debug: add prefill profiling after GPU init for per-layer timing

Author: sauravpanda

examples/benchmark/index.html

@@ -661,6 +661,12 @@ <h2>Comparison Charts</h2>
           const info = JSON.parse(flareEngine.backend_info());
           console.log('[Flare] backend_info:', info);
           log(`Flare backend: ${JSON.stringify(info)}`, gpuOk ? 'success' : 'info');
+          // Prefill profiling — captures per-layer timing
+          flareEngine.enable_prefill_profiling();
+          const profile = JSON.parse(flareEngine.prefill_profile_json());
+          console.log('[Flare] prefill profile:', profile);
+          log(`Flare prefill profile: ${JSON.stringify(profile)}`, 'info');
+          flareEngine.disable_prefill_profiling();
         } catch (e) {
           console.warn('[Flare] GPU init failed:', e);
           log('Flare GPU init failed — using CPU SIMD', 'info');

src/engines/flare-engine-wrapper.ts

@@ -60,6 +60,9 @@ interface FlareEngineInstance {
   add_stop_sequence(seq: string): void;
   clear_stop_sequences(): void;
   backend_info(): string;
+  enable_prefill_profiling(): void;
+  disable_prefill_profiling(): void;
+  prefill_profile_json(): string;
 }
 
 interface FlareModule {
@@ -311,7 +314,11 @@ export class FlareEngineWrapper {
     if (useGpu) {
       try {
         this.gpuEnabled = await this.engine.init_gpu();
-        console.log('[Flare]', JSON.parse(this.engine.backend_info()));
+        console.log('[Flare] backend_info:', JSON.parse(this.engine.backend_info()));
+        // Prefill profiling — captures per-layer timing for the first run
+        this.engine.enable_prefill_profiling();
+        console.log('[Flare] prefill profile:', JSON.parse(this.engine.prefill_profile_json()));
+        this.engine.disable_prefill_profiling();
         if (!this.gpuEnabled) {
           console.info('[Flare] WebGPU unavailable — using CPU SIMD path');
         }