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

[Refactor] Rename n_share_experts_fusion as num_fused_shared_experts (#6735)

Browse Source
This commit is contained in:
Cheng Wan
2025-06-03 17:48:24 -07:00
committed by GitHub
parent b6d0ce9f78
commit 8a5480528d
14 changed files with 82 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
benchmark/kernels/fused_moe_triton/README.md
View File

@@ -27,19 +27,17 @@ python benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py \
--dtype fp8_w8a8 \
--tune
# Tune DeepSeek-V3 with FP8, TP=8 and n_share_experts_fusion=8
# Tune DeepSeek-V3 with FP8 and TP=8
python benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py \
--model deepseek-ai/DeepSeek-V3-0324 \
--tp-size 8 \
--n-share-experts-fusion 8 \
--dtype fp8_w8a8 \
--tune
# Tune DeepSeek-R1 with channel-wise INT8, TP=16 and n_share_experts_fusion=16
# Tune DeepSeek-R1 with channel-wise INT8 and TP=16
python benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py \
--model meituan/DeepSeek-R1-Channel-INT8 \
--tp-size 16 \
--n-share-experts-fusion 16 \
--dtype int8_w8a8 \
--tune
```
@@ -65,11 +63,10 @@ python benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_tri
--model deepseek-ai/DeepSeek-V3-0324 \
--tp-size 8
# Compare with custom TP size and n_share_experts_fusion
# Compare with custom TP size
python benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py \
--model deepseek-ai/DeepSeek-V3-0324 \
--tp-size 8 \
--n-share-experts-fusion 8
--tp-size 8
```
The benchmark results will be saved as plots and data files in the specified output directory (default: `./configs/benchmark_ops/vllm_sglang_fused_moe/`).

10
benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py
View File

@@ -18,7 +18,7 @@ from sglang.srt.layers.moe.fused_moe_triton.fused_moe import (
)
def get_model_config(model_name: str, tp_size: int, n_share_experts_fusion: int = 0):
def get_model_config(model_name: str, tp_size: int):
"""Get model configuration parameters"""
config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
@@ -43,9 +43,8 @@ def get_model_config(model_name: str, tp_size: int, n_share_experts_fusion: int
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // tp_size
elif config.architectures[0] in ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]:
n_share_fusion_experts = n_share_experts_fusion
E = (
config.n_routed_experts + n_share_fusion_experts
config.n_routed_experts + 1
if config.architectures[0] in ["DeepseekV3ForCausalLM"]
else config.n_routed_experts
)
@@ -294,7 +293,6 @@ def main():
"--model", type=str, default="mistralai/Mixtral-8x7B-Instruct-v0.1"
)
parser.add_argument("--tp-size", type=int, default=2)
parser.add_argument("--n-share-experts-fusion", type=int, default=0)
parser.add_argument("--use-fp8-w8a8", action="store_true")
parser.add_argument(
"--save-path",
@@ -325,9 +323,7 @@ def main():
pipeline_model_parallel_size=1,
)
model_config = get_model_config(
args.model, args.tp_size, args.n_share_experts_fusion
)
model_config = get_model_config(args.model, args.tp_size)
benchmark.run(
show_plots=True,
print_data=True,

12
benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py
View File

@@ -399,9 +399,8 @@ def main(args: argparse.Namespace):
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
elif config.architectures[0] in ["DeepseekV2ForCausalLM", "DeepseekV3ForCausalLM"]:
n_share_fusion_experts = args.n_share_experts_fusion
E = (
config.n_routed_experts + n_share_fusion_experts
config.n_routed_experts + 1
if config.architectures[0] in ["DeepseekV3ForCausalLM"]
else config.n_routed_experts
)
@@ -409,8 +408,7 @@ def main(args: argparse.Namespace):
intermediate_size = config.moe_intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
elif config.architectures[0] == "Llama4ForConditionalGeneration":
n_share_fusion_experts = args.n_share_experts_fusion
E = config.text_config.num_local_experts + n_share_fusion_experts
E = config.text_config.num_local_experts + 1
topk = config.text_config.num_experts_per_tok
intermediate_size = config.text_config.intermediate_size
shard_intermediate_size = 2 * intermediate_size // args.tp_size
@@ -570,12 +568,6 @@ if __name__ == "__main__":
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--batch-size", type=int, required=False)
parser.add_argument("--tune", action="store_true")
parser.add_argument(
"--n-share-experts-fusion",
type=int,
default=0,
help="The number of shared_experts need to be replica to fuse with normal experts in deepseek v3/r1",
)
args = parser.parse_args()
main(args)