From 9a23c484562770e6acb721af30d41f201d9b9936 Mon Sep 17 00:00:00 2001 From: Xiaoyu Zhang <35585791+BBuf@users.noreply.github.com> Date: Thu, 26 Dec 2024 19:13:31 +0800 Subject: [PATCH] h100 tuning fused_moe_triton for qwen2 moe (#2560) --- .../benchmark_torch_compile_fused_moe.py | 7 +- ...nchmark_vllm_vs_sglang_fused_moe_triton.py | 7 +- .../tuning_fused_moe_triton.py | 7 +- ...=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json | 146 ++++++++++++++++++ ...280,device_name=NVIDIA_H100_80GB_HBM3.json | 86 +++++------ ...=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json | 146 ++++++++++++++++++ ...=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json | 146 ++++++++++++++++++ ...320,device_name=NVIDIA_H100_80GB_HBM3.json | 146 ++++++++++++++++++ ...=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json | 146 ++++++++++++++++++ ...640,device_name=NVIDIA_H100_80GB_HBM3.json | 42 ++--- 10 files changed, 812 insertions(+), 67 deletions(-) create mode 100644 python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json create mode 100644 python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json create mode 100644 python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json create mode 100644 python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json create mode 100644 python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json diff --git a/benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py b/benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py index 10e258f16..b81a22800 100644 --- a/benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py +++ b/benchmark/kernels/fused_moe_triton/benchmark_torch_compile_fused_moe.py @@ -13,7 +13,7 @@ from sglang.srt.model_executor.cuda_graph_runner import set_torch_compile_config def get_model_config(model_name: str, tp_size: int): """Get model configuration parameters""" - config = AutoConfig.from_pretrained(model_name) + config = AutoConfig.from_pretrained(model_name, trust_remote_code=True) if config.architectures[0] == "DbrxForCausalLM": E = config.ffn_config.moe_num_experts @@ -30,6 +30,11 @@ 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] == "DeepseekV2ForCausalLM": + 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 else: # Default: Mixtral E = config.num_local_experts diff --git a/benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py b/benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py index 1d16b2568..faf5c6b4e 100644 --- a/benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py +++ b/benchmark/kernels/fused_moe_triton/benchmark_vllm_vs_sglang_fused_moe_triton.py @@ -12,7 +12,7 @@ from sglang.srt.layers.moe.fused_moe_triton.fused_moe import ( def get_model_config(model_name: str, tp_size: int): """Get model configuration parameters""" - config = AutoConfig.from_pretrained(model_name) + config = AutoConfig.from_pretrained(model_name, trust_remote_code=True) if config.architectures[0] == "DbrxForCausalLM": E = config.ffn_config.moe_num_experts @@ -29,6 +29,11 @@ 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] == "DeepseekV2ForCausalLM": + 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 else: # Default: Mixtral E = config.num_local_experts diff --git a/benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py b/benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py index da6af2fb8..5e7f7ee1b 100644 --- a/benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py +++ b/benchmark/kernels/fused_moe_triton/tuning_fused_moe_triton.py @@ -307,7 +307,7 @@ def save_configs( def main(args: argparse.Namespace): print(args) - config = AutoConfig.from_pretrained(args.model) + config = AutoConfig.from_pretrained(args.model, trust_remote_code=True) if config.architectures[0] == "DbrxForCausalLM": E = config.ffn_config.moe_num_experts topk = config.ffn_config.moe_top_k @@ -323,6 +323,11 @@ 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] == "DeepseekV2ForCausalLM": + 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 else: # Default: Mixtral E = config.num_local_experts diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 000000000..e2f8164cc --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json index d4c9ddd12..ef6a0479c 100644 --- a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json @@ -3,19 +3,43 @@ "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 16, - "num_warps": 4, + "GROUP_SIZE_M": 32, + "num_warps": 8, "num_stages": 3 }, "2": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 16, + "GROUP_SIZE_M": 64, "num_warps": 4, - "num_stages": 4 + "num_stages": 5 }, "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "24": { "BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, @@ -23,33 +47,9 @@ "num_warps": 4, "num_stages": 3 }, - "8": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, - "BLOCK_SIZE_K": 256, - "GROUP_SIZE_M": 1, - "num_warps": 4, - "num_stages": 2 - }, - "16": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, - "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 16, - "num_warps": 4, - "num_stages": 5 - }, - "24": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 256, - "GROUP_SIZE_M": 32, - "num_warps": 4, - "num_stages": 2 - }, "32": { - "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, @@ -57,32 +57,32 @@ }, "48": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 256, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 1, "num_warps": 4, - "num_stages": 3 + "num_stages": 2 }, "64": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, - "num_stages": 3 + "num_stages": 5 }, "96": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, - "num_stages": 3 + "num_stages": 2 }, "128": { - "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 128, - "BLOCK_SIZE_K": 128, + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 1, "num_warps": 4, "num_stages": 3 @@ -115,7 +115,7 @@ "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 16, "num_warps": 8, "num_stages": 4 }, @@ -131,7 +131,7 @@ "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 32, + "GROUP_SIZE_M": 1, "num_warps": 8, "num_stages": 4 }, @@ -139,7 +139,7 @@ "BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 32, "num_warps": 8, "num_stages": 4 } diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 000000000..14c6f8c1a --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 64, + "num_warps": 8, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 256, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 000000000..869f75127 --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 2 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + } +} diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json new file mode 100644 index 000000000..40d85e43e --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 3 + }, + "16": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 64, + "num_warps": 4, + "num_stages": 3 + }, + "24": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 2 + }, + "128": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json new file mode 100644 index 000000000..96437b5ba --- /dev/null +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json @@ -0,0 +1,146 @@ +{ + "1": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "2": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "4": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "8": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "16": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "24": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "48": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "64": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "96": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 + }, + "128": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 3 + }, + "256": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 3 + }, + "512": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 4, + "num_stages": 4 + }, + "1024": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "1536": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "3072": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 16, + "num_warps": 8, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 128, + "BLOCK_SIZE_N": 256, + "BLOCK_SIZE_K": 128, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 4 + } +} diff --git a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json index b8d3be231..6e6050694 100644 --- a/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json +++ b/python/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json @@ -1,11 +1,11 @@ { "1": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 8, - "num_stages": 4 + "num_warps": 4, + "num_stages": 3 }, "2": { "BLOCK_SIZE_M": 16, @@ -13,15 +13,15 @@ "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 16, "num_warps": 4, - "num_stages": 4 + "num_stages": 3 }, "4": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, - "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 64, - "num_warps": 4, - "num_stages": 4 + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 8, + "num_stages": 3 }, "8": { "BLOCK_SIZE_M": 32, @@ -41,11 +41,11 @@ }, "24": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 64, - "num_warps": 4, - "num_stages": 3 + "GROUP_SIZE_M": 32, + "num_warps": 8, + "num_stages": 4 }, "32": { "BLOCK_SIZE_M": 16, @@ -60,28 +60,28 @@ "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, + "num_warps": 8, "num_stages": 2 }, "64": { "BLOCK_SIZE_M": 16, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, - "num_warps": 4, - "num_stages": 3 + "num_warps": 8, + "num_stages": 2 }, "96": { "BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 16, "num_warps": 4, "num_stages": 2 }, "128": { "BLOCK_SIZE_M": 32, - "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 1, "num_warps": 4, @@ -99,9 +99,9 @@ "BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, - "GROUP_SIZE_M": 1, + "GROUP_SIZE_M": 16, "num_warps": 4, - "num_stages": 3 + "num_stages": 4 }, "1024": { "BLOCK_SIZE_M": 128,