From 570b3536e74b25f438a636f51f4f6ac5368f6951 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Tue, 27 Jan 2026 07:10:09 +0000
Subject: [PATCH 1/8] main2main

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 .github/workflows/bot_pr_create.yaml            | 2 +-
 .github/workflows/dockerfiles/Dockerfile.lint   | 2 +-
 .github/workflows/pr_test_full.yaml             | 2 +-
 .github/workflows/pr_test_light.yaml            | 6 +++---
 .github/workflows/schedule_codecov_refresh.yaml | 2 +-
 docs/source/community/versioning_policy.md      | 2 +-
 vllm_ascend/ops/mm_encoder_attention.py         | 2 --
 7 files changed, 8 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/bot_pr_create.yaml b/.github/workflows/bot_pr_create.yaml
index 1dea09f844d..961b2361a5c 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=dc917cceb877dfd13f98c538c4c96158047d98bd
+          VLLM_COMMIT=cf1167e50b809f18efd21fb3418dd75d2805b14f
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository
diff --git a/.github/workflows/dockerfiles/Dockerfile.lint b/.github/workflows/dockerfiles/Dockerfile.lint
index 9bb2a5b8aa5..1abfbbe08d0 100644
--- a/.github/workflows/dockerfiles/Dockerfile.lint
+++ b/.github/workflows/dockerfiles/Dockerfile.lint
@@ -27,7 +27,7 @@ RUN apt-get update -y && \
 
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
 # For lint purpose, actually we need make a main2main matching.
-ARG VLLM_COMMIT=dc917cceb877dfd13f98c538c4c96158047d98bd
+ARG VLLM_COMMIT=cf1167e50b809f18efd21fb3418dd75d2805b14f
 RUN git clone $VLLM_REPO /vllm-workspace/vllm && \
     cd /vllm-workspace/vllm && \
     git checkout $VLLM_COMMIT
diff --git a/.github/workflows/pr_test_full.yaml b/.github/workflows/pr_test_full.yaml
index 740d2cb1245..2d61361775d 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: [dc917cceb877dfd13f98c538c4c96158047d98bd, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
     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 a698639b277..ce34ba1d7af 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: dc917cceb877dfd13f98c538c4c96158047d98bd
+      vllm: cf1167e50b809f18efd21fb3418dd75d2805b14f
   changes:
     runs-on: linux-aarch64-a2-0
     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: [dc917cceb877dfd13f98c538c4c96158047d98bd, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -99,7 +99,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [dc917cceb877dfd13f98c538c4c96158047d98bd, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
     # 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 98f862918a3..af0c278dcb5 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: [dc917cceb877dfd13f98c538c4c96158047d98bd]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
diff --git a/docs/source/community/versioning_policy.md b/docs/source/community/versioning_policy.md
index 801c8a2a08c..33f7ba5456b 100644
--- a/docs/source/community/versioning_policy.md
+++ b/docs/source/community/versioning_policy.md
@@ -55,7 +55,7 @@ For main branch of vLLM Ascend, we usually make it compatible with the latest vL
 
 | vLLM Ascend | vLLM         | Python           | Stable CANN | PyTorch/torch_npu  |
 |-------------|--------------|------------------|-------------|--------------------|
-|     main    | dc917cceb877dfd13f98c538c4c96158047d98bd, v0.14.1 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
+|     main    | cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
 
 ## Release cadence
 
diff --git a/vllm_ascend/ops/mm_encoder_attention.py b/vllm_ascend/ops/mm_encoder_attention.py
index 19f44066761..aa9a07374bf 100644
--- a/vllm_ascend/ops/mm_encoder_attention.py
+++ b/vllm_ascend/ops/mm_encoder_attention.py
@@ -38,7 +38,6 @@ def __init__(
         scale: float | None = None,
         num_kv_heads: int | None = None,
         prefix: str = "",
-        multimodal_config: MultiModalConfig | None = None,
     ) -> None:
         """
         Args:
@@ -56,7 +55,6 @@ def __init__(
             scale=scale,
             num_kv_heads=num_kv_heads,
             prefix=prefix,
-            multimodal_config=multimodal_config,
         )
 
     def reshape_qkv_to_3d(

From 39a37c53a7570dab22abacb899c8df7ef5313b53 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Tue, 27 Jan 2026 08:12:27 +0000
Subject: [PATCH 2/8] update

