MoE FP8 block scale ds routing topk8 ng1 kg1 e48 h7168 i2048
0 eligible runs
moe
FP8 block-scale MoE (DeepSeek-style). Kimi K2 (EP=8). DeepSeek routing, n_group=1, topk_group=1. 48 local experts per device (384 total / 8 EP).
Current records
No published measurement for the selected workload.
Implementations
Implementation
Runtime
Best latency
Evidence
Availability
Semantics
Inputs and outputs
routing_logitsfp32 [seq_len, num_experts]
routing_biasbf16 [num_experts]
hidden_statesfp8_e4m3 [seq_len, hidden_size]
hidden_states_scalefp32 [num_hidden_blocks, seq_len]
gemm1_weightsfp8_e4m3 [num_local_experts, gemm1_out_size, hidden_size]
gemm1_weights_scalefp32 [num_local_experts, num_gemm1_out_blocks, num_hidden_blocks]
gemm2_weightsfp8_e4m3 [num_local_experts, hidden_size, intermediate_size]
gemm2_weights_scalefp32 [num_local_experts, num_hidden_blocks, num_intermediate_blocks]
local_expert_offsetint32 scalar
routed_scaling_factorfp32 scalar
outputbf16 [seq_len, hidden_size]
Axes and behavior
top_kconstant = 8
seq_lenvariable
hidden_sizeconstant = 7168
num_expertsconstant = 384
gemm1_out_sizeconstant = 4096
intermediate_sizeconstant = 2048
num_hidden_blocksconstant = 56
num_local_expertsconstant = 48
num_gemm1_out_blocksconstant = 32
num_intermediate_blocksconstant = 16
determinismunspecified
constraintsNo mutation or aliasing
Identity
No source imports for this operation yet.How records are decidedJSON