Integrate vLLM 0.30.0 upgrade with current main

.buildkite/cuda/test-merge.yml

@@ -34,4 +34,6 @@ steps:
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-eager-tp2]"
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-dp2]"
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-tp2]"
+            "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-eager-dbo-dp2]"
+            "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-dbo-dp2]"
         mirror_hardwares: l4_4

.buildkite/cuda/test-ready.yml

@@ -34,4 +34,6 @@ steps:
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-eager-tp2]"
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-dp2]"
             "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-tp2]"
+            "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-eager-dbo-dp2]"
+            "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-v2-graph-dbo-dp2]"
         mirror_hardwares: l4_4

.github/ISSUE_TEMPLATE/100-bug-report.yml

@@ -8,7 +8,7 @@ body:
     attributes:
       value: >
         Before submitting, please search existing issues to avoid duplicates.
-        afd-plugin targets vLLM v0.26.0 unless an issue explicitly says otherwise.
+        afd-plugin targets vLLM v0.30.0 unless an issue explicitly says otherwise.
 
   - type: markdown
     attributes:
@@ -76,5 +76,5 @@ body:
       options:
         - label: I searched existing issues for related reports.
           required: true
-        - label: I confirmed whether this reproduces against vLLM v0.26.0 or explained why not.
+        - label: I confirmed whether this reproduces against vLLM v0.30.0 or explained why not.
           required: true

.github/ISSUE_TEMPLATE/200-feature-request.yml

@@ -8,7 +8,7 @@ body:
     attributes:
       value: >
         Before submitting, please search existing issues and RFCs. Keep proposals
-        tied to vLLM v0.26.0 compatibility unless the issue explicitly proposes a
+        tied to vLLM v0.30.0 compatibility unless the issue explicitly proposes a
         version expansion.
 
   - type: textarea
@@ -31,7 +31,7 @@ body:
     id: compatibility
     attributes:
       label: vLLM compatibility and extension points
-      description: Explain how this should work without modifying the vLLM v0.26.0 source tree.
+      description: Explain how this should work without modifying the vLLM v0.30.0 source tree.
       placeholder: |
         Preferred extension point:
         Compat shim needed:

.github/PULL_REQUEST_TEMPLATE.md

@@ -42,7 +42,7 @@ or any behavior that intentionally differs from the original AFD commit.
 
 - [ ] Purpose is clear and linked to public context when possible.
 - [ ] Scope is bounded.
-- [ ] Compatibility with vLLM v0.26.0 is considered.
+- [ ] Compatibility with vLLM v0.30.0 is considered.
 - [ ] No changes are made to the vLLM source checkout.
 - [ ] Plugin-owned classes or explicit dotted class paths are preferred over monkey patches.
 - [ ] Any compat shim or monkey patch is isolated, idempotent, version-guarded, documented, and tested.

README.md

@@ -18,7 +18,8 @@ tests for GPU and Ascend NPU deployments.
 > This project is still experimental and needs more large-scale testing across
 > different hardware backends.
 
-The target runtime is **vLLM `v0.26.0`**. The plugin does not modify the vLLM
+The target runtime is **vLLM `v0.30.0`** at commit `ced6857a`; the Ascend
+integration targets vLLM-Ascend `8d4409d6`. The plugin does not modify the vLLM
 source tree. AFD behavior is installed through the `vllm.general_plugins` entry
 point, `--additional-config`, automatically selected role workers, plugin-owned
 model wrappers, and narrow version-scoped compatibility shims.
@@ -36,7 +37,8 @@ Core runtime support:
   execution for CUDA and Ascend NPU.
 - Eager and `FULL_DECODE_ONLY` graph execution, plus backend-specific profiling
   support.
-- Native DBO with exactly two ubatches on CUDA and the synchronous Ascend path.
+- Native DBO with exactly two ubatches on CUDA V1/V2 and the synchronous
+  Ascend V1 path; Ascend V2 DBO remains unsupported.
 - DeepSeek MoE handoff at the remote-experts boundary on CUDA, with the gate
   placed on either Attention or FFN.
 