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 .github/workflows/_e2e_nightly_multi_node.yaml         | 2 +-
 .github/workflows/pr_test_full.yaml                    | 2 +-
 .github/workflows/pr_test_light.yaml                   | 4 ++--
 .github/workflows/schedule_nightly_test_a2.yaml        | 3 ++-
 .github/workflows/schedule_nightly_test_a3.yaml        | 2 ++
 .github/workflows/schedule_test_benchmarks.yaml        | 2 +-
 Dockerfile                                             | 2 +-
 Dockerfile.310p                                        | 2 +-
 Dockerfile.310p.openEuler                              | 2 +-
 Dockerfile.a3                                          | 2 +-
 Dockerfile.a3.openEuler                                | 2 +-
 Dockerfile.openEuler                                   | 2 +-
 docs/source/community/versioning_policy.md             | 4 ++--
 docs/source/conf.py                                    | 4 ++--
 vllm_ascend/ascend_forward_context.py                  | 4 +---
 vllm_ascend/attention/mla_v1.py                        | 7 ++-----
 vllm_ascend/attention/sfa_v1.py                        | 7 ++-----
 vllm_ascend/patch/platform/patch_multiproc_executor.py | 5 +----
 18 files changed, 25 insertions(+), 33 deletions(-)

diff --git a/.github/workflows/_e2e_nightly_multi_node.yaml b/.github/workflows/_e2e_nightly_multi_node.yaml
index fa3659e5600..1b1758a26e4 100644
--- a/.github/workflows/_e2e_nightly_multi_node.yaml
+++ b/.github/workflows/_e2e_nightly_multi_node.yaml
@@ -32,7 +32,7 @@ on:
         description: how many pods will be pulled up via lws.yaml, indicates number of nodes we need
       vllm_version:
         required: false
-        default: "v0.14.1"
+        default: "releases/v0.15.0"
         type: string
         description: vllm version to use
       vllm_ascend_remote_url:
diff --git a/.github/workflows/pr_test_full.yaml b/.github/workflows/pr_test_full.yaml
index 2d61361775d..34cbea34413 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: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
     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 ce34ba1d7af..257f5c7b5b5 100644
--- a/.github/workflows/pr_test_light.yaml
+++ b/.github/workflows/pr_test_light.yaml
@@ -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: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -99,7 +99,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
     # 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_nightly_test_a2.yaml b/.github/workflows/schedule_nightly_test_a2.yaml
index d25ffd58c66..fb6d2d8797b 100644
--- a/.github/workflows/schedule_nightly_test_a2.yaml
+++ b/.github/workflows/schedule_nightly_test_a2.yaml
@@ -66,6 +66,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/ops/multicard_ops_a2/
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
+      vllm: releases/v0.15.0
       runner: ${{ matrix.test_config.os }}
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
@@ -133,7 +134,7 @@ jobs:
               - Qwen3-Omni-30B-A3B-Instruct
     uses: ./.github/workflows/_e2e_nightly_single_node_models.yaml
     with:
-      vllm: v0.14.1
+      vllm: releases/v0.15.0
       runner: ${{ matrix.test_config.os }}
       model_list: ${{ toJson(matrix.test_config.model_list) }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11'
diff --git a/.github/workflows/schedule_nightly_test_a3.yaml b/.github/workflows/schedule_nightly_test_a3.yaml
index ed0b4383616..043d7b94022 100644
--- a/.github/workflows/schedule_nightly_test_a3.yaml
+++ b/.github/workflows/schedule_nightly_test_a3.yaml
@@ -161,6 +161,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
+      vllm: releases/v0.15.0
       runner: ${{ matrix.test_config.os }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
@@ -180,6 +181,7 @@ jobs:
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       runner: ${{ matrix.test_config.os }}
+      vllm: releases/v0.15.0
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
diff --git a/.github/workflows/schedule_test_benchmarks.yaml b/.github/workflows/schedule_test_benchmarks.yaml
index 8585670ddd7..af694b4d647 100644
--- a/.github/workflows/schedule_test_benchmarks.yaml
+++ b/.github/workflows/schedule_test_benchmarks.yaml
@@ -51,7 +51,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - vllm_branch: v0.14.1
+          - vllm_branch: releases/v0.15.0
             vllm_ascend_branch: main
       max-parallel: 1
     container:
diff --git a/Dockerfile b/Dockerfile
index 02e2820969a..5ee932c1231 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -48,7 +48,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.310p b/Dockerfile.310p
index 07b97bc1ca3..a7b2ffdc942 100644
--- a/Dockerfile.310p
+++ b/Dockerfile.310p
@@ -40,7 +40,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.310p.openEuler b/Dockerfile.310p.openEuler
index abf70a8812c..29370a9517d 100644
--- a/Dockerfile.310p.openEuler
+++ b/Dockerfile.310p.openEuler
@@ -36,7 +36,7 @@ COPY . /vllm-workspace/vllm-ascend/
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.a3 b/Dockerfile.a3
index 7688c14d6a8..c9ce41110b4 100644
--- a/Dockerfile.a3
+++ b/Dockerfile.a3
@@ -47,7 +47,7 @@ RUN apt-get update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.a3.openEuler b/Dockerfile.a3.openEuler
index 6129adc7982..d5b74c6ba79 100644
--- a/Dockerfile.a3.openEuler
+++ b/Dockerfile.a3.openEuler
@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.openEuler b/Dockerfile.openEuler
index 7603372f6af..12ce32a5ffb 100644
--- a/Dockerfile.openEuler
+++ b/Dockerfile.openEuler
@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.14.1
+ARG VLLM_TAG=releases/v0.15.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/docs/source/community/versioning_policy.md b/docs/source/community/versioning_policy.md
index 33f7ba5456b..01ecb86c22c 100644
--- a/docs/source/community/versioning_policy.md
+++ b/docs/source/community/versioning_policy.md
@@ -23,7 +23,7 @@ The table below is the release compatibility matrix for vLLM Ascend release.
 
