@simulatte/doppler 0.1.8 → 0.2.0

package/src/inference/pipelines/diffusion/helpers.js

@@ -1,120 +0,0 @@
-import { getDevice, getKernelCapabilities } from '../../../gpu/device.js';
-import { getBuffer } from '../../../gpu/weight-buffer.js';
-import { releaseBuffer, isBufferActive } from '../../../memory/buffer-pool.js';
-import { selectRuleValue } from '../../../rules/rule-registry.js';
-
-export function resolveDiffusionActivationDtype(runtime) {
-  const caps = getKernelCapabilities();
-  return selectRuleValue('inference', 'dtype', 'diffusionActivationDtype', {
-    requested: runtime?.latent?.dtype,
-    hasF16: caps.hasF16,
-  });
-}
-
-export function createDiffusionBufferReleaser(recorder) {
-  if (!recorder) {
-    return (buffer) => {
-      if (!buffer || !isBufferActive(buffer)) return;
-      releaseBuffer(buffer);
-    };
-  }
-  return (buffer) => {
-    if (!buffer) return;
-    recorder.trackTemporaryBuffer(buffer);
-  };
-}
-
-export function createDiffusionBufferDestroyer(recorder) {
-  if (!recorder) {
-    return (buffer) => {
-      if (!buffer) return;
-      const device = getDevice();
-      if (!device) {
-        buffer.destroy();
-        return;
-      }
-      device.queue.onSubmittedWorkDone()
-        .then(() => {
-          buffer.destroy();
-        })
-        .catch(() => {
-          buffer.destroy();
-        });
-    };
-  }
-  return (buffer) => {
-    if (!buffer) return;
-    recorder.trackTemporaryBuffer(buffer);
-  };
-}
-
-export function createDiffusionIndexBuffer(device, indices, label) {
-  const buffer = device.createBuffer({
-    label,
-    size: indices.byteLength,
-    usage: GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
-  });
-  try {
-    device.queue.writeBuffer(buffer, 0, indices);
-    return buffer;
-  } catch (error) {
-    buffer.destroy();
-    throw error;
-  }
-}
-
-export function expectDiffusionWeight(weight, label) {
-  if (!weight) {
-    throw new Error(`Missing diffusion weight: ${label}`);
-  }
-  return weight;
-}
-
-export function normalizeDiffusionLocationDtype(dtype) {
-  if (!dtype) return null;
-  const normalized = String(dtype).toLowerCase();
-  if (normalized === 'f16' || normalized === 'float16') return 'f16';
-  if (normalized === 'f32' || normalized === 'float32') return 'f32';
-  if (normalized === 'bf16' || normalized === 'bfloat16') return 'f32';
-  return null;
-}
-
-export function normalizeDiffusionMatmulLocationDtype(dtype) {
-  if (!dtype) return null;
-  const normalized = String(dtype).toLowerCase();
-  if (normalized === 'f16' || normalized === 'float16') return 'f16';
-  if (normalized === 'bf16' || normalized === 'bfloat16') return 'bf16';
-  if (normalized === 'f32' || normalized === 'float32') return 'f32';
-  if (normalized === 'q4_k' || normalized === 'q4_k_m') return 'q4k';
-  return normalized;
-}
-
-// Artifact-derived dtype inference: determines actual storage dtype from buffer byte size.
-// This is NOT a config-bypass — it reads physical buffer dimensions (artifact-derived config),
-// which is a valid merge layer per the config merge contract.
-export function inferDiffusionMatmulDtypeFromBuffer(weight, N, K, preferred) {
-  const buffer = getBuffer(weight);
-  if (!buffer || !Number.isFinite(N) || !Number.isFinite(K)) return preferred;
-  if (preferred === 'q4k') return preferred;
-  const expectedF16 = N * K * 2;
-  const expectedF32 = N * K * 4;
-  if (preferred === 'f32' && buffer.size < expectedF32 && buffer.size >= expectedF16) {
-    return 'f16';
-  }
-  if (!preferred) {
-    if (buffer.size >= expectedF32) return 'f32';
-    if (buffer.size >= expectedF16) return 'f16';
-  }
-  return preferred;
-}
-
-export function sumDiffusionProfileTimings(timings) {
-  if (!timings || Object.keys(timings).length === 0) return null;
-  let total = 0;
-  for (const value of Object.values(timings)) {
-    if (Number.isFinite(value)) {
-      total += value;
-    }
-  }
-  return total;
-}

package/src/inference/pipelines/diffusion/index.d.ts

@@ -1,3 +0,0 @@
-export { DiffusionPipeline, createDiffusionPipeline } from './pipeline.js';
-export { mergeDiffusionConfig, initializeDiffusion } from './init.js';
-export { createDiffusionWeightLoader } from './weights.js';

