[AMD] Allow aiter attention backend for Gemma-4 (#38758)

docs/docs/advanced_features/attention_backend.mdx

@@ -477,6 +477,25 @@ Internally, the DSA backend dispatches to different sub-backends for prefill and
   </tbody>
 </table>
 
+#### Gemma 4 on `aiter` (ROCm)
+
+Gemma 4's hybrid attention mixes sliding-window layers (`head_dim` 256) with
+full-attention layers (`head_dim` 512) and shares KV across layers. The `aiter`
+backend supports this and is allow-listed for Gemma 4, so you can serve it with
+`--attention-backend aiter` on MI300X/MI350X. Under the hood `aiter` routes the
+`head_dim`-512 prefill through the Triton `unified_attention` kernel (the CK
+batch-prefill path caps `head_dim` at 256), honors each layer's softmax scale,
+and resolves the sliding-window KV pool for speculative-decode draft workers.
+Non–sliding-window models (`head_dim` ≤ 256) are unaffected and keep using the
+CK path.
+
+<Warning>
+For Gemma 4 on `aiter`, use bf16 KV cache (`--kv-cache-dtype auto`). Do **not**
+use `--kv-cache-dtype fp8_e4m3`: batched decode currently returns incorrect
+output under fp8 KV on this path (a pre-existing aiter fp8-decode issue, tracked
+separately). bf16 KV is correct at every batch size.
+</Warning>
+
 For deployment examples, see the [DeepSeek V3.2 deployment guide](/cookbook/autoregressive/DeepSeek/DeepSeek-V3_2).
 
 ### Hybrid attention (different backends for prefill vs decode) (Experimental)

python/sglang/srt/arg_groups/model_hook.py

@@ -548,12 +548,13 @@ def handle_model_specific_adjustments(server_args: Any):
             "ascend",
             "intel_xpu",
             "intel_amx",
+            "aiter",
         )
         assert (
             prefill_backend in accepted_backends and decode_backend in accepted_backends
         ), (
-            "Gemma4 only supports trtllm_mha, triton, ascend, intel_xpu, or intel_amx "
-            f"attention backend, got prefill={prefill_backend}, decode={decode_backend}"
+            "Gemma4 only supports trtllm_mha, triton, ascend, intel_xpu, intel_amx, or "
+            f"aiter attention backend, got prefill={prefill_backend}, decode={decode_backend}"
         )
 
         # The quantization/moe_runner_backend resolution moved to the override