerNormKernel,
-  runLayerNorm,
-  recordLayerNorm,
-} from './layernorm.js';
-
-// Softmax
-export {
-  runSoftmax,
-  runSoftmaxTopK,
-  recordSoftmax,
-} from './softmax.js';
-
-// KV Quantization
-export {
-  runKVQuantize,
-  recordKVQuantize,
-} from './kv-quantize.js';
-
-// Loss
-export {
-  runCrossEntropyLoss,
-  recordCrossEntropyLoss,
-} from './cross_entropy_loss.js';
-
-// RoPE
-export {
-  runRoPE,
-  recordRoPE,
-} from './rope.js';
-
-// SiLU Activation
-export {
-  runSiLU,
-  runSwiGLURowsplitBias,
-  runSiLURowSplit,
-  recordSiLU,
-  recordSiLURowSplit,
-} from './silu.js';
-
-// GeLU Activation
-export {
-  runGeLU,
-  recordGeLU,
-} from './gelu.js';
-
-// Scale (Element-wise Multiply by Scalar)
-export {
-  runScale,
-  recordScale,
-} from './scale.js';
-
-// Clamp
-export {
-  runClamp,
-  recordClamp,
-} from './clamp.js';
-
-// Energy (EBM helpers)
-export {
-  runEnergyEval,
-  recordEnergyEval,
-  runEnergyUpdate,
-  recordEnergyUpdate,
-  runEnergyQuintelUpdate,
-  recordEnergyQuintelUpdate,
-  runEnergyQuintelReduce,
-  recordEnergyQuintelReduce,
-  runEnergyQuintelGrad,
-  recordEnergyQuintelGrad,
-} from './energy.js';
-
-// Conv2D
-export {
-  runConv2D,
-  recordConv2D,
-} from './conv2d.js';
-
-export {
-  runDepthwiseConv2D,
-  recordDepthwiseConv2D,
-} from './depthwise_conv2d.js';
-
-export {
-  runGroupedPointwiseConv2D,
-  recordGroupedPointwiseConv2D,
-} from './grouped_pointwise_conv2d.js';
-
-// Gather (Embedding Lookup)
-export {
-  runGather,
-  recordGather,
-} from './gather.js';
-
-// GroupNorm
-export {
-  runGroupNorm,
-  recordGroupNorm,
-} from './groupnorm.js';
-
-// Modulate
-export {
-  runModulate,
-  recordModulate,
-} from './modulate.js';
-
-// Residual Connections
-export {
-  runResidualAdd,
-  runBiasAdd,
-  recordResidualAdd,
-  recordBiasAdd,
-} from './residual.js';
-
-// Pixel Shuffle
-export {
-  runPixelShuffle,
-  recordPixelShuffle,
-} from './pixel_shuffle.js';
-
-// Upsample2D
-export {
-  runUpsample2D,
-  recordUpsample2D,
-} from './upsample2d.js';
-
-// Mixture of Experts
-export {
-  runTopK,
-  runMoEGather,
-  runMoEBuildTokenOffsets,
-  recordMoEBuildTokenOffsets,
-  runScatterAdd,
-  runScatterAddDynamic,
-} from './moe.js';
-
-// Type Casting
-export {
-  castF32ToF16,
-  recordCastF32ToF16,
-  castF16ToF32,
-  recordCastF16ToF32,
-  runBF16ToF32,
-  runBF16ToF16,
-} from './cast.js';
-
-// GPU-Side Sampling
-export {
-  runArgmax,
-  runGPUSample,
-  recordArgmax,
-  isGPUSamplingAvailable,
-} from './sample.js';
-
-export {
-  runSanaLinearAttention,
-  recordSanaLinearAttention,
-} from './sana_linear_attention.js';
-
-export {
-  runRepeatChannels,
-  recordRepeatChannels,
-} from './repeat_channels.js';
-
-export {
-  runReLU,
-  recordReLU,
-} from './relu.js';
-
-// Fused FFN (Tier 2 P0)
-export {
-  runFusedFFN,
-  recordFusedFFN,
-  calculateFusedFFNSavings,
-} from './fused_ffn.js';
-
-// Fused Matmul + RMSNorm (P0 - 1.2-1.5x decode speedup)
-export {
-  selectMatmulRMSNormFusedVariant,
-  runMatmulRMSNormFused,
-  recordMatmulRMSNormFused,
-  shouldUseFusedMatmulRMSNorm,
-} from './fused_matmul_rmsnorm.js';
-
-// Re-export for convenience in layer.ts integration
-export { recordMatmulRMSNormFused as doRecordMatmulRMSNormFused } from './fused_matmul_rmsnorm.js';
-
-// Fused Matmul + Residual (P1 - eliminates 1 dispatch per layer for attention output)
-export {
-  runMatmulResidualFused,
-  recordMatmulResidualFused,
-  shouldUseFusedMatmulResidual,
-} from './fused_matmul_residual.js';
-
-// Re-export CommandRecorder types for convenience
-export {
-  CommandRecorder,
-  createCommandRecorder,
-  createProfilingRecorder,
-} from '../command-recorder.js';
-
-// Split QKV
-export {
-  runSplitQKV,
-  recordSplitQKV,
-} from './split_qkv.js';
-
-// Split Q and Gate (de-interleave attentionOutputGate q_proj output)
-export {
-  runSplitQG,
-  recordSplitQG,
-} from './split_qg.js';
-
-// Transpose
-export {
-  runTranspose,
-  recordTranspose,
-} from './transpose.js';
-
-// Training Backward Kernels
-export {
-  runEmbedBackward,
-  recordEmbedBackward,
-  runBiasAddBackward,
-  runUpsample2DBackward,
-  runGroupNormBackward,
-  runConv2DBackward,
-  runMatmulBackward,
-  recordMatmulBackward,
-  runSoftmaxBackward,
-  recordSoftmaxBackward,
-  runRmsNormBackward,
-  recordRmsNormBackward,
-  runLayerNormBackward,
-  recordLayerNormBackward,
-  runAttentionBackward,
-  recordAttentionBackward,
-  runRoPEBackward,
-  recordRoPEBackward,
-  runSiluBackward,
-  recordSiluBackward,
-  runGeluBackward,
-  recordGeluBackward,
-  runScaleBackward,
-  recordScaleBackward,
-  runCrossEntropyBackward,
-  recordCrossEntropyBackward,
-  runAdam,
-  recordAdam,
-} from './backward/index.js';
-
-// Re-export profiling utilities

