[CI][0.9.0][0.9.1] Update CI job (#1149)

1. drop main and add 0.9.1 check  for 0.9.1-dev branch

2. cherry-pick
b75cb78
to fix import error and make 0.9.1 works

3. Fix quantization test failure.

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

Author: wangxiyuan

.github/dependabot.yml

@@ -2,9 +2,6 @@ version: 2
 updates:
   - package-ecosystem: "github-actions"
     directory: "/"
-    schedule:
-      # Check for updates to GitHub Actions every week
-      interval: "weekly"
     open-pull-requests-limit: 2
     reviewers:
       - "Yikun"

.github/workflows/nightly_benchmarks.yaml

@@ -18,11 +18,7 @@
 name: 'Benchmarks / Performance'
 # This workflow runs nightly benchmarks for vllm-ascend.
 
-on:
-  schedule:
-    # Run at 02:00 everyday
-    - cron: '00 18 * * *'
-  
+on:  
   workflow_dispatch:
     # Allow manual triggering of the workflow
 

.github/workflows/vllm_ascend_doctest.yaml

@@ -29,9 +29,6 @@ on:
       - 'tests/e2e/doctests/**'
       - 'tests/e2e/common.sh'
       - 'tests/e2e/run_doctests.sh'
-  schedule:
-    # Runs every 4 hours
-    - cron:  '0 */4 * * *'
 
 # Bash shells do not use ~/.profile or ~/.bashrc so these shells need to be explicitly
 # declared as "shell: bash -el {0}" on steps that need to be properly activated.

.github/workflows/vllm_ascend_test.yaml

@@ -18,8 +18,6 @@
 name: 'e2e test'
 
 on:
-  schedule:
-    - cron: '0 23 * * *'
   pull_request:
     branches:
       - 'main'
@@ -44,7 +42,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-arm64-npu-1, linux-arm64-npu-4]
-        vllm_version: [main, v0.9.0]
+        vllm_version: [v0.9.0, v0.9.1]
     concurrency:
       group: >
         ${{

.github/workflows/vllm_ascend_test_long_term.yaml

@@ -17,9 +17,6 @@
 name: 'e2e test / long-term-test'
 
 on:
-  schedule:
-    # Runs at 23:00 UTC (7:00 AM Beijing) every day
-    - cron: '0 23 * * *'
   pull_request:
     types: [ labeled ]
 
@@ -43,7 +40,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-arm64-npu-1, linux-arm64-npu-4]
-        vllm_version: [main, v0.9.0]
+        vllm_version: [v0.9.0, v0.9.1]
     name: vLLM Ascend long term test
     runs-on: ${{ matrix.os }}
     container:

.github/workflows/vllm_ascend_test_pd.yaml

@@ -17,9 +17,6 @@
 name: 'e2e test / pd-disaggregation'
 
 on:
-  schedule:
-    # Runs at 23:00 UTC (7:00 AM Beijing) every day
-    - cron: '0 23 * * *'
   pull_request:
     types: [ labeled ]
 
@@ -41,7 +38,7 @@ jobs:
     if: ${{ contains(github.event.pull_request.labels.*.name, 'pd-test') && contains(github.event.pull_request.labels.*.name, 'ready-for-test') || github.event_name == 'schedule' }}
     strategy:
       matrix:
-        vllm_verison: [main, v0.9.0]
+        vllm_verison: [v0.9.0, v0.9.1]
     name: vLLM Ascend prefilling decoding disaggregation test
     runs-on: linux-arm64-npu-static-8
 

tests/conftest.py

@@ -78,6 +78,7 @@ def __init__(
         enable_chunked_prefill: bool = False,
         swap_space: int = 4,
         enforce_eager: Optional[bool] = True,
+        quantization: Optional[str] = None,
         **kwargs,
     ) -> None:
         self.model = LLM(
@@ -94,6 +95,7 @@ def __init__(
             max_model_len=max_model_len,
             block_size=block_size,
             enable_chunked_prefill=enable_chunked_prefill,
+            quantization=quantization,
             **kwargs,
         )
 

tests/singlecard/test_offline_inference.py

@@ -25,6 +25,7 @@
 
 import pytest
 import vllm  # noqa: F401
+from modelscope import snapshot_download  # type: ignore[import-untyped]
 from vllm import SamplingParams
 from vllm.assets.image import ImageAsset
 
@@ -33,12 +34,15 @@
 
 MODELS = [
     "Qwen/Qwen2.5-0.5B-Instruct",
-    "vllm-ascend/Qwen2.5-0.5B-Instruct-w8a8",
     "Qwen/Qwen3-0.6B-Base",
 ]
 MULTIMODALITY_MODELS = ["Qwen/Qwen2.5-VL-3B-Instruct"]
 os.environ["PYTORCH_NPU_ALLOC_CONF"] = "max_split_size_mb:256"
 
+QUANTIZATION_MODELS = [
+    "vllm-ascend/Qwen2.5-0.5B-Instruct-W8A8",
+]
+
 
 @pytest.mark.parametrize("model", MODELS)
 @pytest.mark.parametrize("dtype", ["half", "float16"])
@@ -59,6 +63,27 @@ def test_models(model: str, dtype: str, max_tokens: int) -> None:
         vllm_model.generate_greedy(example_prompts, max_tokens)
 
 
+@pytest.mark.parametrize("model", QUANTIZATION_MODELS)
+@pytest.mark.parametrize("max_tokens", [5])
+def test_quantization_models(model: str, max_tokens: int) -> None:
+    prompt = "The following numbers of the sequence " + ", ".join(
+        str(i) for i in range(1024)) + " are:"
+    example_prompts = [prompt]
+
+    # NOTE: Using quantized model repo id from modelscope encounters an issue,
+    # this pr (https://github.com/vllm-project/vllm/pull/19212) fix the issue,
+    # after it is being merged, there's no need to download model explicitly.
+    model_path = snapshot_download(model)
+
+    with VllmRunner(model_path,
+                    max_model_len=8192,
+                    enforce_eager=True,
+                    dtype="auto",
+                    gpu_memory_utilization=0.7,
+                    quantization="ascend") as vllm_model:
+        vllm_model.generate_greedy(example_prompts, max_tokens)
+
+
 @pytest.mark.parametrize("model", MULTIMODALITY_MODELS)
 def test_multimodal(model, prompt_template, vllm_runner):
     image = ImageAsset("cherry_blossom") \

vllm_ascend/compilation/__init__.py

@@ -496,7 +496,11 @@ def _update_states(self, scheduler_output: "SchedulerOutput") -> None:
             # Update the block IDs.
             if not req_data.resumed_from_preemption:
                 # Append the new blocks to the existing block IDs.
-                req_state.block_ids.extend(req_data.new_block_ids)
+                for block_ids, new_block_ids in zip(  # type: ignore[call-overload]
+                        req_state.block_ids,
+                        req_data.new_block_ids,
+                        strict=True):
+                    block_ids.extend(new_block_ids)
             else:
                 # The request is resumed from preemption.
                 # Replace the existing block IDs with the new ones.