diff --git a/tests/kernels/moe/test_deepgemm.py b/tests/kernels/moe/test_deepgemm.py index 298fb23f1d..ae6c9685a4 100644 --- a/tests/kernels/moe/test_deepgemm.py +++ b/tests/kernels/moe/test_deepgemm.py @@ -381,6 +381,7 @@ def run_single_fp4_case(m, n, k, topk, num_experts): FP4_MNKs = [ (128, 4096, 4096), # DeepSeek V4 shape (256, 2048, 2048), # Half-size variant + (128, 384, 3584), # Kimi-K3 TP8 latent-MoE shape ] FP4_TOPKS = [2] diff --git a/vllm/model_executor/layers/fused_moe/deep_gemm_utils.py b/vllm/model_executor/layers/fused_moe/deep_gemm_utils.py index 27659d220d..b25d3e93ae 100644 --- a/vllm/model_executor/layers/fused_moe/deep_gemm_utils.py +++ b/vllm/model_executor/layers/fused_moe/deep_gemm_utils.py @@ -5,6 +5,8 @@ Taken from https://github.com/ModelTC/LightLLM/blob/8ed97c74c18f11505b048b1ba00b and updated to fit vllm needs and terminology. """ +import math + import torch import vllm.model_executor.layers.fused_moe.modular_kernel as mk @@ -424,7 +426,7 @@ def ep_gather( num_warps = 2 num_tokens = output_tensor.shape[0] hidden_size = input_tensor.shape[1] - BLOCK_D = min(hidden_size, 1024) + BLOCK_D = math.gcd(hidden_size, 1024) assert hidden_size % BLOCK_D == 0 grid = (triton.cdiv(hidden_size, BLOCK_D), min(num_tokens, 1024))