A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[benchmark] Add triton version in the moe tuned config (#24769)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Sep 16, 2025 at 14:10 UTC
04ad0dc275ca38f2890b7dd21a9fa75311128aa2
3 files changed
+6
-3
benchmarks/kernels/benchmark_moe.py
+1
-1
@@ -560,7 +560,7 @@ def save_configs(
560
filename = os.path.join(save_dir, filename)
561
print(f"Writing best config to {filename}...")
562
with open(filename, "w") as f:
563
- json.dump(configs, f, indent=4)
563
+ json.dump({"triton_version": triton.__version__, **configs}, f, indent=4)
564
f.write("\n")
565
566
vllm/model_executor/layers/fused_moe/fused_moe.py
+4
-1
@@ -720,7 +720,10 @@ def get_moe_configs(
720
logger.info("Using configuration from %s for MoE layer.",
721
config_file_path)
722
# If a configuration has been found, return it
723
- return {int(key): val for key, val in json.load(f).items()}
723
+ tuned_config = json.load(f)
724
+ # Delete triton_version from tuned_config
725
+ tuned_config.pop("triton_version", None)
726
+ return {int(key): val for key, val in tuned_config.items()}
727
728
# If no optimized configuration is available, we will use the default
729
# configuration
vllm/triton_utils/importing.py
+1
-1
@@ -68,7 +68,7 @@ class TritonPlaceholder(types.ModuleType):
68
69
def __init__(self):
70
super().__init__("triton")
71
- self.__version__ = "3.3.0"
71
+ self.__version__ = "3.4.0"
72
self.jit = self._dummy_decorator("jit")
73
self.autotune = self._dummy_decorator("autotune")
74
self.heuristics = self._dummy_decorator("heuristics")