package/src/inference/pipelines/diffusion/index.js

@@ -1,3 +0,0 @@
-export { DiffusionPipeline, createDiffusionPipeline } from './pipeline.js';
-export { mergeDiffusionConfig, initializeDiffusion } from './init.js';
-export { createDiffusionWeightLoader } from './weights.js';

package/src/inference/pipelines/diffusion/init.d.ts

@@ -1,24 +0,0 @@
-/**
- * Diffusion initialization utilities.
- *
- * @module inference/pipelines/diffusion/init
- */
-
-import type { DiffusionRuntimeConfig, DiffusionModelConfig } from './types.js';
-
-export interface DiffusionInitState {
-  modelConfig: DiffusionModelConfig;
-  runtime: DiffusionRuntimeConfig;
-  latentScale: number;
-  latentChannels: number;
-}
-
-export declare function mergeDiffusionConfig(
-  baseConfig: DiffusionRuntimeConfig | undefined | null,
-  overrideConfig: DiffusionRuntimeConfig | undefined | null
-): DiffusionRuntimeConfig;
-
-export declare function initializeDiffusion(
-  manifest: { config?: Record<string, unknown> },
-  runtimeConfig: { inference?: { diffusion?: DiffusionRuntimeConfig } }
-): DiffusionInitState;

package/src/inference/pipelines/diffusion/init.js

@@ -1,138 +0,0 @@
-import { DEFAULT_DIFFUSION_CONFIG } from '../../../config/schema/index.js';
-
-const SUPPORTED_DIFFUSION_RUNTIME_LAYOUTS = new Set(['sd3', 'flux', 'sana']);
-
-function mergeSection(base, override) {
-  if (!override) return { ...base };
-  return { ...base, ...override };
-}
-
-function mergeDecodeConfig(base, override) {
-  if (!override) return { ...base, tiling: { ...base.tiling } };
-  return {
-    ...base,
-    ...override,
-    tiling: mergeSection(base.tiling || {}, override.tiling || {}),
-  };
-}
-
-function mergeBackendConfig(base, override) {
-  if (!override) return { ...base };
-  return { ...base, ...override };
-}
-
-export function mergeDiffusionConfig(baseConfig, overrideConfig) {
-  const base = baseConfig || DEFAULT_DIFFUSION_CONFIG;
-  const override = overrideConfig || {};
-  return {
-    scheduler: mergeSection(base.scheduler, override.scheduler),
-    latent: mergeSection(base.latent, override.latent),
-    textEncoder: mergeSection(base.textEncoder, override.textEncoder),
-    decode: mergeDecodeConfig(base.decode, override.decode),
-    swapper: mergeSection(base.swapper, override.swapper),
-    quantization: mergeSection(base.quantization, override.quantization),
-    backend: mergeBackendConfig(base.backend, override.backend),
-  };
-}
-
-function resolveSchedulerType(modelScheduler, runtimeScheduler) {
-  const modelClass = modelScheduler?._class_name;
-  if (modelClass === 'FlowMatchEulerDiscreteScheduler') {
-    return 'flowmatch_euler';
-  }
-  if (modelClass === 'SCMScheduler') {
-    return 'scm';
-  }
-  if (modelClass === 'EulerDiscreteScheduler') {
-    return 'euler';
-  }
-  if (modelClass === 'EulerAncestralDiscreteScheduler') {
-    return 'euler_a';
-  }
-  if (modelClass === 'DPMSolverMultistepScheduler') {
-    return 'dpmpp_2m';
-  }
-  return runtimeScheduler?.type;
-}
-
-function mergeSchedulerConfig(modelConfig, runtimeScheduler) {
-  const modelScheduler = modelConfig?.components?.scheduler?.config || {};
-  const type = resolveSchedulerType(modelScheduler, runtimeScheduler);
-  return {
-    ...runtimeScheduler,
-    type,
-    numTrainTimesteps: modelScheduler.num_train_timesteps ?? runtimeScheduler.numTrainTimesteps,
-    shift: modelScheduler.shift ?? runtimeScheduler.shift,
-    predictionType: modelScheduler.prediction_type ?? runtimeScheduler.predictionType,
-    sigmaData: modelScheduler.sigma_data ?? runtimeScheduler.sigmaData,
-  };
-}
-
-function resolveLatentScale(modelConfig, runtimeConfig) {
-  const transformerSize = modelConfig?.components?.transformer?.config?.sample_size;
-  const vaeSize = modelConfig?.components?.vae?.config?.sample_size;
-  if (Number.isFinite(transformerSize) && Number.isFinite(vaeSize) && transformerSize > 0) {
-    const ratio = vaeSize / transformerSize;
-    if (Number.isFinite(ratio) && ratio > 0) {
-      return ratio;
-    }
-  }
-  const runtimeScale = runtimeConfig?.latent?.scale;
-  if (Number.isFinite(runtimeScale) && runtimeScale > 0) return runtimeScale;
-  throw new Error('Diffusion runtime requires a valid latent scale.');
-}
-
-function resolveLatentChannels(modelConfig, runtimeConfig) {
-  const vaeChannels = modelConfig?.components?.vae?.config?.latent_channels;
-  if (Number.isFinite(vaeChannels) && vaeChannels > 0) return vaeChannels;
-  const runtimeChannels = runtimeConfig?.latent?.channels;
-  if (Number.isFinite(runtimeChannels) && runtimeChannels > 0) return runtimeChannels;
-  throw new Error('Diffusion runtime requires valid latent channels.');
-}
-
-export function initializeDiffusion(manifest, runtimeConfig) {
-  const modelConfig = manifest?.config?.diffusion;
-  if (!modelConfig) {
-    const hasInferenceDiffusion = manifest?.inference?.diffusion && typeof manifest.inference.diffusion === 'object';
-    if (hasInferenceDiffusion) {
-      throw new Error(
-        'Diffusion manifest provides inference.diffusion, but runtime expects config.diffusion model contract. ' +
-        'Re-convert the model with config.diffusion populated.'
-      );
-    }
-    throw new Error('Diffusion manifest missing config.diffusion model contract.');
-  }
-  const layout = modelConfig.layout;
-  if (layout && !SUPPORTED_DIFFUSION_RUNTIME_LAYOUTS.has(layout)) {
-    throw new Error(
-      `Diffusion layout "${layout}" is recognized in the manifest, but the GPU runtime is not implemented yet. ` +
-      'Supported runtime layouts: sd3, flux, sana.'
-    );
-  }
-
-  const runtimeBase = mergeDiffusionConfig(DEFAULT_DIFFUSION_CONFIG, runtimeConfig?.inference?.diffusion);
-  const runtime = {
-    ...runtimeBase,
-    scheduler: mergeSchedulerConfig(modelConfig, runtimeBase.scheduler),
-  };
-  if (runtime.backend?.pipeline !== 'gpu') {
-    throw new Error(
-      `Diffusion runtime backend.pipeline must be "gpu"; got "${runtime.backend?.pipeline}".`
-    );
-  }
-  const transformerConfig = modelConfig?.components?.transformer?.config || null;
-  if (!transformerConfig) {
-    throw new Error(
-      'Diffusion runtime requires manifest.config.diffusion.components.transformer.config for GPU execution.'
-    );
-  }
-  const latentScale = resolveLatentScale(modelConfig, runtime);
-  const latentChannels = resolveLatentChannels(modelConfig, runtime);
-
-  return {
-    modelConfig,
-    runtime,
-    latentScale,
-    latentChannels,
-  };
-}

