[v0.21.0] Fix accuracy issue in minimax_m2 with TP > 1#1505

Closed

skavulya wants to merge 29 commits into

vllm-project:releases/v0.21.0from

skavulya:skavulya/minimax2_accuracy

Commits on May 15, 2026

Commits on May 18, 2026

fix: HPU-specific bug fixes for KV-offload + async spec-decode (#1264) (#1401)

authored
[DOC] Fix torchaudio version (#1425)
yangulei
authored
Harden Qwen3.5 CI test to detect regressions (#1443)

authored
Fix decode bucket filter issues from #1122 (#1447)
yangulei
authored
Fix mamba_type comparison for GDN hybrid cache allococation (#1449)
shepark
authored
fix: replace batched_count_greater_than to avoid dynamic shape TypeError on HPU (#1412)
kamil-kaczor
authored
fix: bypass _forward_impl for dp_size==1 to fix DeepSeek R1 FP8 crash (#1441)

authored
Remove num_ctx_tokens_less_or_equal_batched_max_model_len filter (#1454)

yangulei
and
kamil-kaczor
authored

Commits on May 19, 2026

Commits on May 20, 2026

Removal of ray and redundant transformers packages from gaudi requirements (#1445)
iboiko-habana
authored

Commits on May 21, 2026

Commits on May 22, 2026

Commits on May 25, 2026