Fix ONNX WebAssembly Model Loading (Transformers.js / Nginx / CSP)
Budget: $30 – $250 USD
I need an expert to debug and fix issues with running an ONNX language model in the browser using Transformers.js and ONNX Runtime WebAssembly.
The model, configs, and tokenizer files are already hosted on my server behind signed URLs. The problem is that the browser still tries to fetch from HuggingFace, crashes on WASM imports, or reports Unsupported model type: null.
You will not be given access to the production server. All work must be performed remotely on my machine via screen-share (for privacy and security reasons).
Key requirements (must all be met):
1. No HuggingFace dependency at runtime.
The app must not fetch from external endpoints like huggingface.co. Only signed URLs served by my domain are allowed.
2. Local manifest only.
Model files (onnx, tokenizer, config, generation_config) must load exclusively from the manifest/signed URLs I provide.
3. Correct ONNX Runtime WebAssembly setup.
Must handle both threaded (ort-wasm-simd-threaded.wasm) and non-threaded (ort-wasm-simd.wasm) binaries, depending on whether the page is crossOriginIsolated.
4. No WebAssembly import errors.
The fix must resolve runtime errors like
memory import must be a WebAssembly.Memory object and
no available backend found.
5. Cache handling.
Must prevent browser Cache API quota errors (e.g. QuotaExceededError) when large models are cached.
6. Strict Content Security Policy compliance.
All fetch requests must obey CSP headers and not be blocked.
7. Reliable model initialization.
The pipeline must always load with model_type set correctly (e.g. gpt2) so that Unsupported model type: null never occurs.
8. End-to-end working text-generation.
Final setup must allow generating outputs consistently in the browser with no dependency on remote services.
Your profile:
Strong experience with Transformers.js, ONNX Runtime Web, and WebAssembly.
Solid understanding of browser CSP and Nginx config.
Able to debug fetch/network issues in browser devtools.
Comfortable working live on my machine under supervision.
Important:
Do not ask for server credentials. All debugging will happen with me present.
Deliverables = working code/config on my machine that passes all 8 requirements.
The model, configs, and tokenizer files are already hosted on my server behind signed URLs. The problem is that the browser still tries to fetch from HuggingFace, crashes on WASM imports, or reports Unsupported model type: null.
You will not be given access to the production server. All work must be performed remotely on my machine via screen-share (for privacy and security reasons).
Key requirements (must all be met):
1. No HuggingFace dependency at runtime.
The app must not fetch from external endpoints like huggingface.co. Only signed URLs served by my domain are allowed.
2. Local manifest only.
Model files (onnx, tokenizer, config, generation_config) must load exclusively from the manifest/signed URLs I provide.
3. Correct ONNX Runtime WebAssembly setup.
Must handle both threaded (ort-wasm-simd-threaded.wasm) and non-threaded (ort-wasm-simd.wasm) binaries, depending on whether the page is crossOriginIsolated.
4. No WebAssembly import errors.
The fix must resolve runtime errors like
memory import must be a WebAssembly.Memory object and
no available backend found.
5. Cache handling.
Must prevent browser Cache API quota errors (e.g. QuotaExceededError) when large models are cached.
6. Strict Content Security Policy compliance.
All fetch requests must obey CSP headers and not be blocked.
7. Reliable model initialization.
The pipeline must always load with model_type set correctly (e.g. gpt2) so that Unsupported model type: null never occurs.
8. End-to-end working text-generation.
Final setup must allow generating outputs consistently in the browser with no dependency on remote services.
Your profile:
Strong experience with Transformers.js, ONNX Runtime Web, and WebAssembly.
Solid understanding of browser CSP and Nginx config.
Able to debug fetch/network issues in browser devtools.
Comfortable working live on my machine under supervision.
Important:
Do not ask for server credentials. All debugging will happen with me present.
Deliverables = working code/config on my machine that passes all 8 requirements.