 | vLLM Ascend | vLLM              | Python          | Stable CANN |        PyTorch/torch_npu        | Triton Ascend |
 |-------------|-------------------|-----------------|-------------|---------------------------------|---------------|
-| v0.14.0rc1  | v0.14.0/v0.14.1   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
+| v0.14.0rc1  | v0.14.0/releases/v0.15.0   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
 | v0.13.0rc2  | v0.13.0           | >= 3.10, < 3.12 | 8.5.0       | 2.8.0  / 2.8.0.post1            | 3.2.0         |
 | v0.13.0rc1  | v0.13.0           | >= 3.10, < 3.12 | 8.3.RC2     | 2.8.0  / 2.8.0                  |               |
 | v0.11.0     | v0.11.0           | >= 3.9 , < 3.12 | 8.3.RC2     | 2.7.1 / 2.7.1.post1             |               |
@@ -55,7 +55,7 @@ For main branch of vLLM Ascend, we usually make it compatible with the latest vL
 
 | vLLM Ascend | vLLM         | Python           | Stable CANN | PyTorch/torch_npu  |
 |-------------|--------------|------------------|-------------|--------------------|
-|     main    | cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.14.1 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
+|     main    | cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
 
 ## Release cadence
 
diff --git a/docs/source/conf.py b/docs/source/conf.py
index 47ebcec8a09..3e09a63e6eb 100644
--- a/docs/source/conf.py
+++ b/docs/source/conf.py
@@ -65,7 +65,7 @@
     # the branch of vllm, used in vllm clone
     # - main branch: 'main'
     # - vX.Y.Z branch: 'vX.Y.Z'
-    "vllm_version": "v0.14.1",
+    "vllm_version": "releases/v0.15.0",
     # the branch of vllm-ascend, used in vllm-ascend clone and image tag
     # - main branch: 'main'
     # - vX.Y.Z branch: latest vllm-ascend release tag
@@ -77,7 +77,7 @@
     # CANN image tag
     "cann_image_tag": "8.5.0-910b-ubuntu22.04-py3.11",
     # vllm version in ci
-    "ci_vllm_version": "v0.14.1",
+    "ci_vllm_version": "releases/v0.15.0",
 }
 
 # For cross-file header anchors
diff --git a/vllm_ascend/ascend_forward_context.py b/vllm_ascend/ascend_forward_context.py
index faa11d2d43d..61c09afb5ff 100644
--- a/vllm_ascend/ascend_forward_context.py
+++ b/vllm_ascend/ascend_forward_context.py
@@ -19,7 +19,6 @@
     is_drafter_moe_model,
     is_moe_model,
     speculative_enable_dispatch_gmm_combine_decode,
-    vllm_version_is,
 )
 
 
@@ -59,8 +58,7 @@ def set_ascend_forward_context(
         "batch_descriptor": batch_descriptor,
     }
 
-    if not vllm_version_is("0.14.1"):
-        forward_context_kwargs["skip_compiled"] = skip_compiled
+    forward_context_kwargs["skip_compiled"] = skip_compiled
 
     with set_forward_context(**forward_context_kwargs):
         forward_context = get_forward_context()
diff --git a/vllm_ascend/attention/mla_v1.py b/vllm_ascend/attention/mla_v1.py
index 03834212bdf..5c12b8153d4 100644
--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -44,17 +44,14 @@
 from vllm_ascend.ops.rotary_embedding import get_cos_and_sin_mla
 from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
 from vllm_ascend.quantization.methods import AscendW8A8LinearMethod
-from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, maybe_trans_nz, vllm_version_is, weak_ref_tensors
+from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, maybe_trans_nz, weak_ref_tensors
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
 
 # isort: off
-if vllm_version_is("0.14.1"):
-    from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder  # type: ignore
-else:
-    from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
+from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
 # isort: on
 
 MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
diff --git a/vllm_ascend/attention/sfa_v1.py b/vllm_ascend/attention/sfa_v1.py
index 6f283429d27..6d2bb1f859d 100644
--- a/vllm_ascend/attention/sfa_v1.py
+++ b/vllm_ascend/attention/sfa_v1.py
@@ -45,16 +45,13 @@
     enable_dsa_cp,
     enable_dsa_cp_with_layer_shard,
     maybe_trans_nz,
-    vllm_version_is,
 )
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
-if vllm_version_is("0.14.1"):
-    from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder  # type: ignore
-else:
-    from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
+from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
+
 # isort: on
 
