[AMD] Enable FP4 indexer for Deepseek V4

docker/rocm.Dockerfile

@@ -569,10 +569,13 @@ RUN pip uninstall -y aiter
 # produced by a fresh `git clone` above, so there are no real user changes to
 # preserve.
 # cherry pick 8578af1 commit for v4 fp4 indexer kv-cache fix, may be removed in next aiter upgrade
+# apply fix for v4 fp4 indexer, may be removed in next aiter upgrade
 RUN git clone ${AITER_REPO} \
  && cd aiter \
  && git checkout -f ${AITER_COMMIT} \
  && git cherry-pick --no-commit 8578af153f4fa1e007fede7e3c1e1b373f07af4c \
+ && sed -i 's/from functools import lru_cache/from functools import cache/' aiter/ops/flydsl/kernels/mqa_logits/pa_mqa_logits_fp4_prefill.py \
+ && sed -i 's/@lru_cache(maxsize=32)/@cache/' aiter/ops/flydsl/kernels/mqa_logits/pa_mqa_logits_fp4_prefill.py \
  && git submodule update --init --recursive \
  && pip install -r requirements.txt \
  && if [ "${GPU_ARCH_LIST}" = "gfx1250" ]; then \

python/sglang/kernels/ops/attention/dsv4/compress.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import TYPE_CHECKING, Literal, NamedTuple, Optional, Union
+from typing import TYPE_CHECKING, Literal, NamedTuple, Optional, Union, cast
 
 import torch
 
@@ -430,9 +430,33 @@ def compress_norm_rope_store(
     page_size: int,
     use_fp4: bool = False,
     bf16_store: bool = False,
+    # HIP FP4 uses split scale storage and precomputed BF16 RoPE tables.
+    kvcache_scale: Optional[torch.Tensor] = None,
+    rope_cache: Optional[tuple[torch.Tensor, torch.Tensor]] = None,
+    fp4_k_write_metadata=None,
 ) -> None:
     if use_fp4:
         assert kv.shape[-1] == 128
+    if is_hip() and use_fp4:
+        from sglang.kernels.ops.attention.dsv4.fp4_indexer_hip import (
+            aiter_k_indexer_fp4_cache_write,
+        )
+
+        cos, sin = cast(tuple[torch.Tensor, torch.Tensor], rope_cache)
+        aiter_k_indexer_fp4_cache_write(
+            k=kv,
+            norm_weight=norm_weight,
+            norm_epsilon=norm_eps,
+            cos=cos,
+            sin=sin,
+            plan=plan,
+            out_loc=out_loc,
+            k_payload=kvcache,
+            k_scale=cast(torch.Tensor, kvcache_scale),
+            write_metadata=fp4_k_write_metadata,
+        )
+        return
+
     freq_cis = torch.view_as_real(freq_cis).flatten(-2)
     if _is_xpu:
         compress_norm_rope_store_xpu(