MOE\_CFG\_DIR="\$(python \-c 'import vllm, pathlib; print(pathlib.Path(vllm.**file**).parent / "model\_executor/layers/fused\_moe/configs")')" cat \<\< 'JSON' \> "\${MOE\_CFG\_DIR}/E=256,N=512,device\_name=AMD\_Radeon\_AI\_PRO\_R9700,dtype=int4\_w4a16.json" { "1": {"BLOCK\_SIZE\_M": 16, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 1, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "2": {"BLOCK\_SIZE\_M": 16, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 1, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "4": {"BLOCK\_SIZE\_M": 16, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 1, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "8": {"BLOCK\_SIZE\_M": 16, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 1, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "16": {"BLOCK\_SIZE\_M": 32, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 4, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "32": {"BLOCK\_SIZE\_M": 32, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 4, "SPLIT\_K": 1, "num\_warps": 4, "num\_stages": 2}, "64": {"BLOCK\_SIZE\_M": 64, "BLOCK\_SIZE\_N": 64, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 8, "SPLIT\_K": 1, "num\_warps": 8, "num\_stages": 2}, "128": {"BLOCK\_SIZE\_M": 64, "BLOCK\_SIZE\_N": 128, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 8, "SPLIT\_K": 1, "num\_warps": 8, "num\_stages": 2}, "256": {"BLOCK\_SIZE\_M": 64, "BLOCK\_SIZE\_N": 128, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 8, "SPLIT\_K": 1, "num\_warps": 8, "num\_stages": 2}, "512": {"BLOCK\_SIZE\_M": 64, "BLOCK\_SIZE\_N": 128, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 8, "SPLIT\_K": 1, "num\_warps": 8, "num\_stages": 2}, "1024": {"BLOCK\_SIZE\_M": 64, "BLOCK\_SIZE\_N": 128, "BLOCK\_SIZE\_K": 64, "GROUP\_SIZE\_M": 8, "SPLIT\_K": 1, "num\_warps": 8, "num\_stages": 2} } JSON