Run WASM sessions at the default graph optimization level now that ort-web 1.31 loads the 8-bit decoders

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs/Decisions.md

@@ -125,7 +125,8 @@ than ONNX Runtime under Bun; the server engine becomes genuinely useful rather t
 slow convenience.
 
 ## D-014 — WASM sessions run at graphOptimizationLevel "basic"
-**Status:** Accepted (bug workaround, 2026-07-10)
+**Status:** **Superseded by D-WS1c** (bug workaround, 2026-07-10; the ort-web
+build in Transformers.js 4.3.2 no longer has the bug, so the cap is gone)
 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
@@ -1319,3 +1320,43 @@ share the CPU), so read it only as "the seek loop is slower", not as a
 version-to-version speed comparison.
 **Decision:** keep `max_new_tokens: 440` in `src/worker/inference.worker.ts`
 and in `tests/live-real-model.test.ts`.
+
+## D-WS1c — WASM sessions back at the default graph optimization level (D-014's "basic" cap dropped)
+**Status:** Accepted (2026-10-30) — supersedes D-014
+**Re-test** of D-014's revisit condition after the Transformers.js 4.3.2 bump
+(D-WS1a), which brings onnxruntime-web `1.31.0-dev.20260914-8d85527a0`.
+Measured under Bun with onnxruntime-web's own wasm EP (not onnxruntime-node),
+1 thread, AMD Ryzen 9 9950X shared with other workstreams.
+1. **Session creation**, `onnx-community/whisper-tiny`
+   `decoder_model_merged_quantized.onnx` (D-014's repro file), every level:
+
+   | onnxruntime-web | disabled | basic | extended | all |
+   |---|---|---|---|---|
+   | 1.26.0-dev.20260416 (control) | OK | OK | FAIL | FAIL |
+   | 1.31.0-dev.20260914 | OK | OK | OK | OK |
+
+   The control still fails with D-014's exact error (`qdq_actions.cc:137
+   TransposeDQWeightsForMatMulNBits Missing required scale:
+   model.decoder.embed_tokens.weight_merged_0_scale`); 1.31 loads at all four.
+2. **Transcription**, Transformers.js 4.3.2 pipeline on that ort-web build,
+   `whisper-tiny.en` q8 (catalog WASM default), `en_aiproblem` (14.4 s),
+   worker options (30 s chunks, 5 s stride, timestamps, `max_new_tokens:
+   440`). Both pipelines in one process, 1 warm-up each, then 8 interleaved
+   runs (A/B order alternating) so machine-load drift hits both equally:
+
+   | level | warm runs (s) | median | WER | text |
+   |---|---|---|---|---|
+   | `basic` (D-014) | 1.49 1.44 1.36 1.39 1.41 1.47 1.48 1.60 | 1.47 s | 0.000 | identical |
+   | default (`all`) | 1.38 1.47 1.34 1.39 1.37 1.42 1.43 1.43 | 1.42 s | 0.000 | identical |
+
+   Same transcript byte-for-byte at both levels and across every run; the
+   default level is ~3 % faster here (within noise; certainly not slower).
+**Decision:** `src/worker/inference.worker.ts` no longer passes
+`session_options` — WASM and WebGPU both use onnxruntime's default level.
+`tests/transformers-runtime.test.ts` pins the onnxruntime-web version this
+was measured on and that the worker sets no `graphOptimizationLevel`; a
+different ort-web build fails the fast suite until this session-creation
+check is re-run.
+**Not covered here:** a real browser's ort-web wasm build (the same
+`ort-wasm-simd-threaded.*` files, but multithreaded under COOP/COEP) —
+`tests/offline.md` §2's WASM step is the owner's browser check.

docs/WorkBreakdown.md

@@ -342,6 +342,8 @@ cached.
 ## M8 — Transcription quality and speed (spec 001)
 - [x] Transformers.js 4.2.0 → 4.3.2, sequential-session patch regenerated
       for 4.3.2 and guarded by `tests/transformers-runtime.test.ts` (D-WS1a).
+- [x] WASM sessions back at the default graph optimization level: ort-web
+      1.31.0-dev loads the q8 decoders at every level (D-WS1c supersedes D-014).
 
