# SPDX-FileCopyrightText: © 2026 Tenstorrent USA, Inc. # SPDX-License-Identifier: Apache-2.0 """checkpoint_name() must give one name for every way a checkpoint is addressed. The warm weight-cache marker records ``model_name``; the demo seeds it from the HF id while vLLM (under HF_HUB_OFFLINE) hands over the resolved hub snapshot directory. If the two spell the name differently the vLLM server never sees the marker and cold-loads the full checkpoint every start. """ import pytest from models.common.weight_cache import checkpoint_name HUB = "/mnt/MLPerf/huggingface/hub" @pytest.mark.parametrize( "model_path, expected", [ ("google/gemma-4-31B-it", "gemma-4-31B-it"), ("google/gemma-4-31B-it/", "gemma-4-31B-it"), (f"{HUB}/models--google--gemma-4-31B-it/snapshots/842da37", "gemma-4-31B-it"), (f"{HUB}/models--google--gemma-4-31B-it/snapshots/842da37/", "gemma-4-31B-it"), (f"{HUB}/models--meta-llama--Llama-3.1-8B-Instruct/snapshots/0e9e39f", "Llama-3.1-8B-Instruct"), ("/mnt/MLPerf/tt_dnn-models/google/gemma-4-26B-A4B-it", "gemma-4-26B-A4B-it"), ("gemma-4-E2B-it", "gemma-4-E2B-it"), ], ) def test_checkpoint_name(model_path, expected): assert checkpoint_name(model_path) == expected def test_hf_id_and_hub_snapshot_agree(): hf_id = "google/gemma-4-31B-it" snapshot = f"{HUB}/models--google--gemma-4-31B-it/snapshots/842da37" assert checkpoint_name(hf_id) == checkpoint_name(snapshot)