package/src/gpu/kernels/kernel-base.d.ts

@@ -1,33 +0,0 @@
-/**
- * Kernel Base - Shared dispatch and pipeline helpers for kernel wrappers.
- */
-
-import type { CommandRecorder } from '../command-recorder.js';
-
-export declare abstract class KernelBase {
-  protected readonly device: GPUDevice;
-
-  constructor(device: GPUDevice);
-
-  protected getPipelineFor(
-    operation: string,
-    variant: string,
-    bindGroupLayout?: GPUBindGroupLayout | null,
-    constants?: Record<string, number | boolean> | null
-  ): Promise<GPUComputePipeline>;
-
-  protected dispatchKernel(
-    pipeline: GPUComputePipeline,
-    bindGroup: GPUBindGroup,
-    workgroups: number | [number, number, number],
-    label: string
-  ): void;
-
-  protected recordKernel(
-    recorder: CommandRecorder,
-    pipeline: GPUComputePipeline,
-    bindGroup: GPUBindGroup,
-    workgroups: number | [number, number, number],
-    label: string
-  ): void;
-}

package/src/gpu/kernels/kernel-base.js

@@ -1,46 +0,0 @@
-
-
-import { dispatch, recordDispatch } from './dispatch.js';
-import { getPipelineFast } from './utils.js';
-
-
-export class KernelBase {
-
-  device;
-
-
-  constructor(device) {
-    this.device = device;
-  }
-
-
-  async getPipelineFor(
-    operation,
-    variant,
-    bindGroupLayout = null,
-    constants = null
-  ) {
-    return getPipelineFast(operation, variant, bindGroupLayout, constants);
-  }
-
-
-  dispatchKernel(
-    pipeline,
-    bindGroup,
-    workgroups,
-    label
-  ) {
-    dispatch(this.device, pipeline, bindGroup, workgroups, label);
-  }
-
-
-  recordKernel(
-    recorder,
-    pipeline,
-    bindGroup,
-    workgroups,
-    label
-  ) {
-    recordDispatch(recorder, pipeline, bindGroup, workgroups, label);
-  }
-}

