add configs.

wangzaijun · wangzaijun · commit d02b1f1fc654 · 2024-12-26T17:46:00.000+08:00
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=1024,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=1024,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 64, "num_warps": 4}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 8}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "1024": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 1}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=128,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=128,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 64, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "256": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 8}, "512": {"BLOCK_M": 4, "BLOCK_N": 64, "num_warps": 2}, "1024": {"BLOCK_M": 8, "BLOCK_N": 128, "num_warps": 8}, "2048": {"BLOCK_M": 32, "BLOCK_N": 64, "num_warps": 16}, "4096": {"BLOCK_M": 16, "BLOCK_N": 64, "num_warps": 4}, "8192": {"BLOCK_M": 8, "BLOCK_N": 128, "num_warps": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=1408,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=1408,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 64, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "1024": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "4096": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "8192": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=192,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=192,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 2}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "512": {"BLOCK_M": 16, "BLOCK_N": 64, "num_warps": 8}, "1024": {"BLOCK_M": 16, "BLOCK_N": 128, "num_warps": 16}, "2048": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "4096": {"BLOCK_M": 8, "BLOCK_N": 64, "num_warps": 4}, "8192": {"BLOCK_M": 8, "BLOCK_N": 64, "num_warps": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=2048,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=2048,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "1024": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=256,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=256,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "8": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 8}, "64": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "512": {"BLOCK_M": 8, "BLOCK_N": 128, "num_warps": 8}, "1024": {"BLOCK_M": 16, "BLOCK_N": 64, "num_warps": 4}, "2048": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "4096": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "8192": {"BLOCK_M": 4, "BLOCK_N": 256, "num_warps": 8}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=4096,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=4096,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 2}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 1}, "1024": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=512,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=512,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "8": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "64": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 16}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "512": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 2}, "1024": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "2048": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 2}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}}
diff --git a/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=8192,out_dtype=torch.bfloat16}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_silu_and_mul_kernel/{N=8192,out_dtype=torch.bfloat16}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 16}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "1024": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=1024,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=1024,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 2, "NUM_STAGE": 3}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 4, "NUM_STAGE": 2}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 5}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 5}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 1}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 2}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 1}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=1024,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=1024,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 2}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 3}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 8, "NUM_STAGE": 1}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 8, "NUM_STAGE": 1}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 1}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 1}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=2048,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=2048,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 1}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 1, "NUM_STAGE": 4}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 1}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 5}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 3}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 2}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 5}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=2048,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=2048,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 2, "NUM_STAGE": 2}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 4, "NUM_STAGE": 4}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 5}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 1}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=4096,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=4096,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 1, "NUM_STAGE": 4}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 2}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 3}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "2048": {"BLOCK_M": 2, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=4096,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=4096,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 1, "NUM_STAGE": 1}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 2, "NUM_STAGE": 2}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 3}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 4}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 4}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 1}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=5120,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=5120,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 5}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 3}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 4}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=5120,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=5120,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 2, "NUM_STAGE": 1}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 8, "NUM_STAGE": 1}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 3}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 3}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 3}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 3}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=8192,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=8192,out_dtype=torch.bfloat16,topk_num=1}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 1}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 2}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 2}, "1024": {"BLOCK_M": 2, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 4}, "2048": {"BLOCK_M": 2, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 3}, "4096": {"BLOCK_M": 2, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "8192": {"BLOCK_M": 2, "BLOCK_DIM": 512, "num_warps": 4, "NUM_STAGE": 1}}
diff --git a/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=8192,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json b/lightllm/common/all_kernel_configs/moe_sum_reduce_kernel/{hidden_dim=8192,out_dtype=torch.bfloat16,topk_num=6}_NVIDIA_H800.json
@@ -0,0 +1 @@
+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 8, "NUM_STAGE": 3}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 8, "NUM_STAGE": 5}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 3}, "512": {"BLOCK_M": 4, "BLOCK_DIM": 512, "num_warps": 2, "NUM_STAGE": 4}, "1024": {"BLOCK_M": 4, "BLOCK_DIM": 1024, "num_warps": 4, "NUM_STAGE": 1}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 256, "num_warps": 1, "NUM_STAGE": 4}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 1}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 1}}

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	+{"1": {"BLOCK_M": 1, "BLOCK_N": 64, "num_warps": 4}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 8}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "1024": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 1}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	+{"1": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 2}, "8": {"BLOCK_M": 1, "BLOCK_N": 128, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "512": {"BLOCK_M": 16, "BLOCK_N": 64, "num_warps": 8}, "1024": {"BLOCK_M": 16, "BLOCK_N": 128, "num_warps": 16}, "2048": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "4096": {"BLOCK_M": 8, "BLOCK_N": 64, "num_warps": 4}, "8192": {"BLOCK_M": 8, "BLOCK_N": 64, "num_warps": 4}}
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	+{"1": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "8": {"BLOCK_M": 1, "BLOCK_N": 256, "num_warps": 4}, "64": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "128": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 8}, "256": {"BLOCK_M": 1, "BLOCK_N": 512, "num_warps": 4}, "512": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "1024": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "2048": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 2}, "4096": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}, "8192": {"BLOCK_M": 1, "BLOCK_N": 1024, "num_warps": 8}}
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	+{"1": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 2, "NUM_STAGE": 3}, "8": {"BLOCK_M": 1, "BLOCK_DIM": 64, "num_warps": 4, "NUM_STAGE": 2}, "64": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 5}, "128": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 8, "NUM_STAGE": 5}, "256": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 2, "NUM_STAGE": 1}, "512": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "1024": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 5}, "2048": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 2}, "4096": {"BLOCK_M": 1, "BLOCK_DIM": 1024, "num_warps": 1, "NUM_STAGE": 1}, "8192": {"BLOCK_M": 1, "BLOCK_DIM": 512, "num_warps": 1, "NUM_STAGE": 1}}