@@ -46,7 +48,7 @@ Model support:
 | --- | --- | --- | --- |
 | DeepSeekV2 / DeepSeekV3 / DeepSeekV3.2 | `DeepseekForCausalLM`, `DeepseekV2ForCausalLM`, `DeepseekV3ForCausalLM`, `DeepseekV32ForCausalLM` | `AFDDeepseekForCausalLM`, `AFDDeepseekV2ForCausalLM`, `AFDDeepseekV3ForCausalLM` | DeepSeekV3.2 uses `AFDDeepseekV3ForCausalLM`. Each AFD role constructs and loads only its role-required model components, while shared embedding, normalization, and output components remain available where required by the model lifecycle. |
 | Qwen3 MoE | `Qwen3MoeForCausalLM` | `AFDQwen3MoeForCausalLM` | CUDA with `compute_gate_on_attention=false`. |
-| Qwen3.5 / Qwen3.6 MoE | `Qwen3_5MoeForConditionalGeneration` | `AFDQwen3_5MoeForConditionalGeneration` | Qwen3.5/Qwen3.6 adapter family. Repository CUDA E2E evidence currently covers text-only Qwen3.6-35B-A3B with `--language-model-only`, synchronous `P2pNcclAFDConnector`, native DP4/TP1/EP4 baseline, and AFD 2A1F eager/graph/graph+DBO. |
+| Qwen3.5 / Qwen3.6 MoE | `Qwen3_5MoeForConditionalGeneration` | `AFDQwen3_5MoeForConditionalGeneration` | CUDA text-only Qwen3.6-35B-A3B with `--language-model-only`, native DP4/TP1/EP4, and synchronous AFD 2A1F eager/graph. DBO remains excluded by a known FFN CUDA fault; see the [E2E limits](docs/design/module/e2e_testing.md). |
 
 Connector support:
 
@@ -55,17 +57,32 @@ See the [recipe index](recipe/README.md) for deployment and benchmark examples.
 | Connector | Platform | Recommend Stage | Sync or Async | Graph Support | Notes |
 | --- | --- | --- | --- | --- | --- |
 | `P2pNcclAFDConnector` | CUDA | Decode | Sync | `FULL_DECODE_ONLY` CUDA graph | FFN ranks are ordered before Attention ranks. `num_attention_ranks` must be greater than or equal to `num_ffn_ranks`; the Attention ranks are spread over the FFN ranks in blocks that differ in size by at most one. See the [DeepSeek V2 Lite recipe](recipe/gpu/P2pNcclAFDConnector/deepseek_v2_lite/README.md). |
-| `CAMP2pAFDConnector` | Ascend NPU | Decode | Sync | `FULL_DECODE_ONLY` ACL graph | Uses HCCL/CAMP2P custom ops. Ascend ops build by default on NPU platforms. See the [synchronous DeepSeek V3.2 recipe](recipe/npu/CAMP2pAFDConnector/deepseek_v3_2/README.md). |
-| `CAMAsyncAFDConnector` | Ascend NPU | Prefill / decode | Async | Not supported | Experimental v0.26 DP+TP/SP path with AFD-managed two-stage MoE ubatching; native DBO and PCP are unsupported. Post-fix DeepSeek-V3.2 DP2TP8+EP16 token split reached `0.9522` strict match on the complete GSM8K evaluation. The [legacy PCP8 recipe](recipe/npu/CAMAsyncAFDConnector/deepseek_v3_2/README.md) requires `release/v0.19.1rc1`. |
+| `CAMP2pAFDConnector` | Ascend NPU | Decode | Sync | V1 `FULL_DECODE_ONLY`; V2 also `FULL` | Uses HCCL/CAMP2P custom ops. Ascend ops build by default on NPU platforms. See the [synchronous DeepSeek V3.2 recipe](recipe/npu/CAMP2pAFDConnector/deepseek_v3_2/README.md). |
+| `CAMAsyncAFDConnector` | Ascend NPU | Prefill / decode | Async | Not supported | Experimental DP+TP/SP path with AFD-managed two-stage MoE ubatching; native DBO and PCP are unsupported. V2-Lite and DSV4 Flash W4A8 have [representative v0.30 validation](https://github.com/vllm-project/afd-plugin/pull/425#issuecomment-6063910923). The [legacy PCP8 recipe](recipe/npu/CAMAsyncAFDConnector/deepseek_v3_2/README.md) requires `release/v0.19.1rc1`. |
 
 Connector implementations are grouped by backend package:
 `afd_plugin.connectors.gpu` for GPU-only connectors,
 `afd_plugin.connectors.npu` for NPU-only connectors.
 
