Sync changes on io_struct.py and deterministic ops (#11498)

python/sglang/test/test_cutlass_moe.py

@@ -34,7 +34,7 @@ def get_model_config(tp_size: int):
         "topk": topk,
         "hidden_size": config.hidden_size,
         "shard_intermediate_size": shard_intermediate_size,
-        "dtype": config.torch_dtype,
+        "dtype": config.dtype,
         "block_shape": config.quantization_config["weight_block_size"],
     }
 

python/sglang/test/test_deterministic_utils.py

@@ -1,8 +1,5 @@
-import time
 import unittest
 
-import requests
-
 from sglang.srt.utils import kill_process_tree
 from sglang.test.test_deterministic import BenchArgs, test_deterministic
 from sglang.test.test_utils import (
@@ -55,6 +52,7 @@ class TestDeterministicBase(CustomTestCase):
         args.n_start = 10
         args.n_trials = 20
         results = test_deterministic(args)
+        args.temperature = 0.5  # test for deterministic sampling
         for result in results:
             assert result == 1
 
@@ -65,6 +63,7 @@ class TestDeterministicBase(CustomTestCase):
         args.test_mode = "mixed"
         args.n_start = 10
         args.n_trials = 20
+        args.temperature = 0.5  # test for deterministic sampling
         results = test_deterministic(args)
         for result in results:
             assert result == 1
@@ -76,6 +75,7 @@ class TestDeterministicBase(CustomTestCase):
         args.test_mode = "prefix"
         args.n_start = 10
         args.n_trials = 10
+        args.temperature = 0.5  # test for deterministic sampling
         results = test_deterministic(args)
         for result in results:
             assert result == 1