[CI] Upgrade vLLM to 20250920 (c60e613) and address config break (#3067)

### What this PR does / why we need it?
Bump main to
c60e6137f0

- Updated imports in `vllm.config` to
`vllm.config.model`(aed16879a9)
https://github.com/vllm-project/vllm/pull/25252

- Refactored `vllm_ascend/sample/sampler.py` to use string values for
`logprobs_mode` instead of the `LogprobsMode` enum, simplifying logprobs
mode handling and improving compatibility with recent vLLM changes
(aed16879a9)
https://github.com/vllm-project/vllm/pull/25252

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI passed


- vLLM version: v0.10.2
- vLLM main:
6d8246aaff

---------

Signed-off-by: Yikun Jiang <yikunkero@gmail.com>

.github/workflows/format_pr_body.yaml

@@ -36,7 +36,7 @@ jobs:
 
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=6d8246aaffff3ebec84767e373212a7b8da328e2
+          VLLM_COMMIT=c60e6137f0bf2034853919b3a9d705d7e06b93cf
           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: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
+        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
     steps:
       - name: Install packages
         run: |
@@ -140,7 +140,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-1]
-        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
+        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
     name: singlecard e2e test - light
     runs-on: ${{ matrix.os }}
     container:
@@ -206,7 +206,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-2]
-        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
+        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
     name: multicard e2e test - light
     runs-on: ${{ matrix.os }}
     container:

.github/workflows/vllm_ascend_test_full.yaml

@@ -72,7 +72,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-1]
-        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
+        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
     name: singlecard e2e test - full
     runs-on: ${{ matrix.os }}
     container:
@@ -156,7 +156,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-2]
-        vllm_version: [6d8246aaffff3ebec84767e373212a7b8da328e2, v0.10.2]
+        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
     name: multicard e2e test - full
     runs-on: ${{ matrix.os }}
     container:

tests/e2e/conftest.py

@@ -32,7 +32,14 @@ from transformers import (AutoConfig, AutoModelForCausalLM, AutoTokenizer,
                           BatchEncoding, BatchFeature)
 from transformers.models.auto.auto_factory import _BaseAutoModelClass
 from vllm import LLM, SamplingParams
-from vllm.config import TaskOption, _get_and_verify_dtype
+
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.10.2"):
+    from vllm.config import TaskOption, _get_and_verify_dtype
+else:
+    from vllm.config.model import TaskOption, _get_and_verify_dtype
+
 from vllm.inputs import TextPrompt
 from vllm.outputs import RequestOutput
 from vllm.transformers_utils.utils import maybe_model_redirect

vllm_ascend/sample/sampler.py

@@ -1,12 +1,15 @@
 import torch
 import torch_npu
-from vllm.config import LogprobsMode
 from vllm.v1.sample.ops.topk_topp_sampler import TopKTopPSampler, random_sample
 from vllm.v1.sample.sampler import Sampler
 
-from vllm_ascend.utils import is_310p
+from vllm_ascend.utils import is_310p, vllm_version_is
 
-DEFAULT_LOGPROBS_MODE = LogprobsMode.RAW_LOGPROBS
+if vllm_version_is("0.10.2"):
+    from vllm.config import LogprobsMode
+    DEFAULT_LOGPROBS_MODE = LogprobsMode.RAW_LOGPROBS
+else:
+    DEFAULT_LOGPROBS_MODE = "raw_logprobs"
 
 
 class AscendSampler(Sampler):
@@ -65,10 +68,18 @@ class AscendTopKTopPSampler(TopKTopPSampler):
         """Override pytorch native implementation to torch_npu"""
         logits = self._apply_top_k_top_p(logits, k, p)
         logits_to_return = None
-        if self.logprobs_mode == LogprobsMode.PROCESSED_LOGITS:
-            logits_to_return = logits
-        elif self.logprobs_mode == LogprobsMode.PROCESSED_LOGPROBS:
-            logits_to_return = logits.log_softmax(dim=-1, dtype=torch.float32)
+        if vllm_version_is("0.10.2"):
+            if self.logprobs_mode == LogprobsMode.PROCESSED_LOGITS:
+                logits_to_return = logits
+            elif self.logprobs_mode == LogprobsMode.PROCESSED_LOGPROBS:
+                logits_to_return = logits.log_softmax(dim=-1,
+                                                      dtype=torch.float32)
+        else:
+            if self.logprobs_mode == "processed_logits":
+                logits_to_return = logits
+            elif self.logprobs_mode == "processed_logprobs":
+                logits_to_return = logits.log_softmax(dim=-1,
+                                                      dtype=torch.float32)
 
         probs = logits.softmax(dim=-1, dtype=torch.float32)
         return random_sample(probs, generators), logits_to_return