diff --git a/.github/workflows/bot_pr_create.yaml b/.github/workflows/bot_pr_create.yaml
index bb956c2a8fd..f2f878d46e2 100644
--- a/.github/workflows/bot_pr_create.yaml
+++ b/.github/workflows/bot_pr_create.yaml
@@ -37,7 +37,7 @@ jobs:
     steps:
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=15d76f74e2fdb12a95ea00f0ca283acf6219a2b7
+          VLLM_COMMIT=6290470843c131681e3e1318ae71070a34f33225
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> "$GITHUB_ENV"
 
       - name: Checkout repository
diff --git a/.github/workflows/pr_test_full.yaml b/.github/workflows/pr_test_full.yaml
index 9a40f5e344c..a585aa762cc 100644
--- a/.github/workflows/pr_test_full.yaml
+++ b/.github/workflows/pr_test_full.yaml
@@ -75,7 +75,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [15d76f74e2fdb12a95ea00f0ca283acf6219a2b7, v0.16.0]
+        vllm_version: [6290470843c131681e3e1318ae71070a34f33225, v0.16.0]
     needs: [changes]
     if: ${{ needs.changes.outputs.e2e_tracker == 'true' || needs.changes.outputs.e2e_tracker == true }}
     uses: ./.github/workflows/_e2e_test.yaml
diff --git a/.github/workflows/pr_test_light.yaml b/.github/workflows/pr_test_light.yaml
index 94393cf05fa..c5325cb2bed 100644
--- a/.github/workflows/pr_test_light.yaml
+++ b/.github/workflows/pr_test_light.yaml
@@ -41,7 +41,7 @@ jobs:
   lint:
     uses: ./.github/workflows/_pre_commit.yml
     with:
-      vllm: 15d76f74e2fdb12a95ea00f0ca283acf6219a2b7
+      vllm: 6290470843c131681e3e1318ae71070a34f33225
   changes:
     runs-on: linux-aarch64-a2b3-0
     outputs:
@@ -87,7 +87,7 @@ jobs:
     if: ${{ needs.lint.result == 'success' && (needs.changes.outputs.e2e_tracker == 'true' || needs.changes.outputs.ut_tracker == 'true') }}
     strategy:
       matrix:
-        vllm_version: [15d76f74e2fdb12a95ea00f0ca283acf6219a2b7, v0.16.0]
+        vllm_version: [6290470843c131681e3e1318ae71070a34f33225, v0.16.0]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -99,7 +99,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [15d76f74e2fdb12a95ea00f0ca283acf6219a2b7, v0.16.0]
+        vllm_version: [6290470843c131681e3e1318ae71070a34f33225, v0.16.0]
     # Note (yikun): If CI resource are limited we can split job into two chain jobs
     needs: [lint, changes]
     # only trigger e2e test after lint passed and the change is e2e related with pull request.
diff --git a/.github/workflows/schedule_codecov_refresh.yaml b/.github/workflows/schedule_codecov_refresh.yaml
index e693cf7bd18..4611e0db93d 100644
--- a/.github/workflows/schedule_codecov_refresh.yaml
+++ b/.github/workflows/schedule_codecov_refresh.yaml
@@ -33,7 +33,7 @@ jobs:
     name: refresh codecov
     strategy:
       matrix:
-        vllm_version: [15d76f74e2fdb12a95ea00f0ca283acf6219a2b7]
+        vllm_version: [6290470843c131681e3e1318ae71070a34f33225]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
diff --git a/vllm_ascend/core/recompute_scheduler.py b/vllm_ascend/core/recompute_scheduler.py
index a33585582a7..466ed6099fc 100644
--- a/vllm_ascend/core/recompute_scheduler.py
+++ b/vllm_ascend/core/recompute_scheduler.py
@@ -23,7 +23,6 @@
 from dataclasses import dataclass, fields
 
 import numpy as np
-from vllm._bc_linter import bc_linter_include
 from vllm.config import SchedulerConfig, VllmConfig
 from vllm.distributed.ec_transfer.ec_connector.base import ECConnectorMetadata
 from vllm.distributed.kv_events import KVEventBatch
@@ -75,7 +74,6 @@ class RecomputeReqInfo:
     client_index: int = 0
 
 
-@bc_linter_include
 @dataclass
 class RecomputeSchedulerOutput(SchedulerOutput):
     recomputed_reqs: list[RecomputeReqInfo] | None = None
diff --git a/vllm_ascend/ops/mm_encoder_attention.py b/vllm_ascend/ops/mm_encoder_attention.py
index 889b88c42da..c07c6888434 100644
--- a/vllm_ascend/ops/mm_encoder_attention.py
+++ b/vllm_ascend/ops/mm_encoder_attention.py
@@ -97,6 +97,7 @@ def forward_oot(
         value: torch.Tensor,
         cu_seqlens: torch.Tensor | None = None,
         max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
+        sequence_lengths: torch.Tensor | None = None,
     ):
         bsz, q_len = query.size()[:2]
         kv_len = key.size(1)
diff --git a/vllm_ascend/platform.py b/vllm_ascend/platform.py
index c0af4d00054..b206d84f873 100644
--- a/vllm_ascend/platform.py
+++ b/vllm_ascend/platform.py
@@ -355,6 +355,17 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         if get_ascend_device_type() != AscendDeviceType._310P:
             compilation_config.custom_ops = ["all"]
 