 ---
 

docs/plans/001-1-transformers-upgrade.md

@@ -98,12 +98,12 @@ onnxruntime-node 1.30.0), bun:test.
 
 **Files:** `docs/Decisions.md`; `src/worker/inference.worker.ts` only if the outcome is "drop".
 
-- [ ] **Step 1: Measure** (`/tmp/ws1/d014.ts`): `globalThis[Symbol.for("onnxruntime")] =
+- [x] **Step 1: Measure** (`/tmp/ws1/d014.ts`): `globalThis[Symbol.for("onnxruntime")] =
   await import("onnxruntime-web")` before importing transformers, wasm EP, 1 thread; for
   `onnx-community/whisper-tiny` `decoder_model_merged_quantized.onnx` create a session at
   `disabled/basic/extended/all`; record success/error. If all load: transcribe `en_aiproblem` with
   tiny.en q8 at `basic` vs default (`all`), 3 warm runs each, wall-clock + text identity.
-- [ ] **Step 2:** If default fails → keep, `D-WS1c` "still needed on ort-web 1.31.0-dev". If it
+- [x] **Step 2:** If default fails → keep, `D-WS1c` "still needed on ort-web 1.31.0-dev". If it
   loads and is not slower → remove the override (worker comment cites D-WS1c), run a q8 transcribe
   again through the harness with the worker's new options. Commit.
 

src/worker/inference.worker.ts

@@ -260,14 +260,12 @@ async function createPipeline(
       const pipelineOptions: PretrainedModelOptions & { sequentialSessionLoad?: boolean } = {
         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 },
-        }),
+        // No session_options: every backend runs at onnxruntime's default
+        // graph optimization level. D-014's WASM "basic" cap worked around a
+        // level-2 QDQ bug in ort-web 1.26.0-dev that rejected the 8-bit
+        // decoders; the 1.31.0-dev build bundled with Transformers.js 4.3.2
+        // loads them at every level (D-WS1c, pinned in
+        // tests/transformers-runtime.test.ts).
         progress_callback: forwardProgress,
         // D-027/D-029: sequential sub-model session loading trades load speed
         // for a lower peak-memory footprint — only worth it where the UI

tests/offline.md

@@ -69,6 +69,11 @@ The worker now caps WASM sessions at `graphOptimizationLevel: "basic"`
 outside the browser. **Re-run this step with Backend = WASM to confirm in a
 real browser.**
 
+**Update (2026-10-30, D-WS1c):** the Transformers.js 4.3.2 bump brings an
+onnxruntime-web build without that bug, so the cap is gone and WASM sessions
+run at the default level again. Re-run this step with Backend = WASM (Whisper
+Tiny and Tiny.en) to confirm the session still loads in a real browser.
+
 ## 3. Go offline and transcribe
 
 1. DevTools → Network → throttling **Offline** (or OS airplane mode — stronger,

tests/transformers-runtime.test.ts

@@ -54,4 +54,17 @@ describe("@huggingface/transformers runtime", () => {
     const source = read(join(PKG_DIR, "src/models/whisper/modeling_whisper.js"));
     expect(source).toContain("if (generation_config.return_timestamps && !kwargs.max_new_tokens)");
   });
+
+  test("onnxruntime-web is the build D-WS1c re-tested 8-bit decoders on at the default optimization level", () => {
+    // D-014's "basic" cap was dropped because this build loads the q8 merged
+    // decoders at every graphOptimizationLevel. A different build needs the
+    // D-WS1c session-creation check re-run before it ships.
+    const ortWeb = JSON.parse(read(join(ROOT, "node_modules/onnxruntime-web/package.json"))).version;
+    expect(ortWeb).toBe("1.31.0-dev.20260914-8d85527a0");
+  });
+
+  test("worker leaves WASM sessions at the default graph optimization level (D-WS1c)", () => {
+    const worker = read(join(ROOT, "src/worker/inference.worker.ts"));
+    expect(worker).not.toContain("graphOptimizationLevel");
+  });
 });