 # token count limits within bmm_transpose operator
diff --git a/vllm_ascend/patch/platform/patch_multiproc_executor.py b/vllm_ascend/patch/platform/patch_multiproc_executor.py
index abc955e578c..98cc545c9de 100644
--- a/vllm_ascend/patch/platform/patch_multiproc_executor.py
+++ b/vllm_ascend/patch/platform/patch_multiproc_executor.py
@@ -19,8 +19,6 @@
     set_multiprocessing_worker_envs,
 )
 
-from vllm_ascend.utils import vllm_version_is
-
 
 class AscendMultiprocExecutor(MultiprocExecutor):
     def _init_executor(self) -> None:
@@ -178,8 +176,7 @@ def make_worker_process(
             "death_pipe": death_reader,
             "shared_worker_lock": shared_worker_lock,
         }
-        if not vllm_version_is("0.14.1"):
-            process_kwargs["is_driver_worker"] = is_driver_worker
+        process_kwargs["is_driver_worker"] = is_driver_worker
         # Run EngineCore busy loop in background process.
         proc = context.Process(
             target=WorkerProc.worker_main,

From 7d71da618ca5a8621f60e8ba2442b887660df6ba Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Tue, 27 Jan 2026 09:22:31 +0000
Subject: [PATCH 3/8] update

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 .github/workflows/_e2e_nightly_multi_node.yaml  | 2 +-
 .github/workflows/pr_test_full.yaml             | 2 +-
 .github/workflows/pr_test_light.yaml            | 4 ++--
 .github/workflows/schedule_nightly_test_a2.yaml | 4 ++--
 .github/workflows/schedule_nightly_test_a3.yaml | 4 ++--
 .github/workflows/schedule_test_benchmarks.yaml | 2 +-
 Dockerfile                                      | 2 +-
 Dockerfile.310p                                 | 2 +-
 Dockerfile.310p.openEuler                       | 2 +-
 Dockerfile.a3                                   | 2 +-
 Dockerfile.a3.openEuler                         | 2 +-
 Dockerfile.openEuler                            | 2 +-
 docs/source/community/versioning_policy.md      | 4 ++--
 docs/source/conf.py                             | 4 ++--
 14 files changed, 19 insertions(+), 19 deletions(-)

diff --git a/.github/workflows/_e2e_nightly_multi_node.yaml b/.github/workflows/_e2e_nightly_multi_node.yaml
index 1b1758a26e4..84593c57e0f 100644
--- a/.github/workflows/_e2e_nightly_multi_node.yaml
+++ b/.github/workflows/_e2e_nightly_multi_node.yaml
@@ -32,7 +32,7 @@ on:
         description: how many pods will be pulled up via lws.yaml, indicates number of nodes we need
       vllm_version:
         required: false
-        default: "releases/v0.15.0"
+        default: "v0.15.0rc0"
         type: string
         description: vllm version to use
       vllm_ascend_remote_url:
diff --git a/.github/workflows/pr_test_full.yaml b/.github/workflows/pr_test_full.yaml
index 34cbea34413..8cdb4496fa7 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: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.15.0rc0]
     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 257f5c7b5b5..b0a389909c9 100644
--- a/.github/workflows/pr_test_light.yaml
+++ b/.github/workflows/pr_test_light.yaml
@@ -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: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.15.0rc0]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -99,7 +99,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0]
+        vllm_version: [cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.15.0rc0]
     # 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_nightly_test_a2.yaml b/.github/workflows/schedule_nightly_test_a2.yaml
index fb6d2d8797b..ab3fe4684c7 100644
--- a/.github/workflows/schedule_nightly_test_a2.yaml
+++ b/.github/workflows/schedule_nightly_test_a2.yaml
@@ -66,7 +66,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/ops/multicard_ops_a2/
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: releases/v0.15.0
+      vllm: v0.15.0rc0
       runner: ${{ matrix.test_config.os }}
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
@@ -134,7 +134,7 @@ jobs:
               - Qwen3-Omni-30B-A3B-Instruct
     uses: ./.github/workflows/_e2e_nightly_single_node_models.yaml
     with:
-      vllm: releases/v0.15.0
+      vllm: v0.15.0rc0
       runner: ${{ matrix.test_config.os }}
       model_list: ${{ toJson(matrix.test_config.model_list) }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11'
diff --git a/.github/workflows/schedule_nightly_test_a3.yaml b/.github/workflows/schedule_nightly_test_a3.yaml
index 043d7b94022..be84cf96607 100644
--- a/.github/workflows/schedule_nightly_test_a3.yaml
+++ b/.github/workflows/schedule_nightly_test_a3.yaml
@@ -161,7 +161,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: releases/v0.15.0
+      vllm: v0.15.0rc0
       runner: ${{ matrix.test_config.os }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
@@ -181,7 +181,7 @@ jobs:
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       runner: ${{ matrix.test_config.os }}
-      vllm: releases/v0.15.0
+      vllm: v0.15.0rc0
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
diff --git a/.github/workflows/schedule_test_benchmarks.yaml b/.github/workflows/schedule_test_benchmarks.yaml
index af694b4d647..904bee0a84a 100644
--- a/.github/workflows/schedule_test_benchmarks.yaml
+++ b/.github/workflows/schedule_test_benchmarks.yaml
@@ -51,7 +51,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - vllm_branch: releases/v0.15.0
+          - vllm_branch: v0.15.0rc0
             vllm_ascend_branch: main
       max-parallel: 1
     container:
diff --git a/Dockerfile b/Dockerfile
index 5ee932c1231..75ae6462725 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -48,7 +48,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.310p b/Dockerfile.310p
index a7b2ffdc942..f4f142c8ccd 100644
--- a/Dockerfile.310p
+++ b/Dockerfile.310p
@@ -40,7 +40,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.310p.openEuler b/Dockerfile.310p.openEuler
index 29370a9517d..b07fbb83e90 100644
--- a/Dockerfile.310p.openEuler
+++ b/Dockerfile.310p.openEuler
@@ -36,7 +36,7 @@ COPY . /vllm-workspace/vllm-ascend/
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.a3 b/Dockerfile.a3
index c9ce41110b4..94b40bc16f1 100644
--- a/Dockerfile.a3
+++ b/Dockerfile.a3
@@ -47,7 +47,7 @@ RUN apt-get update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.a3.openEuler b/Dockerfile.a3.openEuler
index d5b74c6ba79..65de41960dc 100644
--- a/Dockerfile.a3.openEuler
+++ b/Dockerfile.a3.openEuler
@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/Dockerfile.openEuler b/Dockerfile.openEuler
index 12ce32a5ffb..269e4d7a15f 100644
--- a/Dockerfile.openEuler
+++ b/Dockerfile.openEuler
@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=releases/v0.15.0
+ARG VLLM_TAG=v0.15.0rc0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \
diff --git a/docs/source/community/versioning_policy.md b/docs/source/community/versioning_policy.md
index 01ecb86c22c..73dbb53a65e 100644
--- a/docs/source/community/versioning_policy.md
+++ b/docs/source/community/versioning_policy.md
@@ -23,7 +23,7 @@ The table below is the release compatibility matrix for vLLM Ascend release.
 
