[Tracking] Upstream backlog and merge plan

[krystophny]

This issue tracks the remaining upstream work in waybarrios/vllm-mlx and the order to close it out cleanly from our fork.

Current Upstream PRs (updated 2026-03-26)

Ready to merge (approved / strong +1)

• tokenizer: return successful mlx-lm load result waybarrios/vllm-mlx#215 tokenizer: return successful mlx-lm load result -- APPROVED
• engine: keep SimpleEngine serialized across cancellation waybarrios/vllm-mlx#220 engine: keep SimpleEngine serialized across cancellation -- strong +1, production-validated
• fix: bump mlx-lm minimum to 0.31.0 for hybrid model batching waybarrios/vllm-mlx#227 fix: bump mlx-lm minimum to 0.31.0 for hybrid model batching -- NEW, fixes [Bug] Engine loop error: ArraysCache.__init__() missing 1 required positional argument: 'size' when enabling --continuous-batching or running bench #11

Clean, awaiting review

• cli: expose harmony and gpt-oss tool parsers waybarrios/vllm-mlx#216 cli: expose harmony and gpt-oss tool parsers
• chat: forward chat_template_kwargs on simple-engine paths waybarrios/vllm-mlx#218 chat: forward chat_template_kwargs on simple-engine paths
• simple-engine: keep tool chat on the streaming execution path waybarrios/vllm-mlx#222 simple-engine: keep tool chat on the streaming execution path
• test: make Python 3.13 async suite pass and cover it in CI waybarrios/vllm-mlx#226 test: make Python 3.13 async suite pass and cover it in CI

Previously draft, now ready for review

• scheduler: preserve prompt checkpoints in chunked prefill resume path waybarrios/vllm-mlx#221 scheduler: preserve prompt checkpoints in chunked prefill resume path -- checkpoint callback now wired
• prefix_cache: preserve hybrid recurrent state across blocks waybarrios/vllm-mlx#217 prefix_cache: preserve hybrid recurrent state across blocks -- deduplication safety test added, duplicate tokenizer hunk removed

Largest scope

• server: add OpenAI-compatible /v1/responses endpoint waybarrios/vllm-mlx#214 server: add OpenAI-compatible /v1/responses endpoint

Fork main state

origin/main is 21 commits ahead of upstream/main, containing cherry-picks of all 10 PRs above. All 129 relevant tests pass. All PR branches are 0 behind upstream/main and mergeable.

Recommended merge order

1. fix: bump mlx-lm minimum to 0.31.0 for hybrid model batching waybarrios/vllm-mlx#227 (version bump, unblocks hybrid model users)
2. tokenizer: return successful mlx-lm load result waybarrios/vllm-mlx#215 (approved, small)
3. engine: keep SimpleEngine serialized across cancellation waybarrios/vllm-mlx#220 (production-validated)
4. cli: expose harmony and gpt-oss tool parsers waybarrios/vllm-mlx#216, chat: forward chat_template_kwargs on simple-engine paths waybarrios/vllm-mlx#218, simple-engine: keep tool chat on the streaming execution path waybarrios/vllm-mlx#222 (independent fixes)
5. scheduler: preserve prompt checkpoints in chunked prefill resume path waybarrios/vllm-mlx#221 (scheduler correctness)
6. prefix_cache: preserve hybrid recurrent state across blocks waybarrios/vllm-mlx#217 (cache correctness)
7. test: make Python 3.13 async suite pass and cover it in CI waybarrios/vllm-mlx#226 (test infrastructure)
8. server: add OpenAI-compatible /v1/responses endpoint waybarrios/vllm-mlx#214 (feature, largest)