 Known gaps:
 
-- vLLM versions other than `0.26.0` are not claimed as supported.
-- vLLM/vLLM-Ascend model runner v2 is not supported.
+- The GPU target is exactly vLLM `0.30.0`; other vLLM versions are not
+  claimed as supported on GPU.
+- NPU hardware evidence covers representative DeepSeek-V2-Lite and DSV4
+  Flash W4A8 deployments; see the [runtime matrix](docs/design/module/execution_platforms.md#tested-runtime-matrix).
+  Other models and topologies remain unverified. Historical v0.26 results retain
+  their recorded vLLM-Ascend `80d8c194f` environment.
+- NPU MRV2 many-to-one padding and DSV2 Async SP with TP-sharded shared
+  experts have known correctness defects deferred to separate fixes; see the
+  [runtime matrix](docs/design/module/execution_platforms.md#tested-runtime-matrix).
+- Synchronous DSV4 E2E profiles are smoke cases; A5 concurrent answers remain
+  incorrect. See the [profile limits](tests/e2e/README.md#dsv4-flash-sync-camp2p-concurrent-requests-local-4-or-16-npus).
+- GPU ModelRunnerV2 supports synchronous `P2pNcclAFDConnector` with
+  `compute_gate_on_attention=false`, including two-microbatch DBO and
+  `FULL_DECODE_ONLY`. H20 validation covers DeepSeek-V2-Lite, 2A2F DP2/TP1;
+  see the [MRV2 DBO scenarios](tests/e2e/README.md#gpu-mrv2-dbo-development-comparison).
+  Representative 300-question comparisons do not establish numerical
+  equivalence; see the [runtime matrix](docs/design/module/execution_platforms.md#tested-runtime-matrix). NPU MRV2 DBO is unsupported.
 - GPU and NPU E2E tests are opt-in and require real hardware plus model weights.
 - GPU CUDA graph support is limited to `FULL_DECODE_ONLY`.
 - Native DBO is limited to exactly two ubatches and is not supported by
@@ -79,7 +96,7 @@ Known gaps:
   multi-node execution are unsupported; quantization is unverified; no
   performance claim is made.
 - PCP-based NPU model-runner-v1 deployments from v0.19.1rc1 are not supported
-  on v0.26.
+  on v0.30.
 
 ## Install
 
@@ -103,28 +120,35 @@ command:
 uv sync --group dev --extra vllm
 ```
 
-The optional extra pins `vllm==0.26.0`.
+The optional extra pins `vllm==0.30.0`.
 
 ### Ascend NPU installation
 
-AFD's Ascend path is validated on openEuler 22.03 (aarch64) with
-Ascend 910C / Atlas A3. Install a compatible driver and firmware, and confirm
-the devices with `npu-smi info`. Use this source baseline:
+The Ascend target uses the source pair below. Representative v0.30 validation
+uses Ascend 910C / Atlas A3; see the [recorded environment and results](https://github.com/vllm-project/afd-plugin/pull/425#issuecomment-6063910923).
+Install a compatible driver and firmware, and confirm the devices with
+`npu-smi info`.
 
 | Component | Version |
 | --- | --- |
 | Python | `3.10` or `3.11` |
-| vLLM | `0.26.0` |
-| vLLM-Ascend | commit [`80d8c194f`](https://github.com/vllm-project/vllm-ascend/commit/80d8c194f7584b17fe08065ea99a130916f6b0e7) |
+| vLLM | `0.30.0`, commit `ced6857afa0ea7b2e3f0846a62e1394e90f15607` |
+| vLLM-Ascend | commit [`8d4409d6`](https://github.com/vllm-project/vllm-ascend/commit/8d4409d6256d8a6729140ddcc0d1889e3f96cdd6) |
 | CANN / torch / torch-npu | Use the mutually compatible versions required by that vLLM-Ascend source snapshot. |
 
+> Hardware evidence is scoped to the recorded models, configurations, and
+> sample counts; it does not cover every NPU deployment or guarantee clean
+> device shutdown.
+
 #### Environment
 
-The v0.26 integration was refreshed against vLLM-Ascend commit `80d8c194f`;
-the repository does not currently claim a released v0.26 container tag. Use the
-[installation guide at that source snapshot](https://github.com/vllm-project/vllm-ascend/blob/80d8c194f7584b17fe08065ea99a130916f6b0e7/docs/source/installation.md)
+The Ascend environment targets vLLM-Ascend commit `8d4409d6` and vLLM
+`0.30.0`. The recorded A3 environment uses CANN 9.1.0, torch 2.10.0+cpu,
+and torch-npu 2.10.0.post4. The repository does not claim a released container
+tag for either target. Use the
+[installation guide at that source snapshot](https://github.com/vllm-project/vllm-ascend/blob/8d4409d6256d8a6729140ddcc0d1889e3f96cdd6/docs/source/getting_started/installation.md)
 to prepare a matching A3/openEuler environment, then install AFD from the
-repository root. Do not reuse the former v0.19.1rc1 image as a v0.26 runtime.
+repository root. Do not reuse the former v0.19.1rc1 image as a v0.30 runtime.
 
 #### Install AFD
 
@@ -177,8 +201,11 @@ or standard Ascend NPU platform. Explicit AFD worker paths remain accepted for
 compatibility with existing commands, but are not required or stable launch
 interfaces.
 
-GPU model runner v2 is not supported. Select model runner v1 before starting
-either GPU role:
+The examples below use model runner v1. For GPU MRV2, set
+`VLLM_USE_V2_MODEL_RUNNER=1` on both roles and use the
+[dedicated E2E scenarios](tests/e2e/README.md#gpu-mrv2-dbo-development-comparison).
+MRV2 DBO requires two microbatches and Attention DP > 1.
+For the MRV1 examples below, select:
 
 ```bash
 export VLLM_USE_V2_MODEL_RUNNER=0

afd_plugin/__init__.py

@@ -197,8 +197,17 @@ def register_afd() -> None:
             exc_info=True,
         )
 
-    # NPU compatibility patches are applied during AFD config construction and
-    # worker startup, after vLLM-Ascend completes its platform initialization.
+    # EngineCore constructs Ascend schedulers before handshake revalidation or
+    # AFD worker startup. Install the namespace factory at plugin registration
+    # so direct scheduler calls and already imported aliases accept AFD settings.
+    from vllm.platforms import current_platform
+
+    if current_platform.device_type == "npu":
+        from afd_plugin.compat.patches.npu.ascend_config import (
+            apply_afd_ascend_config_patch,
+        )
+
+        apply_afd_ascend_config_patch()
 
     from vllm.model_executor.models import ModelRegistry
 

afd_plugin/compat/__init__.py

@@ -6,12 +6,14 @@
     TARGET_VLLM_VERSION,
     assert_vllm_version_supported,
     get_installed_vllm_version,
+    is_target_vllm_compatible,
     is_vllm_version_supported,
 )
 
 __all__ = [
     "TARGET_VLLM_VERSION",
     "assert_vllm_version_supported",
     "get_installed_vllm_version",
+    "is_target_vllm_compatible",
     "is_vllm_version_supported",
 ]

afd_plugin/compat/npu/__init__.py

@@ -20,6 +20,7 @@
 )
 from afd_plugin.compat.npu.runtime import (
     apply_afd_ascend_config_patch_if_needed,
+    apply_afd_ascend_engine_core_config_patch_if_needed,
     apply_afd_ascend_patches_if_needed,
     apply_afd_async_dp_engine_patch_if_needed,
     ascend_forward_context,
@@ -30,6 +31,7 @@
 
 __all__ = [
     "apply_afd_ascend_config_patch_if_needed",
+    "apply_afd_ascend_engine_core_config_patch_if_needed",
     "apply_afd_ascend_patches_if_needed",
     "apply_afd_async_dp_engine_patch_if_needed",
     "ascend_forward_context",

afd_plugin/compat/npu/runtime.py

@@ -27,10 +27,14 @@
 def apply_afd_ascend_config_patch_if_needed() -> None:
     """Apply patches required while vLLM builds an AFD NPU config."""
 
+    from afd_plugin.compat.patches.npu.ascend_config import (
+        apply_afd_ascend_config_patch,
+    )
     from afd_plugin.compat.patches.npu.ascend_platform import (
         apply_afd_ascend_dbo_config_patch,
     )
 
+    apply_afd_ascend_config_patch()
     if not apply_afd_ascend_dbo_config_patch():
         raise RuntimeError(
             "AFD NPU DBO config patch requires vLLM-Ascend NPUPlatform",
@@ -59,6 +63,26 @@ def apply_afd_async_dp_engine_patch_if_needed(vllm_config: VllmConfig) -> bool:
     return True
 
 
+def apply_afd_ascend_engine_core_config_patch_if_needed(
+    vllm_config: VllmConfig,
+) -> bool:
+    """Prepare AFD sync Attention and FFN DP children for Ascend config."""
+    afd_config = parse_optional_afd_config(vllm_config, validate=False)
+    if afd_config is None or (
+        afd_config.role == "attention" and is_afd_async_dp(vllm_config)
+    ):
+        return False
+
+    from vllm.v1.engine.core import EngineCoreProc
+
+    from afd_plugin.compat.patches.npu.ascend_config import (
+        run_afd_ascend_engine_core,
+    )
+
+    EngineCoreProc.run_engine_core = staticmethod(run_afd_ascend_engine_core)
+    return True
+
+
 def apply_afd_ascend_patches_if_needed() -> None:
     """Apply plugin-owned runtime patches after Ascend initialization."""
 
@@ -80,6 +104,7 @@ def apply_afd_ascend_patches_if_needed() -> None:
 
 __all__ = [
     "apply_afd_async_dp_engine_patch_if_needed",
+    "apply_afd_ascend_engine_core_config_patch_if_needed",
     "apply_afd_ascend_config_patch_if_needed",
     "apply_afd_ascend_patches_if_needed",
     "ascend_forward_context",

afd_plugin/compat/npu/runtime_config.py

@@ -10,6 +10,9 @@
     from vllm.config import VllmConfig
 
 FLASHINFER_ALL2ALLV_BACKEND = "flashinfer_all2allv"
+DEEPSEEK_V2_MODEL_TYPES = frozenset(
+    ("deepseek", "deepseek_v2", "deepseek_v3", "deepseek_v32", "glm_moe_dsa")
+)
 
 
 def npu_afd_num_ubatches(vllm_config: VllmConfig) -> int:
@@ -19,27 +22,49 @@ def npu_afd_num_ubatches(vllm_config: VllmConfig) -> int:
     return 1
 
 
+def npu_model_uses_sequence_parallel_moe(vllm_config: VllmConfig) -> bool:
+    """Match the DSV2 decoder's effective SP setting under pipeline parallelism."""
+    parallel_config = vllm_config.parallel_config
+    model_type = vllm_config.model_config.hf_text_config.model_type
+    return parallel_config.use_sequence_parallel_moe and (
+        model_type not in DEEPSEEK_V2_MODEL_TYPES
+        or parallel_config.pipeline_parallel_size == 1
+    )
+
+
+def npu_model_uses_sharded_pp_tensors(vllm_config: VllmConfig) -> bool:
+    """Return the model's PP wire layout independently of its internal SP.
+
+    Target DSV4 gathers the complete sequence before each PP boundary. Target
+    DSV2 disables MoE SP when PP is enabled; its PP tensors are global too.
+    """
+    return (
+        npu_model_uses_sequence_parallel_moe(vllm_config)
+        and vllm_config.model_config.hf_text_config.model_type != "deepseek_v4"
+    )
+
+
 def fix_all2all_backend_for_afd(vllm_config: VllmConfig) -> None:
-    """Apply vLLM-Ascend's default-worker all2all fix to AFD workers.
-
-    vLLM-Ascend normally rewrites ``all2all_backend`` to
-    ``"flashinfer_all2allv"`` when sequence parallelism is disabled, but that
-    compatibility rewrite is gated on the default ``worker_cls == "auto"``.
-    Automatic AFD worker selection preserves that upstream rewrite. Legacy
-    commands that explicitly select an AFD worker miss it and can keep the
-    default ``"allgather_reducescatter"`` backend.
-
-    Leaving that backend in place can make the Ascend MoE path think sequence
-    parallel MoE is enabled and split tokens through ``sequence_parallel_chunk``,
-    which is not the layout AFD's NPU connector path sends. Mirror the upstream
-    rewrite here before AFD creates the NPU model runner.
+    """Keep configured model SP while preserving the non-SP worker fallback.
+
+    Target Ascend derives SP from ParallelConfig.use_sequence_parallel_moe.
+    AscendConfig first applies the effective FlashComm config switch to the
+    backend. Preserve that result, including for explicit AFD worker classes.
+    Rewriting its all2all backend disables model-owned token sharding even
+    when EP/DP/TP requested it. Non-SP workers keep the legacy backend fix.
     """
     parallel_config = vllm_config.parallel_config
     if (
-        not vllm_config.compilation_config.pass_config.enable_sp
+        not parallel_config.use_sequence_parallel_moe
+        and not vllm_config.compilation_config.pass_config.enable_sp
         and parallel_config.all2all_backend != FLASHINFER_ALL2ALLV_BACKEND
     ):
         parallel_config.all2all_backend = FLASHINFER_ALL2ALLV_BACKEND
 
 
-__all__ = ["fix_all2all_backend_for_afd", "npu_afd_num_ubatches"]
+__all__ = [
+    "fix_all2all_backend_for_afd",
+    "npu_afd_num_ubatches",
+    "npu_model_uses_sequence_parallel_moe",
+    "npu_model_uses_sharded_pp_tensors",
+]