[AMD] Fix accuracy issue when running TP4 dsv3 model with mtp (#18607)

Co-authored-by: YC Tseng <yctseng@amd.com>
Co-authored-by: kkHuang-amd <wunhuang@amd.com>

docker/rocm.Dockerfile

@@ -21,7 +21,7 @@ ENV BUILD_TRITON="0"
 ENV BUILD_LLVM="0"
 ENV BUILD_AITER_ALL="1"
 ENV BUILD_MOONCAKE="1"
-ENV AITER_COMMIT="v0.1.10.post2"
+ENV AITER_COMMIT="v0.1.10.post3"
 
 # ===============================
 # Base image 950 and args
@@ -31,7 +31,7 @@ ENV BUILD_TRITON="0"
 ENV BUILD_LLVM="0"
 ENV BUILD_AITER_ALL="1"
 ENV BUILD_MOONCAKE="1"
-ENV AITER_COMMIT="v0.1.10.post2"
+ENV AITER_COMMIT="v0.1.10.post3"
 # ===============================
 # Chosen arch and args
 FROM ${GPU_ARCH}

python/sglang/srt/layers/attention/aiter_backend.py

@@ -195,11 +195,15 @@ class AiterAttnBackend(AttentionBackend):
             )
             global _use_mla_ps_kernel, fast_mode, intra_batch_mode
 
+            if self.num_head == 32:
+                fast_mode = True
+                intra_batch_mode = False
+
             # current persist a16w16 mla_decode kernel does not support head_num = 128
             # need to fall back to non-persist
             # only use mla_ps_kernel when fp8 kv_cache
-            # for non-fp8 kv_cache, use non-persist kernel to avoid performance degradation
+            # for non-fp8 kv_cache on tp8, use non-persist kernel to avoid performance degradation
-            if self.kv_cache_dtype is not fp8_dtype:
+            if self.num_head == 16 and self.kv_cache_dtype is not fp8_dtype:
                 _use_mla_ps_kernel = False
                 fast_mode = False
                 intra_batch_mode = False
@@ -301,7 +305,7 @@ class AiterAttnBackend(AttentionBackend):
             kv_last_page_len,
             self.num_head // nhead_kv,
             nhead_kv,
-            True,
+            False,
             work_metadata,
             work_info_set,
             work_indptr,