[Kernel][Perf] Tune H20 block-FP8 MoE for low-batch decode

Retune the E=256, N=256 H20-3e block-FP8 configuration used by GLM-5.3 TP8 and DeepSeek models. Narrower N tiles improve SM utilization for M=1 through M=8 while leaving larger batches unchanged.

AI assistance was used for analysis, benchmarking, and edits.

Co-authored-by: Codex <noreply@openai.com>

Signed-off-by: liuyao0322 <yaoliu548926@gmail.com>

Author: liuyao0322

vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json

@@ -1,35 +1,35 @@
 {
     "1": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
-        "num_stages": 3
+        "num_stages": 4
     },
     "2": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
         "num_stages": 4
     },
     "4": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
         "num_stages": 3
     },
     "8": {
         "BLOCK_SIZE_M": 16,
-        "BLOCK_SIZE_N": 128,
+        "BLOCK_SIZE_N": 64,
         "BLOCK_SIZE_K": 128,
         "GROUP_SIZE_M": 1,
         "num_warps": 4,
-        "num_stages": 4
+        "num_stages": 3
     },
     "16": {
         "BLOCK_SIZE_M": 16,