package/src/gpu/kernels/kernel-configs.d.ts

@@ -1,65 +0,0 @@
-/**
- * Kernel Configurations - Table-driven kernel metadata
- *
- * Contains all kernel configurations including shader files, entry points,
- * workgroup sizes, and feature requirements.
- *
- * @module gpu/kernels/kernel-configs
- */
-
-// ============================================================================
-// Types
-// ============================================================================
-
-/** Variant-specific metadata for table-driven kernel configuration */
-export interface VariantMetadata {
-  /** Columns processed per workgroup (matmul multicol variants) */
-  colsPerWg?: number;
-  /** Tile size for M dimension (batched matmul variants) */
-  tileM?: number;
-  /** Output buffer binding index (gather F16 output variants) */
-  outputBinding?: number;
-  /** Maximum KV length for chunked attention */
-  maxKVLen?: number;
-}
-
-/** Kernel configuration */
-export interface KernelConfig {
-  shaderFile: string;
-  entryPoint: string;
-  workgroupSize: [number, number, number];
-  requires: string[];
-  validate?: (seqLen: number, numHeads: number, headDim: number) => void;
-  /** Output dtype for variants that output to specific precision */
-  outputDtype?: 'f16' | 'f32';
-  /** Tile size for attention/matmul kernels */
-  tileSize?: number;
-  /** Additional variant-specific configuration */
-  variantMetadata?: VariantMetadata;
-}
-
-// ============================================================================
-// Kernel Configurations
-// ============================================================================
-
-/** All kernel configurations by operation and variant */
-export declare const KERNEL_CONFIGS: Record<string, Record<string, KernelConfig>>;
-
-// ============================================================================
-// Config Helpers
-// ============================================================================
-
-/**
- * Get kernel configuration
- */
-export declare function getKernelConfig(operation: string, variant: string): KernelConfig;
-
-/**
- * Set a validator function on a kernel config.
- * Used to set attention validators after import to avoid circular dependencies.
- */
-export declare function setKernelValidator(
-  operation: string,
-  variant: string,
-  validator: (seqLen: number, numHeads: number, headDim: number) => void
-): void;

package/src/gpu/kernels/kernel-configs.js

@@ -1,50 +0,0 @@
-import { loadJson } from '../../utils/load-json.js';
-
-const registry = await loadJson('../../config/kernels/registry.json', import.meta.url, 'Failed to load registry');
-import { resolveKernelConfig } from '../../config/schema/kernel-registry.schema.js';
-
-export const KERNEL_CONFIGS = Object.fromEntries(
-  Object.entries(registry.operations).map(([operation, opSchema]) => {
-    const variants = Object.fromEntries(
-      Object.entries(opSchema.variants).map(([variant, variantSchema]) => {
-        const resolved = resolveKernelConfig(operation, variant, opSchema, variantSchema);
-        return [
-          variant,
-          {
-            shaderFile: resolved.wgsl,
-            entryPoint: resolved.entryPoint,
-            workgroupSize: resolved.workgroup,
-            requires: resolved.requires,
-            bindings: resolved.bindings,
-            uniforms: resolved.uniforms,
-            wgslOverrides: resolved.wgslOverrides,
-            sharedMemory: resolved.sharedMemory,
-            outputDtype: resolved.outputDtype ?? undefined,
-            variantMetadata: resolved.variantMetadata ?? undefined,
-          },
-        ];
-      })
-    );
-
-    return [operation, variants];
-  })
-);
-
-export function getKernelConfig(operation, variant) {
-  const config = KERNEL_CONFIGS[operation]?.[variant];
-  if (!config) {
-    throw new Error(`Unknown kernel: ${operation}/${variant}`);
-  }
-  return config;
-}
-
-export function setKernelValidator(
-  operation,
-  variant,
-  validator
-) {
-  const config = KERNEL_CONFIGS[operation]?.[variant];
-  if (config) {
-    config.validate = validator;
-  }
-}

package/src/gpu/kernels/kernel-tuning.d.ts

@@ -1,42 +0,0 @@
-/**
- * Kernel Tuning - Auto-tuning and kernel prewarming
- *
- * Provides utilities for tuning kernel workgroup sizes and
- * prewarming kernel pipelines for optimal performance.
- *
- * @module gpu/kernels/kernel-tuning
- */
-
-// ============================================================================
-// Workgroup Size Tuning
-// ============================================================================
-
-/**
- * Get tuned workgroup size for an operation
- */
-export declare function getTunedWorkgroupSize(
-  operation: string,
-  inputSizes?: Record<string, number>
-): Promise<[number, number, number]>;
-
-// ============================================================================
-// Auto-Tuning
-// ============================================================================
-
-/**
- * Run auto-tuning for all kernels with given model config
- */
-export declare function autoTuneKernels(
-  modelConfig?: Record<string, number>
-): Promise<Record<string, any>>;
-
-// ============================================================================
-// Pipeline Prewarming
-// ============================================================================
-
-/**
- * Prewarm all supported kernel pipelines
- */
-export declare function prewarmKernels(
-  options?: { mode?: 'parallel' | 'sequential' }
-): Promise<void>;

