Skip to content

Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠

Audit local function bodies; take long-prompt path and dtype from the presets

Claudef61514320de3
diff --git a/scripts/check-webgpu-long-prompts.mjs b/scripts/check-webgpu-long-prompts.mjs
index 657a729b8..e2a245dba 100644
--- a/scripts/check-webgpu-long-prompts.mjs
+++ b/scripts/check-webgpu-long-prompts.mjs
@@ -20,8 +20,12 @@
// node scripts/check-webgpu-long-prompts.mjs --model webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision \
// --vision --profile .build/bonsai/chunked-2b-profile --offline --lengths 1000,4000,8000 --slices 0,512
//
-// A repository whose name ends in -Vision always uses the multimodal path, as
-// in production; --vision forces it for other repositories.
+// Shipped presets use their production path and dtype from
+// src/chrome/src/providers/webgpu.js: `onnx-vl` presets run multimodal chat,
+// and other ONNX presets run text chat with their preset dtype (q4f16 by
+// default, as for custom repositories). --vision forces multimodal chat and
+// --dtype (a name or a JSON object) overrides the dtype. The `bitgpu` runtime
+// does not use this worker and is rejected.
//
// Options mirror check-bonsai-webgpu.mjs: --local-repo owner/repo=dir serves
// unpublished files, --offline requires a cached model and blocks Hugging Face,
@@ -34,17 +38,21 @@ import http from 'node:http';
import path from 'node:path';
import { fileURLToPath } from 'node:url';
import { chromium } from 'playwright';
+import {
+ WEBGPU_RUNTIME_BITGPU, WEBGPU_RUNTIME_ONNX_VL, webgpuModelDtype, webgpuModelRuntime,
+} from '../src/chrome/src/providers/webgpu.js';
const root = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '..');
const args = process.argv.slice(2);
const option = (name, fallback) => { const i = args.indexOf(name); return i < 0 ? fallback : args[i + 1]; };
const model = option('--model', 'webbrain-one/Ternary-Bonsai-2-2B-ONNX-WebGPU-Vision');
-// The Vision presets always run through the multimodal path in production.
-const vision = args.includes('--vision') || /-Vision$/i.test(model);
+const runtime = webgpuModelRuntime(model);
+assert(runtime !== WEBGPU_RUNTIME_BITGPU, `${model} uses the ${runtime} runtime, which this ONNX worker check does not cover`);
+// As in production, onnx-vl presets always use the multimodal path.
+const vision = args.includes('--vision') || runtime === WEBGPU_RUNTIME_ONNX_VL;
const offline = args.includes('--offline');
-const dtype = option('--dtype', null)
- ? JSON.parse(option('--dtype'))
- : vision ? { embed_tokens: 'q2f16', vision_encoder: 'fp16', decoder_model_merged: 'q2f16' } : 'q2f16';
+const parseDtype = value => { try { return JSON.parse(value); } catch { return value; } };
+const dtype = option('--dtype', null) ? parseDtype(option('--dtype')) : webgpuModelDtype(model);
const lengths = option('--lengths', '1000,4000,8000').split(',').map(Number);
const slices = option('--slices', '0,512').split(',').map(Number);
const maxTokens = Number(option('--max-tokens', 48));
diff --git a/scripts/models/audit_webgpu_onnx.py b/scripts/models/audit_webgpu_onnx.py
index 229672c3f..e434aaa4f 100644
--- a/scripts/models/audit_webgpu_onnx.py
+++ b/scripts/models/audit_webgpu_onnx.py
@@ -4,9 +4,10 @@ Catches the problems that made Ternary Bonsai 2 prefill take minutes in the
browser (see docs/webgpu-models.md):
* control-flow ops that ONNX Runtime Web's WebGPU execution provider does not
- implement (`Scan`, `Loop`), including inside `If` branches. ONNX Runtime
- places them on the CPU and their bodies then run once per iteration with a
- GPU round trip, so a per-token recurrence becomes a per-token CPU loop;
+ implement (`Scan`, `Loop`), including inside `If` branches and model-local
+ functions. ONNX Runtime places them on the CPU and their bodies then run
+ once per iteration with a GPU round trip, so a per-token recurrence becomes
+ a per-token CPU loop;
* a `logits` output that keeps the sequence dimension, so prefill computes and
downloads `[sequence_length, vocab]` scores (0.5 MB per prompt token for a
248k vocabulary in FP16) and can exceed a 2 GB WebGPU buffer limit.
@@ -62,7 +63,11 @@ def audit(path, text_generate_only=False):
control_flow = collections.defaultdict(list)
op_counts = collections.Counter()
- for node, location in walk(graph):
+ # Model-local functions are inlined by ONNX Runtime, so audit their bodies too.
+ bodies = [(graph, "")] + [(function, f"function:{function.domain}::{function.name}")
+ for function in model.functions]
+ nodes = [entry for body, location in bodies for entry in walk(body, location)]
+ for node, location in nodes:
domain = node.domain or "ai.onnx"
op_counts[f"{domain}::{node.op_type}"] += 1
if node.op_type in CPU_ONLY_CONTROL_FLOW and domain == "ai.onnx":
@@ -104,11 +109,13 @@ def audit(path, text_generate_only=False):
"the only caller.")
blockers.append(entry)
- contrib = sorted(k for k in op_counts if not k.startswith("ai.onnx::"))
+ local_functions = {f"{function.domain}::{function.name}" for function in model.functions}
+ contrib = sorted(k for k in op_counts if not k.startswith("ai.onnx::") and k not in local_functions)
info.update({
"opsets": {o.domain or "ai.onnx": o.version for o in model.opset_import},
"nodes": sum(op_counts.values()),
"if_nodes": op_counts.get("ai.onnx::If", 0),
+ "local_functions": sorted(local_functions),
"contrib_ops": {k: op_counts[k] for k in contrib},
"external_data_files": sorted({
entry.value for tensor in graph.initializer if tensor.data_location == onnx.TensorProto.EXTERNAL