[Misc]feat: adapt to vLLM main (5af684c3...6e6d182d)

[Meihan-chen]

Automated adaptation to upstream vLLM main branch changes.
Commit range: 5af684c31912232e5c89484c2e8259e0fac6c55b...6e6d182d18dbe5c81c9c9058a5571241f649ae7b

phase: detect

round: 0

git_commits:

• sha: ccd7922c056d8ee443bb763b56698633ec1ff1b4
  subject: fix: Adapt to upstream vLLM changes (5af684c3..6e6d182d1)
  body:

  Changes Adapted

  P0 - Breaking Changes

  1. Platform Interface - New is_integrated_gpu method

     • Added is_integrated_gpu() method to NPUPlatform class
     • Returns False as Ascend NPUs are discrete devices, not integrated GPUs

  2. CacheConfig - Removed cpu_kvcache_space_bytes field

     • Removed code that checked/reset this field since it was removed from upstream

  P4 - Configuration Updates

  3. Updated vLLM commit references from 5af684c3 to 6e6d182d1
     • docs/source/conf.py
     • .github/workflows/pr_test_full.yaml
     • .github/workflows/pr_test_light.yaml
     • .github/workflows/dockerfiles/Dockerfile.lint

  Files Changed

  • vllm_ascend/platform.py: Added is_integrated_gpu method, removed cpu_kvcache_space_bytes handling
  • docs/source/conf.py: Updated main_vllm_commit reference
  • .github/workflows/pr_test_full.yaml: Updated vllm_version matrix
  • .github/workflows/pr_test_light.yaml: Updated vllm_version references
  • .github/workflows/dockerfiles/Dockerfile.lint: Updated VLLM_COMMIT

  Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com
  Signed-off-by: main2main-bot main2main-bot@users.noreply.github.com

phase: fix

round: 1

git_commits:

• sha: d7ba05e9dfd8c0bfd1fd7bc76a37a4746e074dfe
  subject: fix: Adapt to MoE runner rename (DefaultMoERunner -> MoERunner)
  body:

  Error Fixed

  ModuleNotFoundError: No module named 'vllm.model_executor.layers.fused_moe.runner.default_moe_runner'

  Root Cause

  Upstream commit b55b26520 "[MoE] Make MoERunnerInterface a PluggableLayer for OOT support (#35178)"
  renamed DefaultMoERunner to MoERunner and moved it from default_moe_runner.py to moe_runner.py.

  Files deleted:

  • vllm/model_executor/layers/fused_moe/runner/default_moe_runner.py
  • vllm/model_executor/layers/fused_moe/runner/moe_runner_base.py

  New structure:

  • MoERunner class now in moe_runner.py, inheriting from MoERunnerInterface

  Fix Applied

  Added vllm_version_is() guard to import MoERunner from the correct path:

  • v0.19.0 (pinned): DefaultMoERunner from default_moe_runner.py
  • main branch: MoERunner from moe_runner.py

  Changed class inheritance from DefaultMoERunner to MoERunnerBase (aliased).

  Files Changed

  • vllm_ascend/ops/fused_moe/fused_moe.py

  Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com
  Signed-off-by: main2main-bot main2main-bot@users.noreply.github.com

phase: bisect

round: 1

git_commits:

• sha: ba21d8e2a7fa72256a7c3a38fda24d71228408c4
  subject: fix: Adapt to CompilationTimes return type change in compile_or_warm_up_model
  body:

  Error Fixed

  TypeError caused by return type mismatch in compile_or_warm_up_model method.
  Upstream executor expects CompilationTimes NamedTuple but Ascend worker returned float.

  Root Cause

  Upstream commit c08f3b2a62 "Measure encoder compile time seperate from llm backbone (#39240)"
  changed compile_or_warm_up_model return type from float to CompilationTimes NamedTuple.

  Changes in upstream:

  • Added CompilationTimes NamedTuple with fields: language_model, encoder
  • WorkerBase.compile_or_warm_up_model now returns CompilationTimes
  • Executor.collective_rpc expects list[CompilationTimes] instead of list[float]

  Fix Applied

  Added version-guarded import and return logic:

  • v0.19.0 (pinned): Import not needed, return float directly
  • main branch: Import CompilationTimes, return CompilationTimes(language_model=..., encoder=0.0)

  Files Changed

  • vllm_ascend/worker/worker.py: Added vllm_version_is import, CompilationTimes import, updated return logic

  Bisect Result

  First bad commit: c08f3b2a62fd8ddcb644ae91697a7981f2db4097
  Test: tests/e2e/multicard/2-cards/test_offline_inference_distributed.py::test_deepseek3_2_w8a8_pruning_mtp_tp2_ep

  Co-Authored-By: Claude Opus 4.7 noreply@anthropic.com
  Signed-off-by: main2main-bot main2main-bot@users.noreply.github.com

• vLLM version: v0.19.0

• vLLM main: vllm-project/vllm@5af684c