AMD/ROCm: update AITER repo to ROCm/aiter (#3747)

docker/Dockerfile.rocm

@@ -17,8 +17,8 @@ ARG TRITON_REPO="https://github.com/ROCm/triton.git"
 ARG TRITON_COMMIT="improve_fa_decode_3.0.0"
 
 
-ARG ATER_REPO="https://github.com/HaiShaw/ater"
-ARG CK_COMMITS="fa05ae"
+ARG AITER_REPO="https://github.com/ROCm/aiter.git"
+ARG AITER_COMMIT="dev/testx"
 
 RUN git clone ${SGL_REPO} \
     && cd sglang \
@@ -53,8 +53,9 @@ RUN git clone ${TRITON_REPO} \
     && cd python \
     && python3 setup.py install
 
-RUN git clone ${ATER_REPO} \
-    && cd ater \
+RUN git clone ${AITER_REPO} \
+    && cd aiter \
+    && git checkout ${AITER_COMMIT} \
     && git submodule update --init --recursive \
     && PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py develop
 

python/sglang/srt/layers/moe/fused_moe_triton/layer.py

@@ -167,8 +167,8 @@ class UnquantizedFusedMoEMethod(FusedMoEMethodBase, CustomOp):
         )
 
         if is_hip_ and get_bool_env_var("CK_MOE"):
-            import ater
-            from ater.fused_moe import fused_experts_ck
+            import aiter
+            from aiter.fused_moe import fused_experts_ck
 
             assert activation == "silu", f"{activation=} is not supported."
 

python/sglang/srt/layers/quantization/fp8.py

@@ -789,8 +789,8 @@ class Fp8MoEMethod:
         )
 
         if is_hip_ and get_bool_env_var("CK_MOE"):
-            import ater
-            from ater.fused_moe import fused_experts_ck
+            import aiter
+            from aiter.fused_moe import fused_experts_ck
 
             assert activation == "silu", f"{activation=} is not supported."