[Feature, Hardware] Enable DeepseekV3 on AMD GPUs (#2601)

Co-authored-by: root <root@banff-cyxtera-s83-5.amd.com>
Co-authored-by: HAI <hixiao@gmail.com>
Co-authored-by: Bruce Xue <yigex@xilinx.com>
Co-authored-by: Yineng Zhang <me@zhyncs.com>

python/sglang/srt/layers/attention/triton_ops/decode_attention.py

@@ -406,6 +406,10 @@ def _decode_grouped_att_m_fwd(
     Lk = k_buffer.shape[-1]
     Lv = v_buffer.shape[-1]
 
+    # [TODO] work around shmem limit on MI3xx
+    if is_hip_ and Lk >= 576:
+        BLOCK = 16
+
     if Lk == 576:
         BLOCK_DMODEL = 512
         BLOCK_DPE = 64