[AMD] Annotate ATOM source for imported v4 unified attention kernels (#28392)

python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/paged_decode.py

@@ -1,6 +1,9 @@
 # SPDX-License-Identifier: MIT
 # Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
 
+# The following kernel is imported from ATOM.
+# Source: atom/model_ops/v4_kernels/paged_decode.py
+
 """Sparse decode attention over a unified KV pool with per-token paged indices.
 
 Designed for V4 decode + CUDAGraph: replaces the per-fwd `kv_flat_sa`

python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/paged_decode_indices.py

@@ -1,6 +1,9 @@
 # SPDX-License-Identifier: MIT
 # Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
 
+# The following kernel is imported from ATOM.
+# Source: atom/model_ops/v4_kernels/paged_decode_indices.py
+
 """V4 paged-decode index scatter — single Triton kernel writes SWA window-
 prefix paged offsets into the three ragged-packed destination buffers
 (`kv_indices_swa` / `kv_indices_csa` / `kv_indices_hca`).

python/sglang/srt/layers/attention/dsv4/unified_kv_kernels/paged_prefill.py

@@ -1,6 +1,9 @@
 # SPDX-License-Identifier: MIT
 # Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
 
+# The following kernel is imported from ATOM.
+# Source: atom/model_ops/v4_kernels/paged_prefill.py
+
 """Sparse prefill attention with two KV sources: paged `unified_kv` (history)
 and per-fwd flat `kv` (current chunk's input).
 

python/sglang/srt/mem_cache/deepseek_v4_memory_pool.py

@@ -375,6 +375,7 @@ class DeepSeekV4LayerItem(NamedTuple):
     compress_kv_pool: Optional[DeepSeekV4SingleKVPool] = None
 
 
+# The following kv pool follows ATOM's unified_kv kernel layout.
 class DeepSeekV4UnifiedKVPool:
     """
     Layout: