mlx-bun 0.0.4

package/src/model/factory.ts

@@ -0,0 +1,27 @@
+// Model construction with generated-specialization dispatch
+// (docs/design/optimization_plan.md Phase C): pick the generated class whose config
+// fingerprint matches, else the monolith. Generated classes subclass
+// Gemma4Model and only override forwardLayers (with their own
+// cache-signature guard), so the choice is always safe.
+
+import type { ModelConfig } from "../config";
+import type { Weights } from "../weights";
+import { Gemma4Model } from "./gemma4";
+import { configFingerprint } from "./fingerprint";
+import { GENERATED } from "./generated";
+import { MiniCPM5Model } from "./minicpm5";
+import { Qwen35Model } from "./qwen3_5";
+import { isMiniCPM5Config, isQwen35Config } from "./support";
+
+export type RuntimeModel = Gemma4Model | MiniCPM5Model | Qwen35Model;
+
+export function createModel(weights: Weights, config: ModelConfig): RuntimeModel {
+  if (isMiniCPM5Config(config)) return new MiniCPM5Model(weights, config);
+  if (isQwen35Config(config)) return new Qwen35Model(weights, config);
+  if (config.modelType === "llama")
+    throw new Error("unsupported llama config: only MiniCPM5-1B-OptiQ-4bit is wired");
+  if (config.modelType.startsWith("qwen3_5"))
+    throw new Error(`unsupported qwen3_5 config (MoE variants are deferred)`);
+  const cls = GENERATED.get(configFingerprint(config)) ?? Gemma4Model;
+  return new cls(weights, config);
+}

package/src/model/fingerprint.ts

@@ -0,0 +1,33 @@
+// Config fingerprint for generated-model dispatch (docs/design/optimization_plan.md
+// Phase C): a stable hash over every config field that changes the
+// decode graph structure, plus the kv_config quant layout. A generated
+// specialization is used only when the fingerprint of the loaded config
+// matches the one it was generated from; everything else runs the
+// monolith (slow, never broken).
+
+import type { ModelConfig } from "../config";
+
+export function configFingerprint(config: ModelConfig): string {
+  const t = config.text;
+  const payload = JSON.stringify({
+    layerTypes: t.layerTypes,
+    hiddenSize: t.hiddenSize,
+    headDim: t.headDim,
+    globalHeadDim: t.globalHeadDim,
+    numAttentionHeads: t.numAttentionHeads,
+    numKeyValueHeads: t.numKeyValueHeads,
+    numGlobalKeyValueHeads: t.numGlobalKeyValueHeads,
+    attentionKEqV: t.attentionKEqV,
+    numKvSharedLayers: t.numKvSharedLayers,
+    enableMoeBlock: t.enableMoeBlock,
+    numExperts: t.enableMoeBlock ? t.numExperts : 0,
+    topKExperts: t.enableMoeBlock ? t.topKExperts : 0,
+    hiddenSizePerLayerInput: t.hiddenSizePerLayerInput,
+    slidingWindow: t.slidingWindow,
+    finalLogitSoftcapping: t.finalLogitSoftcapping,
+    kvQuant: config.kvQuant ?? null,
+  });
+  const h = new Bun.CryptoHasher("sha256");
+  h.update(payload);
+  return h.digest("hex").slice(0, 16);
+}