fix WASM (cpu)

Decisions.md

@@ -117,3 +117,23 @@ sidecar is absent).
 **Why:** CTranslate2 CPU inference is several times faster and more memory-efficient
 than ONNX Runtime under Bun; the server engine becomes genuinely useful rather than a
 slow convenience.
+
+## D-014 — WASM sessions run at graphOptimizationLevel "basic"
+**Status:** Accepted (bug workaround, 2026-07-10)
+On the WASM backend, the inference worker passes
+`session_options: { graphOptimizationLevel: "basic" }` to the Transformers.js
+pipeline. The onnxruntime-web build bundled with Transformers.js 4.2
+(1.26.0-dev.20260416) has a broken level-2 QDQ graph transform
+(`qdq_actions.cc` `TransposeDQWeightsForMatMulNBits`) that fails session
+creation for the 8-bit Whisper decoders (`_quantized`/`_int8`/`_uint8`), whose
+tied `embed_tokens` weight is both gathered (embedding) and matmul'd (lm_head):
+"Missing required scale: model.decoder.embed_tokens.weight_merged_0_scale".
+Reproduced outside the browser against `decoder_model_merged_quantized.onnx`
+of `onnx-community/whisper-tiny`; levels "all"/"extended" fail, "basic"/
+"disabled" load fine, and onnxruntime-node 1.24.3 loads the same file at every
+level (regression in the ort-web dev build, not the model).
+**Scope:** WASM only — WebGPU keeps the default level. The q4 conversions use
+native `MatMulNBits` ops (no QDQ), so skipping level-2 costs the larger WASM
+models nothing.
+**Revisit:** when Transformers.js bumps its onnxruntime-web dependency past
+1.26.0-dev.20260416, retest q8-on-WASM at the default level and drop this.

src/worker/inference.worker.ts

@@ -136,6 +136,14 @@ async function createPipeline(
       const asr = await pipeline("automatic-speech-recognition", message.modelId, {
         device,
         dtype,
+        // WASM sessions cap graph optimization at "basic" (D-014): the level-2
+        // QDQ transform in the bundled onnxruntime-web build rejects the 8-bit
+        // decoders' tied embed_tokens weight ("Missing required scale …
+        // DequantizeLinear"). q4 models use native MatMulNBits ops, so the
+        // lower level costs them nothing. WebGPU keeps the default level.
+        ...(device === "wasm" && {
+          session_options: { graphOptimizationLevel: "basic" as const },
+        }),
         progress_callback: forwardProgress,
       });
       post({ type: "load-progress", status: "ready", device });

tests/offline.md

@@ -61,6 +61,14 @@ model.decoder.embed_tokens.weight_merged_0_scale for node:
 model.decoder.embed_tokens.weight_transposed_DequantizeLinear
 ```
 
+**Diagnosed & fixed (2026-07-10, D-014):** a level-2 graph-optimizer bug in
+the onnxruntime-web build bundled with Transformers.js 4.2, triggered by the
+8-bit decoders' tied `embed_tokens` weight — not a model or cache problem.
+The worker now caps WASM sessions at `graphOptimizationLevel: "basic"`
+(`src/worker/inference.worker.ts`), verified against the same ort-web build
+outside the browser. **Re-run this step with Backend = WASM to confirm in a
+real browser.**
+
 ## 3. Go offline and transcribe
 
 1. DevTools → Network → throttling **Offline** (or OS airplane mode — stronger,