From 583cfc4d55dda55b4b4de713b1ba7c728f095e34 Mon Sep 17 00:00:00 2001 From: Baizhou Zhang Date: Tue, 20 Oct 2026 14:57:01 -0700 Subject: [PATCH] [CI] muse_spark: Use complete local checkpoints in model tests (#45428) --- python/sglang/test/muse_spark_utils.py | 56 +++++++++++++++++++ .../e2e/models/test_muse_spark_12_b300_pd.py | 10 ++-- .../models/test_muse_spark_unified_memory.py | 10 ++-- 3 files changed, 66 insertions(+), 10 deletions(-) create mode 100644 python/sglang/test/muse_spark_utils.py diff --git a/python/sglang/test/muse_spark_utils.py b/python/sglang/test/muse_spark_utils.py new file mode 100644 index 0000000000..1c12a6d0e5 --- /dev/null +++ b/python/sglang/test/muse_spark_utils.py @@ -0,0 +1,56 @@ +"""Local checkpoint resolution for the Muse-Spark serving tests.""" + +import json +from pathlib import Path + +from sglang.srt.utils import find_local_repo_dir +from sglang.test.test_utils import try_cached_model + + +def try_cached_muse_spark_model(model_repo: str, *, draft: bool = False) -> str: + """Use a complete local snapshot, or retain the Hub ID for downloading.""" + required_files = ( + () + if draft + else ( + "tokenizer.json", + "tokenizer_config.json", + "special_tokens_map.json", + "l4_200k_base", + "processor_config.json", + "chat_template.jinja", + "generation_config.json", + "hf_quant_config.json", + ) + ) + model_dir = try_cached_model(model_repo) + candidates = [model_dir] if model_dir != model_repo else [] + try: + candidates.append(find_local_repo_dir(model_repo)) + except OSError: + pass + for model_dir in candidates: + if not model_dir: + continue + snapshot = Path(model_dir) + try: + index = snapshot / "model.safetensors.index.json" + if index.exists() or index.is_symlink(): + weight_map = json.loads(index.read_text())["weight_map"] + if not isinstance(weight_map, dict) or not weight_map: + continue + weights = set(weight_map.values()) + else: + weights = {"model.safetensors"} + files = {"config.json", *required_files, *weights} + if not all( + (snapshot / name).is_file() and (snapshot / name).stat().st_size > 0 + for name in files + ): + continue + except (OSError, ValueError, KeyError, TypeError): + continue + + print(f"Using complete local checkpoint for {model_repo}: {snapshot}") + return str(snapshot.absolute()) + return model_repo diff --git a/test/registered/e2e/models/test_muse_spark_12_b300_pd.py b/test/registered/e2e/models/test_muse_spark_12_b300_pd.py index e0c2da0952..63e9718d92 100644 --- a/test/registered/e2e/models/test_muse_spark_12_b300_pd.py +++ b/test/registered/e2e/models/test_muse_spark_12_b300_pd.py @@ -11,6 +11,7 @@ import requests from sglang.test.ci.ci_register import register_cuda_ci from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin +from sglang.test.muse_spark_utils import try_cached_muse_spark_model from sglang.test.server_fixtures.disaggregation_fixture import ( PDDisaggregationServerBase, ) @@ -18,7 +19,6 @@ from sglang.test.sgl_eval_utils import run_sgl_eval from sglang.test.test_utils import ( _wait_for_gpu_idle_in_ci, is_in_ci, - try_cached_model, write_github_step_summary, ) @@ -53,12 +53,12 @@ class TestMuseSpark12B300PD(SpecDecodingMixin, PDDisaggregationServerBase): @classmethod def setUpClass(cls): super().setUpClass() - # Use the gated-model credential only in this test and its server children. - if token := os.environ.get("HF_MUSE_SPARK_TOKEN"): + cls.model = try_cached_muse_spark_model(MODEL_PATH) + # A complete local checkpoint needs no gated-model credential. + if cls.model == MODEL_PATH and (token := os.environ.get("HF_MUSE_SPARK_TOKEN")): token_env = patch.dict(os.environ, {"HF_TOKEN": token}) token_env.start() cls.addClassCleanup(token_env.stop) - cls.model = try_cached_model(MODEL_PATH) cls.transfer_backend = ["--disaggregation-transfer-backend", "mooncake"] common_args = [ "--quantization", @@ -80,7 +80,7 @@ class TestMuseSpark12B300PD(SpecDecodingMixin, PDDisaggregationServerBase): "--speculative-algorithm", "DFLASH", "--speculative-draft-model-path", - try_cached_model(DRAFT_MODEL_PATH), + try_cached_muse_spark_model(DRAFT_MODEL_PATH, draft=True), "--speculative-draft-attention-backend", "fa4", "--speculative-num-draft-tokens", diff --git a/test/registered/e2e/models/test_muse_spark_unified_memory.py b/test/registered/e2e/models/test_muse_spark_unified_memory.py index d6d89f7295..12a43bc1a9 100644 --- a/test/registered/e2e/models/test_muse_spark_unified_memory.py +++ b/test/registered/e2e/models/test_muse_spark_unified_memory.py @@ -8,12 +8,12 @@ from types import SimpleNamespace from unittest.mock import patch from sglang.test.ci.ci_register import register_cuda_ci +from sglang.test.muse_spark_utils import try_cached_muse_spark_model from sglang.test.server_fixtures.default_fixture import DefaultServerBase from sglang.test.sgl_eval_utils import run_sgl_eval from sglang.test.test_utils import ( _wait_for_gpu_idle_in_ci, is_in_ci, - try_cached_model, write_github_step_summary, ) @@ -29,12 +29,12 @@ class TestMuseSparkUnifiedMemory(DefaultServerBase): @classmethod def setUpClass(cls): - # Use the gated-model credential only in this test and its server children. - if token := os.environ.get("HF_MUSE_SPARK_TOKEN"): + cls.model = try_cached_muse_spark_model(MODEL_PATH) + # A complete local checkpoint needs no gated-model credential. + if cls.model == MODEL_PATH and (token := os.environ.get("HF_MUSE_SPARK_TOKEN")): token_env = patch.dict(os.environ, {"HF_TOKEN": token}) token_env.start() cls.addClassCleanup(token_env.stop) - cls.model = try_cached_model(MODEL_PATH) cls.other_args = [ "--tp-size", "4", @@ -57,7 +57,7 @@ class TestMuseSparkUnifiedMemory(DefaultServerBase): "--speculative-algorithm", "DFLASH", "--speculative-draft-model-path", - try_cached_model(DRAFT_MODEL_PATH), + try_cached_muse_spark_model(DRAFT_MODEL_PATH, draft=True), "--speculative-draft-attention-backend", "fa4", "--speculative-num-draft-tokens",