Fix gemlite import

python/sglang/bench_offline_throughput.py

@@ -322,18 +322,6 @@ def throughput_test(
         )
         time.sleep(0.5)
 
-    try:
-        import os
-        import pwd
-
-        from gemlite.core import GemLiteLinearTriton
-
-        GemLiteLinearTriton.cache_config(
-            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
-        )
-    except ImportError:
-        pass
-
     logging.info("\nBenchmark...")
     result = throughput_test_once(
         backend_name=bench_args.backend,

python/sglang/bench_one_batch.py

@@ -63,7 +63,12 @@
 from sglang.srt.sampling.sampling_params import SamplingParams
 from sglang.srt.server import _set_envs_and_config
 from sglang.srt.server_args import PortArgs, ServerArgs
-from sglang.srt.utils import configure_logger, kill_process_tree, suppress_other_loggers
+from sglang.srt.utils import (
+    configure_logger,
+    kill_process_tree,
+    load_gemlite_cache,
+    suppress_other_loggers,
+)
 
 
 @dataclasses.dataclass
@@ -386,18 +391,6 @@ def latency_test(
         server_args.device,
     )
 
-    try:
-        import os
-        import pwd
-
-        from gemlite.core import GemLiteLinearTriton
-
-        GemLiteLinearTriton.cache_config(
-            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
-        )
-    except ImportError:
-        pass
-
     rank_print("Benchmark ...")
 
     # Run the sweep

python/sglang/srt/layers/torchao_utils.py

@@ -2,8 +2,14 @@
 Common utilities for torchao.
 """
 
+import logging
+import os
+import pwd
+
 import torch
 
+logger = logging.getLogger(__name__)
+
 
 def apply_torchao_config_to_model(
     model: torch.nn.Module, torchao_config: str, filter_fn=None
@@ -50,27 +56,17 @@ def filter_fn(module, fqn):
     elif "gemlite" in torchao_config:
         # gemlite-<packing_bitwidth>-<bit_width>-<group_size> or
         # gemlite-<bit_width>-<group_size> (packing_bitwidth defaults to 32)
-        import os
-        import pwd
-
-        import gemlite
-        from gemlite.core import GemLiteLinearTriton, set_autotune
-
-        try:
-            from torchao.quantization import gemlite_uintx_weight_only
-        except:
-            print(
-                f"import `gemlite_uintx_weight_only` failed, please use torchao nightly to use gemlite quantization"
-            )
-            return model
+        from gemlite.core import GemLiteLinearTriton
+        from torchao.quantization import gemlite_uintx_weight_only
 
         _quant_args = torchao_config.split("-")
         bit_width = int(_quant_args[-2])
         group_size = None if _quant_args[-1] == "None" else int(_quant_args[-1])
+
         try:
             packing_bitwidth = int(_quant_args[-3])
-        except:
-            # if only 2 inputs found, use default value
+        except (ValueError, IndexError):
+            # if only 2 inputs found or conversion fails, use default value
             packing_bitwidth = 32
 
         quantize_(

python/sglang/srt/utils.py

@@ -21,6 +21,7 @@
 import logging
 import os
 import pickle
+import pwd
 import random
 import re
 import resource
@@ -1273,3 +1274,14 @@ def dataclass_to_string_truncated(data, max_length=2048):
         )
     else:
         return str(data)
+
+
+def load_gemlite_cache():
+    try:
+        from gemlite.core import GemLiteLinearTriton
+
+        GemLiteLinearTriton.cache_config(
+            f"/tmp/{pwd.getpwuid(os.getuid()).pw_gecos}_gemlite.json"
+        )
+    except ImportError:
+        pass