diff --git a/.github/workflows/pr-test-amd.yml b/.github/workflows/pr-test-amd.yml
index 1c52f0c7bb5..718a4f9076c 100644
--- a/.github/workflows/pr-test-amd.yml
+++ b/.github/workflows/pr-test-amd.yml
@@ -38,12 +38,12 @@ jobs:
           else
             DEVICE_FLAG="--device /dev/dri"
           fi
-          docker pull lmsysorg/sglang:v0.4.5-rocm630
+          docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
           docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
             -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
             --cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
             -w /sglang-checkout --name ci_sglang \
-            lmsysorg/sglang:v0.4.5-rocm630
+            lmsysorg/sglang:v0.4.5.post2-rocm630
 
       - name: Install dependencies
         run: |
@@ -82,12 +82,12 @@ jobs:
           else
             DEVICE_FLAG="--device /dev/dri"
           fi
-          docker pull lmsysorg/sglang:v0.4.5-rocm630
+          docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
           docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
             -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
             --cap-add=SYS_PTRACE -e HF_TOKEN=${{ secrets.AMD_HF_TOKEN }} --security-opt seccomp=unconfined \
             -w /sglang-checkout --name ci_sglang \
-            lmsysorg/sglang:v0.4.5-rocm630
+            lmsysorg/sglang:v0.4.5.post2-rocm630
 
       - name: Install dependencies
         run: |
@@ -120,12 +120,12 @@ jobs:
           else
             DEVICE_FLAG="--device /dev/dri"
           fi
-          docker pull lmsysorg/sglang:v0.4.5-rocm630
+          docker pull lmsysorg/sglang:v0.4.5.post2-rocm630
           docker run -dt --user root --device=/dev/kfd $DEVICE_FLAG \
             -v ${{ github.workspace }}:/sglang-checkout --ipc=host --group-add video \
             --cap-add=SYS_PTRACE -e HF_TOKEN=${HF_TOKEN} --security-opt seccomp=unconfined \
             -w /sglang-checkout --name ci_sglang \
-            lmsysorg/sglang:v0.4.5-rocm630
+            lmsysorg/sglang:v0.4.5.post2-rocm630
 
       - name: Install dependencies
         run: |
@@ -149,7 +149,7 @@ jobs:
   finish:
     if: always()
     needs: [
-      accuracy-test-1-gpu-amd, mla-test-1-gpu-amd
+      accuracy-test-1-gpu-amd, mla-test-1-gpu-amd, bench-test-2-gpu-amd
     ]
     runs-on: ubuntu-latest
     steps:
diff --git a/python/sglang/srt/layers/rotary_embedding.py b/python/sglang/srt/layers/rotary_embedding.py
index 95ae14b7246..e3a99655a0e 100644
--- a/python/sglang/srt/layers/rotary_embedding.py
+++ b/python/sglang/srt/layers/rotary_embedding.py
@@ -665,6 +665,7 @@ def forward_native(
         offsets: Optional[torch.Tensor] = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         """PyTorch-native implementation equivalent to forward()."""
+        dtype = query.dtype
         query_rot = query[..., : self.rotary_dim]
         key_rot = key[..., : self.rotary_dim]
         if self.rotary_dim < self.head_size:
@@ -695,7 +696,7 @@ def forward_native(
         else:
             query = query_rot
             key = key_rot
-        return query, key
+        return query.to(dtype), key.to(dtype)
 
 
 class Llama3RotaryEmbedding(RotaryEmbedding):
diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 3daeab95c08..a9cb3c5b91c 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -682,10 +682,6 @@ def forward_absorb(
         forward_batch: ForwardBatch,
         zero_allocator: BumpAllocator,
     ) -> torch.Tensor:
-        q_len = hidden_states.shape[0]
-        q_input = hidden_states.new_empty(
-            q_len, self.num_local_heads, self.kv_lora_rank + self.qk_rope_head_dim
-        )
         if self.q_lora_rank is not None:
             q = self.q_a_proj(hidden_states)[0]
             q = self.q_a_layernorm(q)
@@ -729,20 +725,20 @@ def forward_absorb(
             )
         else:
             q_nope_out = torch.bmm(q_nope.transpose(0, 1), self.w_kc)
-        q_input[..., : self.kv_lora_rank] = q_nope_out.transpose(0, 1)
+
+        q_nope_out = q_nope_out.transpose(0, 1)
 
         latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
-        v_input = latent_cache[..., : self.kv_lora_rank]
-        v_input = self.kv_a_layernorm(v_input.contiguous()).unsqueeze(1)
-        k_input = latent_cache.unsqueeze(1)
-        k_input[..., : self.kv_lora_rank] = v_input
-        k_pe = k_input[..., self.kv_lora_rank :]
+        k_nope = latent_cache[..., : self.kv_lora_rank]
+        k_nope = self.kv_a_layernorm(k_nope).unsqueeze(1)
+        k_pe = latent_cache[..., self.kv_lora_rank :].unsqueeze(1)
 
         q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
-        q_input[..., self.kv_lora_rank :] = q_pe
-        k_input[..., self.kv_lora_rank :] = k_pe
 
-        attn_output = self.attn_mqa(q_input, k_input, v_input, forward_batch)
+        q = torch.cat([q_nope_out, q_pe], dim=-1)
+        k = torch.cat([k_nope, k_pe], dim=-1)
+
+        attn_output = self.attn_mqa(q, k, k_nope, forward_batch)
         attn_output = attn_output.view(-1, self.num_local_heads, self.kv_lora_rank)
 
         if self.use_deep_gemm_bmm: