[CI] Main2main upgrade to 0324 #7787

Merged

MengqingCao merged 21 commits into

vllm-project:mainfrom

leo-pony:main2main_0324_qz

Apr 3, 2026

Commits on Apr 2, 2026

upgrade to 0324

22dimensions
authored and
leo-pony
committed
fix: add vllm_is_batch_invariant compatibility wrapper

claude
authored and
leo-pony
committed
fix issued caused by [Async][Spec Decoding]

authored and
leo-pony
committed
Fix kv_cache and batch inference issues and Fix MLA kv_cache IndexError and kv_offload import
leo-pony
committed
fix rebase also use batch_invariant from vllm
leo-pony
committed
ci format fix
leo-pony
committed
Fix pooling/embedding model hang by computing optimistic_seq_lens_cpu
leo-pony
committed
Fix arange_np corruption in _dummy_run causing IndexError on second inference

leo-pony
and
claude
committed
set continue on error
leo-pony
committed
Fix eagle_proposer accessing .gpu/.cpu on plain tensor attributes
leo-pony
committed
Disabled async_scheduling when speculative decoding is active on Ascend, since the GPU-side num_computed_tokens correction needed for async spec decode is not implemented.
leo-pony
committed
ci format fix
leo-pony
committed
seq_lens 和 positions 始终是 plain GPU tensor，所有 hasattr(self.seq_lens, 'np'/'cpu'/'gpu') 和 hasattr(self.positions, 'np'/'copy_to_gpu') 的分支判断都是多余的 — 永远走 else 分支。已全部简化为直接使用 plain tensor API。
leo-pony
committed
1. seq_lens、positions — plain GPU tensor，删除所有 guards
leo-pony
committed
Cancel unnecessary changes in model_runner_v1.py
Your Name
committed
Fix model_runner_v1.py and cancel changes in attention_v1.py
wangbj127
committed
Optmize the unneeded dtype translate
leo-pony
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[CI] Main2main upgrade to 0324 #7787

[CI] Main2main upgrade to 0324 #7787
MengqingCao merged 21 commits into
vllm-project:mainfrom
leo-pony:main2main_0324_qz

Commits on Apr 2, 2026

upgrade to 0324

fix: add vllm_is_batch_invariant compatibility wrapper

fix issued caused by [Async][Spec Decoding]

Fix kv_cache and batch inference issues and Fix MLA kv_cache IndexError and kv_offload import

fix rebase also use batch_invariant from vllm

ci format fix

Fix pooling/embedding model hang by computing optimistic_seq_lens_cpu

Fix arange_np corruption in _dummy_run causing IndexError on second inference

set continue on error

Fix eagle_proposer accessing .gpu/.cpu on plain tensor attributes

Disabled async_scheduling when speculative decoding is active on Ascend, since the GPU-side num_computed_tokens correction needed for async spec decode is not implemented.

ci format fix

seq_lens 和 positions 始终是 plain GPU tensor，所有 hasattr(self.seq_lens, 'np'/'cpu'/'gpu') 和 hasattr(self.positions, 'np'/'copy_to_gpu') 的分支判断都是多余的 — 永远走 else 分支。已全部简化为直接使用 plain tensor API。

1. seq_lens、positions — plain GPU tensor，删除所有 guards

Cancel unnecessary changes in model_runner_v1.py

Fix model_runner_v1.py and cancel changes in attention_v1.py

Optmize the unneeded dtype translate

Commits on Apr 3, 2026

Replace vllm_is_batch_invariant with envs, to keep consist with upstream, see vllm pr:35007

fix 310P _prepare_input_ids less num_reqs param, break import in vllm PR:32951

Skip the 310P qwen3.5 test case and make test case more stable

Optimize the performance: remove seqs_len device to host