EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts (#3988)

Browse Source 
Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <hanming_lu@berkeley.edu>
This commit is contained in:
Lianmin Zheng
2025-03-03 00:12:04 -08:00
parent 0194948fd9
commit ac2387279e
86 changed files with 4116 additions and 2015 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py
View File

@@ -30,11 +30,20 @@ def get_model_config(model_name: str, tp_size: int):
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // tp_size
elif config.architectures[0] in [
"Grok1ForCausalLM",
"Grok1ImgGen",
"Grok1AForCausalLM",
]:
E = config.num_local_experts
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // tp_size
elif config.architectures[0] in ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]:
E = config.n_routed_experts
topk = config.num_experts_per_tok
intermediate_size = config.intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
shard_intermediate_size = 2 * intermediate_size // tp_size
else:
# Default: Mixtral
E = config.num_local_experts

9
benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py
View File

@@ -35,6 +35,15 @@ def get_model_config(model_name: str, tp_size: int):
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // tp_size
elif config.architectures[0] in [
"Grok1ForCausalLM",
"Grok1ImgGen",
"Grok1AForCausalLM",
]:
E = config.num_local_experts
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // tp_size
else:
# Default: Mixtral
E = config.num_local_experts

9
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
View File

@@ -397,6 +397,15 @@ def main(args: argparse.Namespace):
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
elif config.architectures[0] in [
"Grok1ForCausalLM",
"Grok1ImgGen",
"Grok1AForCausalLM",
]:
E = config.num_local_experts
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
else:
# Default: Mixtral
E = config.num_local_experts