[ROCm] Guard AITER GDN decode fast path by layout (#42880)

Signed-off-by: Tuukka Sarvi <tuukka.sarvi@amd.com>

vllm/model_executor/layers/mamba/gdn_linear_attn.py

@@ -995,9 +995,9 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
         """ROCm AITER fast path: conv1d + recurrent attention from packed
         qkvz/ba layout.
 
-        For decode-only (no spec, no prefill), dispatches directly to
-        ``_forward_core_decode_fast``.  Otherwise unpacks the packed
-        layout and falls through to ``_forward_core``.
+        For decode-only (no spec, no prefill) interleaved-GQA layouts,
+        dispatches directly to ``_forward_core_decode_fast``. Otherwise unpacks
+        the packed layout and falls through to ``_forward_core``.
 
         Args:
             qkvz: packed [q, k, v, z] projection (num_tokens, qkvz_dim)
@@ -1018,8 +1018,12 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
         attn_metadata = attn_metadata_raw[self.prefix]  # type: ignore[index]
         assert isinstance(attn_metadata, GDNAttentionMetadata)
 
+        # The AITER fused reshape/conv kernel expects Qwen3-Next's interleaved
+        # GQA layout. Qwen3.5 uses a non-interleaved q/k/v/z layout and must use
+        # the generic path below to split/rearrange inputs correctly.
         if (
-            attn_metadata.spec_sequence_masks is None
+            self.gqa_interleaved_layout
+            and attn_metadata.spec_sequence_masks is None
             and attn_metadata.num_prefills == 0
             and attn_metadata.num_decodes > 0
         ):