[CI] Re-enable B300 jobs (#35607)

Co-authored-by: Mohammad Angkad <mohammad.angkad@radixark.ai>

.github/workflows/pr-test.yml

@@ -471,21 +471,20 @@ jobs:
       skip_prebuilt_rust_ext: true
     secrets: inherit
 
-  # Temporarily disabled.
-  # base-c-test-8-gpu-b300:
-  #   needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
-  #   if: ${{ !failure() && !cancelled() }}
-  #   uses: ./.github/workflows/_pr-test-stage.yml
-  #   with:
-  #     self_name: base-c-test-8-gpu-b300
-  #     runner_config: 8-gpu-b300
-  #     check_changes: ${{ toJson(needs.check-changes.outputs) }}
-  #     caller_inputs: ${{ toJson(inputs) }}
-  #     partitions: ${{ needs.check-changes.outputs.partitions }}
-  #     run_timeout_minutes: '60'
-  #     timeout_per_file: '3600'
-  #     rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
-  #   secrets: inherit
+  base-c-test-8-gpu-b300:
+    needs: [check-changes, call-gate, wait-for-base-b, sgl-kernel-build-wheels, rust-ext-build]
+    if: ${{ !failure() && !cancelled() }}
+    uses: ./.github/workflows/_pr-test-stage.yml
+    with:
+      self_name: base-c-test-8-gpu-b300
+      runner_config: 8-gpu-b300
+      check_changes: ${{ toJson(needs.check-changes.outputs) }}
+      caller_inputs: ${{ toJson(inputs) }}
+      partitions: ${{ needs.check-changes.outputs.partitions }}
+      run_timeout_minutes: '60'
+      timeout_per_file: '3600'
+      rust_ext_artifact: ${{ needs.rust-ext-build.outputs.artifact_name }}
+    secrets: inherit
 
   # List every build and test job: `skipped` passes here, so an omission turns that
   # job's failure into a green run with no tests.
@@ -519,7 +518,7 @@ jobs:
         base-c-test-8-gpu-h200,
         base-c-test-4-gpu-b200,
         base-c-test-4-gpu-gb300,
-        # base-c-test-8-gpu-b300,  # Temporarily disabled.
+        base-c-test-8-gpu-b300,
       ]
     if: always()
     runs-on: ubuntu-latest

test/registered/models_e2e/test_dsa_glm52_pd_mtp_cp_layersplit.py

@@ -22,14 +22,16 @@ from sglang.test.server_fixtures.disaggregation_fixture import (
 
 register_cuda_ci(est_time=750, stage="base-c", runner_config="8-gpu-b300")
 
+MODEL_LOADER_EXTRA_CONFIG = '{"enable_multithread_load": true, "num_threads": 6}'
+
 
 class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
-    model = "/data/radixark/model-cache/hub/models--nvidia--GLM-5.2-NVFP4/snapshots/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa"
+    model = "nvidia/GLM-5.2-NVFP4"
 
     # Full GSM8K test set (1319 questions) with a tight accuracy floor.
     gsm8k_accuracy_thres = 0.935
     gsm8k_num_questions = 1319
-    gsm8k_num_threads = 200
+    gsm8k_num_threads = 48
     gsm8k_num_shots = 20
 
     # Prefill worker: interleave prefill-CP + DSA cache layer split on 4 GPUs
@@ -49,6 +51,8 @@ class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
         "interleave",
         "--mem-fraction-static",
         "0.85",
+        "--model-loader-extra-config",
+        MODEL_LOADER_EXTRA_CONFIG,
         "--chunked-prefill-size",
         "4096",
         "--max-prefill-tokens",
@@ -73,6 +77,8 @@ class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
         "fp8_e4m3",
         "--mem-fraction-static",
         "0.85",
+        "--model-loader-extra-config",
+        MODEL_LOADER_EXTRA_CONFIG,
         "--speculative-algorithm",
         "EAGLE",
         "--speculative-num-steps",

test/registered/models_e2e/test_kimi_k3_b300.py

@@ -20,12 +20,10 @@ from sglang.test.test_utils import (
 
 register_cuda_ci(est_time=1200, stage="base-c", runner_config="8-gpu-b300")
 
-MODEL_PATH = (
-    "/data/radixark/model-cache/hub/models--moonshotai--Kimi-K3/"
-    "snapshots/9f62e4e9fffbd0a83ddd60e1c209d828994b3569"
-)
+MODEL_PATH = "moonshotai/Kimi-K3"
 DSPARK_DRAFT_MODEL = "RadixArk/Kimi-K3-DSpark"
 MEGAMOE_URL = "http://0.0.0.0:30000"
+MODEL_LOADER_EXTRA_CONFIG = '{"enable_multithread_load": true, "num_threads": 12}'
 MEGAMOE_ENV = {
     "SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK": "8320",
 }
@@ -44,6 +42,7 @@ class TestKimiK3B300LowLatency(GSM8KMixin, SpecDecodingMixin, CustomTestCase):
 
     gsm8k_score_threshold = 0.95
     gsm8k_num_examples = 200
+    gsm8k_num_threads = 37
     # Gated on GSM8K rather than on test_bs_1_speed below: a 200-question
     # average holds steady when a numerics change moves where the single
     # greedy prompt hits EOS.
@@ -68,7 +67,8 @@ class TestKimiK3B300LowLatency(GSM8KMixin, SpecDecodingMixin, CustomTestCase):
                 "8",
                 "--mem-fraction-static",
                 "0.85",
-                "--weight-loader-prefetch-checkpoints",
+                "--model-loader-extra-config",
+                MODEL_LOADER_EXTRA_CONFIG,
                 "--reasoning-parser",
                 "kimi_k3",
                 "--tool-call-parser",
@@ -95,6 +95,7 @@ class TestKimiK3B300Balanced(GSM8KMixin, CustomTestCase):
 
     gsm8k_score_threshold = 0.95
     gsm8k_num_examples = 200
+    gsm8k_num_threads = 98
 
     @classmethod
     def setUpClass(cls):
@@ -112,7 +113,8 @@ class TestKimiK3B300Balanced(GSM8KMixin, CustomTestCase):
                 "8",
                 "--mem-fraction-static",
                 "0.85",
-                "--weight-loader-prefetch-checkpoints",
+                "--model-loader-extra-config",
+                MODEL_LOADER_EXTRA_CONFIG,
                 "--reasoning-parser",
                 "kimi_k3",
                 "--tool-call-parser",
@@ -133,6 +135,7 @@ class TestKimiK3B300MegaMoE(GSM8KMixin, CustomTestCase):
 
     gsm8k_score_threshold = 0.95
     gsm8k_num_examples = 200
+    gsm8k_num_threads = 22
 
     @classmethod
     def setUpClass(cls):
@@ -154,6 +157,8 @@ class TestKimiK3B300MegaMoE(GSM8KMixin, CustomTestCase):
                 "8",
                 "--mem-fraction-static",
                 "0.85",
+                "--model-loader-extra-config",
+                MODEL_LOADER_EXTRA_CONFIG,
                 "--reasoning-parser",
                 "kimi_k3",
                 "--tool-call-parser",