### Check for existing issues

I searched the existing issues; this is not a duplicate. The closest is the now-closed #28084 (a missing `vertexai` import), which is a different root cause — the problem described here persists with the dependency installed.

### What happened?

`POST /v1/messages/count_tokens` returns **HTTP 500** when a provider-specific token counter *raises* an exception, instead of falling back to the local tokenizer.

The provider dispatch in `_try_provider_token_count` (`litellm/proxy/proxy_server.py`) only handles two failure modes gracefully:

- `httpx.HTTPStatusError` → re-raised as a `ProxyException`. - the returned `result.error is True` → warning + fall back to the local tokenizer (unless `disable_token_counter` is set).

Any **other** exception raised by `provider_counter.count_tokens()` is not caught, so it propagates out of the `count_tokens` endpoint handler and becomes a 500.

The Vertex AI partner-models token counter (`litellm/llms/vertex_ai/common_utils.py` → `vertex_ai_partner_models`) raises a raw exception on failure (e.g. `litellm.Timeout`, or an ADC/credentials error when credentials are absent), which hits this uncaught path.

By contrast, …