package/src/inference/pipelines/diffusion/pipeline.d.ts

@@ -1,38 +0,0 @@
-/**
- * Diffusion pipeline.
- *
- * @module inference/pipelines/diffusion/pipeline
- */
-
-import type { DiffusionRequest, DiffusionResult, DiffusionStats, DiffusionRuntimeConfig } from './types.js';
-import type { DiffusionWeightLoader, DiffusionWeightEntry } from './weights.js';
-
-export declare class DiffusionPipeline {
-  runtimeConfig: { inference?: { diffusion?: DiffusionRuntimeConfig } } | null;
-  manifest: Record<string, unknown> | null;
-  diffusionState: Record<string, unknown> | null;
-  tokenizers: Record<string, unknown> | null;
-  stats: DiffusionStats;
-  weightLoader: DiffusionWeightLoader | null;
-  vaeWeights: DiffusionWeightEntry | null;
-  textEncoderWeights: Record<string, DiffusionWeightEntry> | null;
-  transformerWeights: DiffusionWeightEntry | null;
-
-  initialize(contexts?: Record<string, unknown>): Promise<void>;
-  loadModel(manifest: Record<string, unknown>): Promise<void>;
-  getStats(): DiffusionStats;
-  getMemoryStats(): { used: number; kvCache: null };
-  unload(): Promise<void>;
-  ensureTextEncoderWeights(): Promise<Record<string, DiffusionWeightEntry>>;
-  ensureTransformerWeights(): Promise<DiffusionWeightEntry>;
-  releaseTextEncoderWeights(): void;
-  releaseTransformerWeights(): void;
-  generate(request: DiffusionRequest): Promise<DiffusionResult>;
-  generateCPU(request: DiffusionRequest): Promise<DiffusionResult>;
-  generateGPU(request: DiffusionRequest): Promise<DiffusionResult>;
-}
-
-export declare function createDiffusionPipeline(
-  manifest: Record<string, unknown>,
-  contexts?: Record<string, unknown>
-): Promise<DiffusionPipeline>;