+        # Upstream removed automatic +rms_norm forcing for SP (PR #35410),
+        # but Ascend needs it because torch.ops._C.rms_norm (CUDA kernel) is
+        # not available. Re-force it so the custom op wrapper routes to the
+        # OOT (NPU) implementation.
+        if (
+            compilation_config.pass_config.enable_sp
+            and "+rms_norm" not in compilation_config.custom_ops
+            and "-rms_norm" not in compilation_config.custom_ops
+        ):
+            compilation_config.custom_ops.append("+rms_norm")
+
         if ascend_config.recompute_scheduler_enable:
             from vllm_ascend.core.recompute_scheduler import RecomputeSchedulerConfig
 
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index ead030c2eff..5e7e23b1863 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -119,6 +119,7 @@
     is_moe_model,
     lmhead_tp_enable,
     set_weight_prefetch_method,
+    vllm_version_is,
 )
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 from vllm_ascend.worker.pcp_utils import PCPManager
@@ -1812,16 +1813,30 @@ def _determine_batch_execution_and_padding(
         has_lora = len(self.input_batch.lora_id_to_lora_request) > 0 if force_has_lora is None else force_has_lora
 
         # ruff: noqa: E731
-        dispatch_cudagraph = (
-            lambda num_tokens, disable_full: self.cudagraph_dispatcher.dispatch(
-                num_tokens=num_tokens,
-                has_lora=has_lora,
-                uniform_decode=uniform_decode,
-                disable_full=disable_full,
-            )
-            if not force_eager
-            else (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
-        )
+        if vllm_version_is("0.16.0"):
+
+            def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
+                if force_eager:
+                    return (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
+                return self.cudagraph_dispatcher.dispatch(
+                    num_tokens=num_tokens,
+                    has_lora=has_lora,
+                    uniform_decode=uniform_decode,
+                    disable_full=disable_full,
+                )
+        else:
+
+            def dispatch_cudagraph(num_tokens, disable_full=False, valid_modes=None):
+                if force_eager:
+                    return (CUDAGraphMode.NONE, BatchDescriptor(num_tokens_padded))
+                return self.cudagraph_dispatcher.dispatch(
+                    num_tokens=num_tokens,
+                    has_lora=has_lora,
+                    uniform_decode=uniform_decode,
+                    valid_modes=valid_modes,
+                    invalid_modes={CUDAGraphMode.FULL} if disable_full else None,
+                )
+
         cudagraph_mode, batch_descriptor = dispatch_cudagraph(num_tokens_padded, use_cascade_attn or has_encoder_output)
         num_tokens_padded = batch_descriptor.num_tokens
         if enable_sp(self.vllm_config):
@@ -1842,10 +1857,16 @@ def _determine_batch_execution_and_padding(
                 dp_rank = self.parallel_config.data_parallel_rank
                 num_tokens_padded = int(num_tokens_across_dp[dp_rank].item())
                 # Re-dispatch with DP padding
-                cudagraph_mode, batch_descriptor = dispatch_cudagraph(
-                    num_tokens_padded,
-                    disable_full=synced_cudagraph_mode <= CUDAGraphMode.PIECEWISE.value,
-                )
+                if vllm_version_is("0.16.0"):
+                    cudagraph_mode, batch_descriptor = dispatch_cudagraph(
+                        num_tokens_padded,
+                        disable_full=synced_cudagraph_mode <= CUDAGraphMode.PIECEWISE.value,
+                    )
+                else:
+                    cudagraph_mode, batch_descriptor = dispatch_cudagraph(
+                        num_tokens_padded,
+                        valid_modes={CUDAGraphMode(synced_cudagraph_mode)},
+                    )
                 # Assert to make sure the agreed upon token count is correct otherwise
                 # num_tokens_across_dp will no-longer be valid
                 assert batch_descriptor.num_tokens == num_tokens_padded
diff --git a/vllm_ascend/worker/worker.py b/vllm_ascend/worker/worker.py
index 1cadf05b233..f35ac135e05 100644
--- a/vllm_ascend/worker/worker.py
+++ b/vllm_ascend/worker/worker.py
@@ -428,7 +428,7 @@ def load_model(self) -> None:
         with context, set_current_vllm_config(self.vllm_config):
             self.model_runner.load_model()
 
-    def compile_or_warm_up_model(self) -> None:
+    def compile_or_warm_up_model(self) -> float:
         # Note: need to adapt for graph mode.
         warmup_sizes = (self.vllm_config.compilation_config.compile_sizes or []).copy()
         if not self.model_config.enforce_eager:
@@ -460,6 +460,9 @@ def compile_or_warm_up_model(self) -> None:
         # Reset the seed to ensure that the random state is not affected by
         # the model initialization and profiling.
         set_random_seed(self.model_config.seed)
+        return self.vllm_config.compilation_config.compilation_time
+
+        return self.vllm_config.compilation_config.compilation_time
 
     def _warm_up_atb(self):
         x = torch.rand((2, 4), dtype=torch.float16).npu()