Skip to content

TROUBLESHOOTING.md

Latest commit

 

History

History
50 lines (35 loc) · 1.21 KB

TROUBLESHOOTING.md

File metadata and controls

50 lines (35 loc) · 1.21 KB
Raw

Troubleshooting

WebGPU not available

Symptom:

  • App shows: WebGPU not available. Use Chrome 113+ or Edge 113+.

Fix:

  1. Use latest Chrome/Edge desktop.
  2. Verify hardware acceleration is enabled.
  3. Retry in a clean tab/profile.

Compare says prompt exceeds context window

Symptom:

  • Prompt tokens exceed context window size

Cause:

  • Prompt is longer than selected model context (e.g., >4096 on 4k model).

Fix:

  1. Use custom 8k model for long prompts.
  2. Reduce prompt length.
  3. Use sliding window mode where applicable.

Local custom model init fails

Symptom:

  • Errors around local files/wasm load.

Fix:

  1. Ensure required files exist: mlc-chat-config.json, tokenizer.json, tensor-cache.json, params_shard_*.bin.
  2. Re-add local files after page reload (session-only).
  3. Use matching .wasm for your model build.

Safari is much slower

Symptom:

  • Lower throughput and larger TTFT vs Chrome/Edge.

Fix:

  1. Use Safari Stable preset.
  2. Reduce prompt/output window.
  3. For publishable benchmarks, run Chrome/Edge primary and include Safari secondary.

VRAM numbers seem inconsistent

Explanation:

  • Browser WebGPU does not expose exact live GPU VRAM usage.
  • Treat VRAM as metadata/proxy signals.