@simulatte/doppler 0.1.7 → 0.1.8

package/src/tooling/conversion-config-materializer.js

@@ -2,6 +2,7 @@ import path from 'node:path';
 
 import { createConverterConfig } from '../config/schema/index.js';
 import { resolveConversionPlan } from '../converter/conversion-plan.js';
+import { normalizeQuantTag } from '../converter/quantization-info.js';
 
 function toSafeString(value) {
   if (typeof value !== 'string') return '';
@@ -10,10 +11,7 @@ function toSafeString(value) {
 }
 
 function normalizeQuantizationTag(value) {
-  const raw = toSafeString(value).toUpperCase();
-  if (!raw) return 'f16';
-  if (raw === 'Q4_K_M' || raw === 'Q4_K') return 'q4k';
-  return raw.toLowerCase();
+  return normalizeQuantTag(toSafeString(value));
 }
 
 function resolveArchitectureHint(architecture) {
@@ -37,7 +35,7 @@ function extractSourceQuantization(manifest) {
   if (explicitWeights) return explicitWeights;
   const explicitQuant = toSafeString(manifest?.quantization);
   if (explicitQuant) return explicitQuant;
-  return 'f16';
+  return normalizeQuantTag(null);
 }
 
 function buildRefreshRawConfig(manifest) {

package/src/tooling/node-converter.js

@@ -875,6 +875,7 @@ export async function convertSafetensorsDirectory(options) {
   let sourceQuantization = null;
   let tokenizerJson = null;
   let tokenizerConfig = null;
+  let generationConfig = null;
   let hasTokenizerModel = false;
   let tokenizerModelPath = null;
   let diffusionAuxFiles = [];
@@ -1101,6 +1102,7 @@ export async function convertSafetensorsDirectory(options) {
       },
     });
     config = parsedTransformer.config;
+    generationConfig = parsedTransformer.generationConfig ?? null;
     tensors = parsedTransformer.tensors;
     architectureHint = parsedTransformer.architectureHint;
     architecture = extractArchitecture(config, null);
@@ -1169,6 +1171,7 @@ export async function convertSafetensorsDirectory(options) {
     quantization: targetQuantization,
     tokenizerJson,
     tokenizerConfig,
+    generationConfig,
     tokenizerModel: hasTokenizerModel ? 'tokenizer.model' : null,
   };
 

package/src/tooling/node-source-runtime.js

@@ -411,6 +411,39 @@ function buildNodeFileReaders() {
   };
 }
 
+// Source dtype → compute precision mapping for source-runtime inference.
+// BF16/F32 sources require f32 compute (BF16 has no native WebGPU support).
+// Quantized formats require f32 compute for dequantization accuracy.
+// F16 sources can use f16 compute directly.
+const SOURCE_QUANT_COMPUTE_MAP = {
+  'F16': 'f16',
+  'BF16': 'f32',
+  'F32': 'f32',
+  'Q4_K': 'f32',
+  'Q4_K_M': 'f32',
+  'Q6_K': 'f32',
+};
+const SOURCE_COMPUTE_DEFAULT = 'f16';
+
+function resolveSourceRuntimeComputePrecision(tensors, sourceQuantization) {
+  const dtypes = new Set();
+  for (const tensor of Array.isArray(tensors) ? tensors : []) {
+    const dtype = String(tensor?.dtype || '').trim().toUpperCase();
+    if (dtype) {
+      dtypes.add(dtype);
+    }
+  }
+  // If any tensor requires f32 compute, use f32 for all.
+  for (const dtype of dtypes) {
+    if (SOURCE_QUANT_COMPUTE_MAP[dtype] === 'f32') {
+      return 'f32';
+    }
+  }
+
+  const normalized = String(sourceQuantization || '').trim().toUpperCase();
+  return SOURCE_QUANT_COMPUTE_MAP[normalized] ?? SOURCE_COMPUTE_DEFAULT;
+}
+
 async function addHashesToFileEntries(entries, hashAlgorithm) {
   const normalized = [];
   for (const entry of Array.isArray(entries) ? entries : []) {
@@ -473,6 +506,9 @@ export async function resolveNodeSourceRuntimeBundle(options = {}) {
   assertSupportedSourceDtypes(parsed.tensors, parsed.sourceKind);
 
   const converterConfig = createConverterConfig({
+    quantization: {
+      computePrecision: resolveSourceRuntimeComputePrecision(parsed.tensors, parsed.sourceQuantization),
+    },
     output: {
       modelBaseId: options.modelId || null,
     },

package/src/types/model.d.ts

@@ -9,7 +9,11 @@ export type ModelArchitecture =
   | 'gemma'
   | 'gemma2'
   | 'gemma3'
+  | 'embeddinggemma'
   | 'functiongemma'
+  | 'janus_text'
+  | 'lfm2'
+  | 'modernbert'
   | 'qwen2'
   | 'qwen3'
   | 'phi3'
@@ -19,6 +23,7 @@ export type ModelArchitecture =
   | 'deepseek'
   | 'mamba'
   | 'kimi_k2'
+  | 'translategemma'
   | 'transformer';
 
 /** Attention type variants */

package/tools/doppler-cli.js

@@ -1,5 +1,6 @@
 #!/usr/bin/env node
 
+import { existsSync } from 'node:fs';
 import fs from 'node:fs/promises';
 import path from 'node:path';
 import { fileURLToPath, pathToFileURL } from 'node:url';
@@ -13,7 +14,8 @@ import { createToolingErrorEnvelope } from '../src/tooling/command-envelope.js';
 
 const NODE_WEBGPU_INCOMPLETE_MESSAGE = 'node command: WebGPU runtime is incomplete in Node';
 const CLI_POLICY_PATH = fileURLToPath(new URL('./configs/cli/doppler-cli-policy.json', import.meta.url));
-const DEFAULT_EXTERNAL_MODELS_ROOT = process.env.DOPPLER_EXTERNAL_MODELS_ROOT || '/media/x/models';
+const DEFAULT_EXTERNAL_MODELS_ROOT = process.env.DOPPLER_EXTERNAL_MODELS_ROOT
+  || (existsSync('/Volumes/models') ? '/Volumes/models' : '/media/x/models');
 const DEFAULT_EXTERNAL_RDRR_ROOT = path.join(DEFAULT_EXTERNAL_MODELS_ROOT, 'rdrr');
 const DEFAULT_CLI_POLICY = {
   defaults: {
@@ -1260,6 +1262,9 @@ function printMetricsSummary(result) {
       `prefill=${formatNumber(metrics.prefillTokensPerSec)} ` +
       `decode=${formatNumber(metrics.decodeTokensPerSec)}`
     );
+    if (typeof result.output === 'string' && result.output.length > 0) {
+      console.log(`[output] ${quoteOneLine(result.output)}`);
+    }
     printExecutionContractSummary(result);
     printExecutionV0GraphSummary(metrics.executionV0GraphContractArtifact);
     return;