 | vLLM Ascend | vLLM              | Python          | Stable CANN |        PyTorch/torch_npu        | Triton Ascend |
 |-------------|-------------------|-----------------|-------------|---------------------------------|---------------|
-| v0.14.0rc1  | v0.14.0/releases/v0.15.0   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
+| v0.14.0rc1  | v0.14.0/v0.15.0rc0   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
 | v0.13.0rc2  | v0.13.0           | >= 3.10, < 3.12 | 8.5.0       | 2.8.0  / 2.8.0.post1            | 3.2.0         |
 | v0.13.0rc1  | v0.13.0           | >= 3.10, < 3.12 | 8.3.RC2     | 2.8.0  / 2.8.0                  |               |
 | v0.11.0     | v0.11.0           | >= 3.9 , < 3.12 | 8.3.RC2     | 2.7.1 / 2.7.1.post1             |               |
@@ -55,7 +55,7 @@ For main branch of vLLM Ascend, we usually make it compatible with the latest vL
 
 | vLLM Ascend | vLLM         | Python           | Stable CANN | PyTorch/torch_npu  |
 |-------------|--------------|------------------|-------------|--------------------|
-|     main    | cf1167e50b809f18efd21fb3418dd75d2805b14f, releases/v0.15.0 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
+|     main    | cf1167e50b809f18efd21fb3418dd75d2805b14f, v0.15.0rc0 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
 
 ## Release cadence
 
diff --git a/docs/source/conf.py b/docs/source/conf.py
index 3e09a63e6eb..a00bccc5d56 100644
--- a/docs/source/conf.py
+++ b/docs/source/conf.py
@@ -65,7 +65,7 @@
     # the branch of vllm, used in vllm clone
     # - main branch: 'main'
     # - vX.Y.Z branch: 'vX.Y.Z'
-    "vllm_version": "releases/v0.15.0",
+    "vllm_version": "v0.15.0rc0",
     # the branch of vllm-ascend, used in vllm-ascend clone and image tag
     # - main branch: 'main'
     # - vX.Y.Z branch: latest vllm-ascend release tag
@@ -77,7 +77,7 @@
     # CANN image tag
     "cann_image_tag": "8.5.0-910b-ubuntu22.04-py3.11",
     # vllm version in ci
-    "ci_vllm_version": "releases/v0.15.0",
+    "ci_vllm_version": "v0.15.0rc0",
 }
 
 # For cross-file header anchors

From 46d18d38946f10c326b588d43f8b73f2930920f8 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Wed, 28 Jan 2026 06:30:33 +0000
Subject: [PATCH 4/8] update

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 docs/source/community/versioning_policy.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/source/community/versioning_policy.md b/docs/source/community/versioning_policy.md
index 73dbb53a65e..3637d89cc5f 100644
--- a/docs/source/community/versioning_policy.md
+++ b/docs/source/community/versioning_policy.md
@@ -23,7 +23,7 @@ The table below is the release compatibility matrix for vLLM Ascend release.
 
