Fix external_models import path and migrate model loading tests (#16458)

.github/workflows/pr-test.yml

@@ -594,7 +594,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        partition: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        partition: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -622,7 +622,7 @@ jobs:
           if [[ "${{ needs.check-changes.outputs.continue_on_error }}" == "true" ]]; then
             CONTINUE_ON_ERROR_FLAG="--continue-on-error"
           fi
-          python3 run_suite.py --hw cuda --suite stage-b-test-small-1-gpu --auto-partition-id ${{ matrix.partition }} --auto-partition-size 11 $CONTINUE_ON_ERROR_FLAG
+          python3 run_suite.py --hw cuda --suite stage-b-test-small-1-gpu --auto-partition-id ${{ matrix.partition }} --auto-partition-size 14 $CONTINUE_ON_ERROR_FLAG
 
   stage-b-test-large-1-gpu:
     needs: [check-changes, call-gate, stage-a-test-1, sgl-kernel-build-wheels]

python/sglang/srt/layers/moe/moe_runner/deep_gemm.py

@@ -45,7 +45,7 @@ _DEEPGEMM_ON_H20 = get_bool_env_var("SGLANG_DEEPGEMM_ON_H20")
 
 # TODO(kaixih@nvidia): ideally we should merge this logic into
 # `fill_gateup_input_triton_kernel` to directly generate e8m0 scale.
-@torch.compile
+@torch.compile(disable=_is_hip or _is_npu)
 def _cast_to_e8m0_with_rounding_up(x: torch.Tensor) -> torch.Tensor:
     temp = x.to(torch.float32).view(torch.int32)
     exp = torch.bitwise_right_shift(temp, 23)

test/registered/model_loading/test_external_models.py

@@ -2,13 +2,17 @@ import unittest
 
 import sglang as sgl
 from sglang.srt.environ import envs
+from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import CustomTestCase
 
+register_cuda_ci(est_time=30, suite="stage-b-test-small-1-gpu")
+register_amd_ci(est_time=45, suite="stage-b-test-small-1-gpu")
+
 
 class TestExternalModels(CustomTestCase):
     def test_external_model(self):
-        envs.SGLANG_EXTERNAL_MODEL_PACKAGE.set("external_models")
-        envs.SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE.set("external_models")
+        envs.SGLANG_EXTERNAL_MODEL_PACKAGE.set("sglang.test.external_models")
+        envs.SGLANG_EXTERNAL_MM_PROCESSOR_PACKAGE.set("sglang.test.external_models")
         prompt = "Today is a sunny day and I like"
         model_path = "Qwen/Qwen2-VL-2B-Instruct"
 

test/registered/model_loading/test_modelopt_export.py

@@ -17,6 +17,9 @@ from sglang.srt.configs.device_config import DeviceConfig
 from sglang.srt.configs.load_config import LoadConfig
 from sglang.srt.configs.model_config import ModelConfig
 from sglang.srt.model_loader.loader import ModelOptModelLoader
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=9, suite="stage-b-test-small-1-gpu")
 
 # Note: PYTHONPATH=python should be set when running tests
 

test/registered/model_loading/test_modelopt_loader.py

@@ -10,6 +10,14 @@ from unittest.mock import MagicMock, patch
 
 import torch.nn as nn
 
+from sglang.srt.configs.device_config import DeviceConfig
+from sglang.srt.configs.load_config import LoadConfig
+from sglang.srt.configs.model_config import ModelConfig
+from sglang.srt.layers.modelopt_utils import QUANT_CFG_CHOICES
+from sglang.srt.model_loader.loader import ModelOptModelLoader
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.test_utils import CustomTestCase
+
 # Note: PYTHONPATH=python should be set when running tests
 
 # Constants for calibration parameters to avoid hard-coded values
@@ -17,17 +25,7 @@ CALIBRATION_BATCH_SIZE = 36
 CALIBRATION_NUM_SAMPLES = 512
 DEFAULT_DEVICE = "cuda:0"
 
-# Constants for calibration parameters to avoid hard-coded values
-CALIBRATION_BATCH_SIZE = 36
-CALIBRATION_NUM_SAMPLES = 512
-DEFAULT_DEVICE = "cuda:0"
-
-from sglang.srt.configs.device_config import DeviceConfig
-from sglang.srt.configs.load_config import LoadConfig
-from sglang.srt.configs.model_config import ModelConfig
-from sglang.srt.layers.modelopt_utils import QUANT_CFG_CHOICES
-from sglang.srt.model_loader.loader import ModelOptModelLoader
-from sglang.test.test_utils import CustomTestCase
+register_cuda_ci(est_time=11, suite="stage-b-test-small-1-gpu")
 
 
 class TestModelOptModelLoader(CustomTestCase):

test/registered/model_loading/test_utils_update_weights.py

@@ -9,8 +9,11 @@ from transformers import AutoModelForCausalLM
 
 from sglang.srt.entrypoints.engine import Engine
 from sglang.srt.weight_sync.utils import update_weights
+from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import DEFAULT_SMALL_MODEL_NAME_FOR_TEST
 
+register_cuda_ci(est_time=29, suite="stage-b-test-small-1-gpu")
+
 
 class AsyncEngine(Engine):
     def __init__(self, **kwargs):

test/srt/run_suite.py

@@ -10,11 +10,7 @@ from sglang.test.ci.ci_utils import TestFile, run_unittest_files
 suites = {
     "per-commit-1-gpu": [
         TestFile("test_evs.py", 20),
-        TestFile("test_external_models.py", 30),
-        TestFile("test_modelopt_loader.py", 11),
-        TestFile("test_utils_update_weights.py", 29),
         TestFile("test_video_utils.py", 5),
-        TestFile("test_modelopt_export.py", 9),
     ],
     "per-commit-4-gpu": [
         TestFile("models/test_qwen3_next_models.py", 650),
@@ -92,7 +88,6 @@ suite_amd = {
         # TestFile("lora/test_lora_cuda_graph.py", 250), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
         # TestFile("lora/test_lora_qwen3.py", 97), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
         TestFile("test_bench_typebaseddispatcher.py", 10),
-        TestFile("test_external_models.py", 45),
         TestFile("test_rope_rocm.py", 3),
         # TestFile("test_torch_compile_moe.py", 210), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
         TestFile("test_type_based_dispatcher.py", 10),