diff --git a/docs/npu.md b/docs/npu.md index 507a0fe3e..6947d3c33 100644 --- a/docs/npu.md +++ b/docs/npu.md @@ -2681,7 +2681,11 @@ that runs the phase is private, like the kernels it runs. `tests/data/qwen3_tokenizer_golden.json` holds 14 cases encoded by `tokenizers` 0.23.1, and all 14 match, ids and round trip. - **Endpoints:** `/v1/health` (503 while loading), `/v1/models`, `/v1/chat/completions` - (streaming or not) and `/v1/completions`. + (streaming or not) and `/v1/completions`. The serve also answers `/health` beside `/v1/health`, and there is + **no `/api/v1/*` prefix**: a probe against `/api/v1/health` is a 404, `curl -s` exits 0 on a 404, and a + readiness check written that way therefore passes while the weights are still loading -- the request then + comes back `{"error":{"message":"model is loading"}}`. Use `curl -sf /health` (measured: `/api/v1/health` + 404, `/health` 503 while loading then 200; `scripts/npu-window11.sh` is the house form of this probe). - **Prompt:** the **container's own chat template** — `tokenizer_config.json`'s `chat_template` string, else `chat_template.jinja` beside it — rendered with the pinned llama.cpp fork's Jinja engine (`app/npu_template.h`, `docs/npu-chat-template.md`). A container that declares no template keeps ChatML