 | vLLM Ascend | vLLM              | Python          | Stable CANN |        PyTorch/torch_npu        | Triton Ascend |
 |-------------|-------------------|-----------------|-------------|---------------------------------|---------------|
-| v0.14.0rc1  | v0.14.0/v0.15.0rc0   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
+| v0.14.0rc1  | v0.14.0/v0.14.1   | >= 3.10, < 3.12 | 8.5.0       | 2.9.0  / 2.9.0                  | 3.2.0         |
 | v0.13.0rc2  | v0.13.0           | >= 3.10, < 3.12 | 8.5.0       | 2.8.0  / 2.8.0.post1            | 3.2.0         |
 | v0.13.0rc1  | v0.13.0           | >= 3.10, < 3.12 | 8.3.RC2     | 2.8.0  / 2.8.0                  |               |
 | v0.11.0     | v0.11.0           | >= 3.9 , < 3.12 | 8.3.RC2     | 2.7.1 / 2.7.1.post1             |               |

From 516c1702f66935862c8904ee5d448db24a366425 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Thu, 29 Jan 2026 07:40:42 +0000
Subject: [PATCH 5/8] split attn forward and kv_cache update

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 vllm_ascend/attention/attention_v1.py | 12 +++--
 vllm_ascend/ops/kv_cache.py           | 78 +++++++++++++++++++++++++++
 2 files changed, 85 insertions(+), 5 deletions(-)
 create mode 100644 vllm_ascend/ops/kv_cache.py

diff --git a/vllm_ascend/attention/attention_v1.py b/vllm_ascend/attention/attention_v1.py
index 768b908224a..66e34feadd8 100644
--- a/vllm_ascend/attention/attention_v1.py
+++ b/vllm_ascend/attention/attention_v1.py
@@ -854,13 +854,13 @@ def _forward_encoder_attention(
                 actual_seq_kvlen=attn_metadata.actual_seq_lengths_q,
             )[0]
 
-    def reshape_and_cache(
+    def do_kv_cache_update(
         self,
         key: torch.Tensor,
         value: torch.Tensor,
-        kv_cache: tuple[torch.Tensor],
+        kv_cache: torch.Tensor,
         attn_metadata: AscendMetadata,
-    ):
+    ) -> None:
         if len(kv_cache) > 1:
             if self.is_kv_producer:
                 attn_metadata.reshape_cache_event = torch.npu.Event()
@@ -932,8 +932,10 @@ def forward(
         num_tokens = query.shape[0]
         if attn_metadata is None:
             return output.fill_(0)
-        if key is not None and value is not None:
-            key, value = self.reshape_and_cache(key, value, kv_cache, attn_metadata)
+
+        if len(kv_cache) > 1 and self.key_cache is None:
+            self.key_cache, self.value_cache = kv_cache[0], kv_cache[1]
+
         # pooling model branch
         if attn_metadata.model_runner_type == "pooling":
             attn_output = self._forward_encoder_attention(query, key, value, attn_metadata, output)
diff --git a/vllm_ascend/ops/kv_cache.py b/vllm_ascend/ops/kv_cache.py
new file mode 100644
index 00000000000..cf66349f542
--- /dev/null
+++ b/vllm_ascend/ops/kv_cache.py
@@ -0,0 +1,78 @@
+# SPDX-License-Identifier: Apache-2.0
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+# Copyright 2023 The vLLM team.
+# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+from vllm.forward_context import get_forward_context
+from vllm.utils.torch_utils import (
+    direct_register_custom_op,
+)
+
+
+def _unified_kv_cache_update_impl(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    layer_name: str,
+) -> torch.Tensor:
+    """
+    Returns a dummy that is passed to unified_attention to signal a side effect and
+    the data dependency between them to ensure torch.compile preserves ordering.
+    """
+    forward_context = get_forward_context()
+    attn_layer = forward_context.no_compile_layers[layer_name]
+    kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]
+    attn_metadata = forward_context.attn_metadata
+
+    slot_mapping = forward_context.slot_mapping
+    assert isinstance(slot_mapping, dict), f"Expected slot_mapping to be a dict, got {type(slot_mapping)}. "
+    layer_slot_mapping = slot_mapping.get(layer_name)
+
+    if layer_slot_mapping is not None and key is not None and value is not None:
+        assert hasattr(attn_layer.impl, "do_kv_cache_update"), (
+            f"{attn_layer.impl.__class__.__name__} does not support kv cache update"
+        )
+
+        attn_layer.impl.do_kv_cache_update(
+            key,
+            value,
+            kv_cache,
+            attn_metadata,
+        )
+
+    return torch.empty(0, device=kv_cache.device, dtype=kv_cache.dtype)
+
+
+def _unified_kv_cache_update_impl_fake(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    layer_name: str,
+) -> torch.Tensor:
+    return torch.empty(0, device=key.device, dtype=key.dtype)
+
+
+direct_register_custom_op(
+    op_name="unified_kv_cache_update",
+    op_func=_unified_kv_cache_update_impl,
+    fake_impl=_unified_kv_cache_update_impl_fake,
+    mutates_args=[],
+    dispatch_key="PrivateUse1",
+)

