From 8e1f97e70984192cc63c51a8af3e313cab8c5c73 Mon Sep 17 00:00:00 2001 From: Canlin Guo Date: Tue, 8 Sep 2026 22:45:18 +0800 Subject: [PATCH] [Qwen3.8-Flash-Next] Tune FP8 TP2/TP4 Triton MoE on B200 (#55890) Signed-off-by: Canlin Co-authored-by: OpenAI Codex --- ...00,dtype=fp8_w8a8,block_shape=[32,32].json | 115 ++++++++++++++++++ ...00,dtype=fp8_w8a8,block_shape=[64,64].json | 115 ++++++++++++++++++ 2 files changed, 230 insertions(+) create mode 100644 vllm/model_executor/layers/fused_moe/configs/E=512,N=160,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[32,32].json create mode 100644 vllm/model_executor/layers/fused_moe/configs/E=512,N=320,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[64,64].json diff --git a/vllm/model_executor/layers/fused_moe/configs/E=512,N=160,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[32,32].json b/vllm/model_executor/layers/fused_moe/configs/E=512,N=160,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[32,32].json new file mode 100644 index 0000000000..8cbe4026cc --- /dev/null +++ b/vllm/model_executor/layers/fused_moe/configs/E=512,N=160,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[32,32].json @@ -0,0 +1,115 @@ +{ + "triton_version": "3.7.1", + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 6 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 4 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 4 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 4 + }, + "8192": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 32, + "GROUP_SIZE_M": 8, + "num_warps": 2, + "num_stages": 4 + } +} diff --git a/vllm/model_executor/layers/fused_moe/configs/E=512,N=320,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[64,64].json b/vllm/model_executor/layers/fused_moe/configs/E=512,N=320,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[64,64].json new file mode 100644 index 0000000000..a65fa34b98 --- /dev/null +++ b/vllm/model_executor/layers/fused_moe/configs/E=512,N=320,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[64,64].json @@ -0,0 +1,115 @@ +{ + "triton_version": "3.7.1", + "1": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "2": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "4": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 32, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "8": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 64, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 5 + }, + "16": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 3 + }, + "32": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 4 + }, + "64": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "128": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "256": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 4, + "num_stages": 5 + }, + "512": { + "BLOCK_SIZE_M": 16, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 1, + "num_warps": 2, + "num_stages": 3 + }, + "1024": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 8, + "num_warps": 4, + "num_stages": 5 + }, + "2048": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "4096": { + "BLOCK_SIZE_M": 32, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 5 + }, + "8192": { + "BLOCK_SIZE_M": 64, + "BLOCK_SIZE_N": 128, + "BLOCK_SIZE_K": 64, + "GROUP_SIZE_M": 32, + "num_warps": 4, + "num_stages": 6 + } +}