package/src/gpu/kernels/kernel-tuning.js

@@ -1,149 +0,0 @@
-
-
-import { getKernelCapabilities } from '../device.js';
-import { getKernelTuner, getTunerConfig } from '../kernel-tuner.js';
-import { KERNEL_CONFIGS } from './kernel-configs.js';
-import { createPipeline } from './pipeline-cache.js';
-import { hasRequiredFeatures } from './feature-check.js';
-import { log } from '../../debug/index.js';
-
-// ============================================================================
-// Workgroup Size Tuning
-// ============================================================================
-
-
-export async function getTunedWorkgroupSize(
-  operation,
-  inputSizes = {}
-) {
-  try {
-    const tuner = await getKernelTuner();
-    const result = tuner.getCachedResult(operation, inputSizes);
-
-    if (result) {
-      return result.optimalWorkgroupSize;
-    }
-
-    // Run tuning if not cached
-    const tuneResult = await tuner.tuneKernel(operation, inputSizes);
-    return tuneResult.optimalWorkgroupSize;
-  } catch (e) {
-    log.warn('KernelTuning', `Tuning failed for ${operation}, using defaults: ${e.message}`);
-    const { fallbackWorkgroupSizes } = getTunerConfig();
-    const fallback = fallbackWorkgroupSizes?.[operation] ?? fallbackWorkgroupSizes?.default;
-    if (!fallback) {
-      throw new Error(`KernelTuning: missing fallback workgroup size for "${operation}".`);
-    }
-    return fallback;
-  }
-}
-
-// ============================================================================
-// Auto-Tuning
-// ============================================================================
-
-
-export async function autoTuneKernels(
-  modelConfig = {}
-) {
-  const {
-    hiddenSize = 4096,
-    intermediateSize = 14336,
-    numHeads = 32,
-    headDim = 128,
-    maxSeqLen = 4096,
-    vocabSize = 32000,
-  } = modelConfig;
-
-  const tuner = await getKernelTuner();
-  
-  const results = {};
-
-  // Tune matmul for common sizes
-  results.matmul_hidden = await tuner.tuneKernel('matmul', {
-    M: 1, N: hiddenSize, K: hiddenSize,
-  });
-  results.matmul_ffn = await tuner.tuneKernel('matmul', {
-    M: 1, N: intermediateSize, K: hiddenSize,
-  });
-
-  // Tune attention
-  results.attention = await tuner.tuneKernel('attention', {
-    seqLen: 1, numHeads, headDim,
-  });
-
-  // Tune softmax (LM head output)
-  results.softmax = await tuner.tuneKernel('softmax', {
-    innerSize: vocabSize, outerSize: 1,
-  });
-
-  // Tune RMSNorm
-  results.rmsnorm = await tuner.tuneKernel('rmsnorm', {
-    hiddenSize, numTokens: 1,
-  });
-
-  // Tune dequant
-  results.dequant = await tuner.tuneKernel('dequant', {
-    numBlocks: 1000,
-  });
-
-  log.debug('KernelTuning', `Auto-tuning complete: ${JSON.stringify(results)}`);
-  return results;
-}
-
-// ============================================================================
-// Pipeline Prewarming
-// ============================================================================
-
-
-export async function prewarmKernels(
-  options = {}
-) {
-  const caps = getKernelCapabilities();
-  const mode = options.mode ?? 'parallel';
-  const entries = Object.entries(KERNEL_CONFIGS)
-    .sort(([a], [b]) => a.localeCompare(b))
-    .map(([operation, variants]) => [
-      operation,
-      Object.entries(variants).sort(([a], [b]) => a.localeCompare(b))
-    ]);
-
-  if (mode === 'sequential') {
-    let count = 0;
-    for (const [operation, variants] of entries) {
-      for (const [variant, cfg] of variants) {
-        if (cfg.requires && !hasRequiredFeatures(cfg.requires, caps)) {
-          continue;
-        }
-        try {
-          await createPipeline(operation, variant);
-          count += 1;
-        } catch (e) {
-          log.warn('KernelTuning', `Prewarm failed for ${operation}/${variant}: ${e.message}`);
-        }
-      }
-    }
-    log.debug('KernelTuning', `Prewarmed ${count} kernel pipelines`);
-    return;
-  }
-
-  
-  const jobs = [];
-  for (const [operation, variants] of entries) {
-    for (const [variant, cfg] of variants) {
-      if (cfg.requires && !hasRequiredFeatures(cfg.requires, caps)) {
-        continue;
-      }
-      jobs.push(
-        createPipeline(operation, variant)
-          .then(() => {}) // Ignore the pipeline result
-          .catch((e) => {
-            log.warn('KernelTuning', `Prewarm failed for ${operation}/${variant}: ${e.message}`);
-          })
-      );
-    }
-  }
-
-  await Promise.all(jobs);
-  log.debug('KernelTuning', `Prewarmed ${jobs.length} kernel pipelines`);
-}