From b110014b2d728913643580aa2966d1ec61c02ef6 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Fri, 30 Jan 2026 03:38:22 +0000
Subject: [PATCH 6/8] fix none shared experts case

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 vllm_ascend/ops/fused_moe/fused_moe.py | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/vllm_ascend/ops/fused_moe/fused_moe.py b/vllm_ascend/ops/fused_moe/fused_moe.py
index 50618ae8152..7e1bde15483 100644
--- a/vllm_ascend/ops/fused_moe/fused_moe.py
+++ b/vllm_ascend/ops/fused_moe/fused_moe.py
@@ -512,6 +512,15 @@ def forward(
         hidden_states: torch.Tensor,
         router_logits: torch.Tensor,
     ) -> tuple[torch.Tensor, torch.Tensor]:
+        if self._shared_experts is None:
+            fused_out = AscendFusedMoE.forward(
+                self,
+                hidden_states=hidden_states,
+                router_logits=router_logits,
+            )
+            shared_out = None
+            return shared_out, fused_out
+
         shared_out, fused_out = AscendFusedMoE.forward(
             self,
             hidden_states=hidden_states,
@@ -571,6 +580,9 @@ def forward_impl(  # type: ignore[override]
         )
         routed_out = fused_moe_results.routed_out
 
+        if self._shared_experts is None:
+            return routed_out
+        
         if self.multistream_overlap_gate:
             fc3_context = get_flash_common3_context()
             assert fc3_context is not None

From 850c6d6bd724d81341a4a1862cca0b812c36f5c3 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Mon, 2 Feb 2026 03:04:36 +0000
Subject: [PATCH 7/8] fix

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 vllm_ascend/attention/attention_v1.py   | 15 +++--
 vllm_ascend/ops/kv_cache.py             | 78 -------------------------
 vllm_ascend/ops/mm_encoder_attention.py | 14 ++---
 3 files changed, 16 insertions(+), 91 deletions(-)
 delete mode 100644 vllm_ascend/ops/kv_cache.py

diff --git a/vllm_ascend/attention/attention_v1.py b/vllm_ascend/attention/attention_v1.py
index 66e34feadd8..d174cabc598 100644
--- a/vllm_ascend/attention/attention_v1.py
+++ b/vllm_ascend/attention/attention_v1.py
@@ -66,6 +66,8 @@
 class AscendAttentionBackend(AttentionBackend):
     accept_output_buffer: bool = True
 
+    forward_includes_kv_cache_update: bool = True
+
     @staticmethod
     def get_name() -> str:
         # HACK(Ronald1995): vllm `initialize_kv_cache` method in model runner v2 make
@@ -854,13 +856,13 @@ def _forward_encoder_attention(
                 actual_seq_kvlen=attn_metadata.actual_seq_lengths_q,
             )[0]
 
-    def do_kv_cache_update(
+    def reshape_and_cache(
         self,
         key: torch.Tensor,
         value: torch.Tensor,
-        kv_cache: torch.Tensor,
+        kv_cache: tuple[torch.Tensor],
         attn_metadata: AscendMetadata,
-    ) -> None:
+    ):
         if len(kv_cache) > 1:
             if self.is_kv_producer:
                 attn_metadata.reshape_cache_event = torch.npu.Event()
@@ -868,6 +870,9 @@ def do_kv_cache_update(
                 self.key_cache, self.value_cache = kv_cache[0], kv_cache[1]
             slots = attn_metadata.slot_mapping
             encoder_decoder = self.attn_type == AttentionType.ENCODER_DECODER
+
+            slots = slots.to(torch.int32)
+
             DeviceOperator.reshape_and_cache(
                 key=key[: attn_metadata.num_actual_tokens] if not encoder_decoder else key,
                 value=value[: attn_metadata.num_actual_tokens] if not encoder_decoder else value,
@@ -933,8 +938,8 @@ def forward(
         if attn_metadata is None:
             return output.fill_(0)
 
-        if len(kv_cache) > 1 and self.key_cache is None:
-            self.key_cache, self.value_cache = kv_cache[0], kv_cache[1]
+        if key is not None and value is not None:
+            key, value = self.reshape_and_cache(key, value, kv_cache, attn_metadata)
 
         # pooling model branch
         if attn_metadata.model_runner_type == "pooling":
diff --git a/vllm_ascend/ops/kv_cache.py b/vllm_ascend/ops/kv_cache.py
deleted file mode 100644
index cf66349f542..00000000000
--- a/vllm_ascend/ops/kv_cache.py
+++ /dev/null
@@ -1,78 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
-# Copyright 2023 The vLLM team.
-# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
-#
-# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
-# and OPT implementations in this library. It has been modified from its
-# original forms to accommodate minor architectural differences compared
-# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-
-import torch
-from vllm.forward_context import get_forward_context
-from vllm.utils.torch_utils import (
-    direct_register_custom_op,
-)
-
-
-def _unified_kv_cache_update_impl(
-    key: torch.Tensor,
-    value: torch.Tensor,
-    layer_name: str,
-) -> torch.Tensor:
-    """
-    Returns a dummy that is passed to unified_attention to signal a side effect and
-    the data dependency between them to ensure torch.compile preserves ordering.
-    """
-    forward_context = get_forward_context()
-    attn_layer = forward_context.no_compile_layers[layer_name]
-    kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]
-    attn_metadata = forward_context.attn_metadata
-
-    slot_mapping = forward_context.slot_mapping
-    assert isinstance(slot_mapping, dict), f"Expected slot_mapping to be a dict, got {type(slot_mapping)}. "
-    layer_slot_mapping = slot_mapping.get(layer_name)
-
-    if layer_slot_mapping is not None and key is not None and value is not None:
-        assert hasattr(attn_layer.impl, "do_kv_cache_update"), (
-            f"{attn_layer.impl.__class__.__name__} does not support kv cache update"
-        )
-
-        attn_layer.impl.do_kv_cache_update(
-            key,
-            value,
-            kv_cache,
-            attn_metadata,
-        )
-
-    return torch.empty(0, device=kv_cache.device, dtype=kv_cache.dtype)
-
-
-def _unified_kv_cache_update_impl_fake(
-    key: torch.Tensor,
-    value: torch.Tensor,
-    layer_name: str,
-) -> torch.Tensor:
-    return torch.empty(0, device=key.device, dtype=key.dtype)
-
-
-direct_register_custom_op(
-    op_name="unified_kv_cache_update",
-    op_func=_unified_kv_cache_update_impl,
-    fake_impl=_unified_kv_cache_update_impl_fake,
-    mutates_args=[],
-    dispatch_key="PrivateUse1",
-)
diff --git a/vllm_ascend/ops/mm_encoder_attention.py b/vllm_ascend/ops/mm_encoder_attention.py
index aa9a07374bf..5ef46cd9c42 100644
--- a/vllm_ascend/ops/mm_encoder_attention.py
+++ b/vllm_ascend/ops/mm_encoder_attention.py
@@ -47,7 +47,6 @@ def __init__(
             num_kv_heads: number of kv heads.
             prefix: This has no effect, it is only here to make it easier to
                     swap between Attention and MMEncoderAttention.
-            multimodal_config: configs for multi-modal.
         """
         super().__init__(
             num_heads=num_heads,
@@ -82,13 +81,12 @@ def reshape_qkv_to_3d(
         return query, key, value
 
     def forward_oot(
-            self,
-            query: torch.Tensor,
-            key: torch.Tensor,
-            value: torch.Tensor,
-            cu_seqlens: torch.Tensor | None = None,
-            max_seqlen: torch.Tensor
-        | None = None,  # Only used for Flash Attention
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        cu_seqlens: torch.Tensor | None = None,
+        max_seqlen: torch.Tensor | None = None,  # Only used for Flash Attention
     ):
         bsz, q_len = query.size()[:2]
         kv_len = key.size(1)

From 1fe1c542cb023173a947acc3796dccece6fabf15 Mon Sep 17 00:00:00 2001
From: shen-shanshan <467638484@qq.com>
Date: Mon, 2 Feb 2026 03:22:44 +0000
Subject: [PATCH 8/8] fix lint

Signed-off-by: shen-shanshan <467638484@qq.com>
---
 .github/workflows/schedule_nightly_test_a2.yaml | 1 -
 .github/workflows/schedule_nightly_test_a3.yaml | 2 --
 2 files changed, 3 deletions(-)

diff --git a/.github/workflows/schedule_nightly_test_a2.yaml b/.github/workflows/schedule_nightly_test_a2.yaml
index ab3fe4684c7..9577d0e4120 100644
--- a/.github/workflows/schedule_nightly_test_a2.yaml
+++ b/.github/workflows/schedule_nightly_test_a2.yaml
@@ -66,7 +66,6 @@ jobs:
             tests: tests/e2e/nightly/single_node/ops/multicard_ops_a2/
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: v0.15.0rc0
       runner: ${{ matrix.test_config.os }}
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
diff --git a/.github/workflows/schedule_nightly_test_a3.yaml b/.github/workflows/schedule_nightly_test_a3.yaml
index be84cf96607..ed0b4383616 100644
--- a/.github/workflows/schedule_nightly_test_a3.yaml
+++ b/.github/workflows/schedule_nightly_test_a3.yaml
@@ -161,7 +161,6 @@ jobs:
             tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: v0.15.0rc0
       runner: ${{ matrix.test_config.os }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
@@ -181,7 +180,6 @@ jobs:
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       runner: ${{ matrix.test_config.os }}
-      vllm: v0.15.0rc0
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}