DO NOT MERGE - CI sandbox for stateless scheduler b#25172

Open

fzyzcjy wants to merge 46 commits into

sgl-project:mainfrom

fzyzcjy:feat/stateless_scheduler_b

Commits on May 14, 2026

Prep: abort_request dedup for chunked-resume dual-queue holding
fzyzcjy
committed
Prep: subtract prefix_indices from waiting_queue pending tokens sum
fzyzcjy
committed
Prep: document filter_batch chunked-exclusion invariant
fzyzcjy
committed
Bound cache_unfinished_req row read by kv_committed_len
fzyzcjy
committed
Drop is_chunked from req_to_token_pool alloc assert
fzyzcjy
committed
Unify chunked admission via add_one_req reuse branch + add has_pending_chunk
fzyzcjy
committed
Switch chunked-resume to waiting_queue holding; delete chunked_req fields
fzyzcjy
committed
refactor: rename Req.is_chunked -> Req.pending_middle_outputs
fzyzcjy
committed
Fix retract_all passing List[Req] to filter_batch as keep_indices
fzyzcjy
committed
Refactor filter_batch to use explicit exclude_chunked_req flag
fzyzcjy
committed
Tighten add_one_req reuse gate to has_pending_chunk
fzyzcjy
committed
Reset host_hit_length unconditionally in prepare_for_extend
fzyzcjy
committed
Skip chunked-resume reqs in calc_priority prefix matching
fzyzcjy
committed
Skip chunked-resume reqs in _abort_on_waiting_timeout
fzyzcjy
committed
Bypass LoRA scheduling gate for chunked-resume reqs
fzyzcjy
committed
Skip mamba_pool_idx cleanup for chunked-resume on NO_TOKEN
fzyzcjy
committed
Widen merge_batch assert to match filter_batch predicate
fzyzcjy
committed
Bound streaming-session chunked stash by kv_committed_len
fzyzcjy
committed
Release row + KV + lock_ref when aborting a chunked-resume req from waiting_queue
fzyzcjy
committed
Give chunked-resume reqs priority in LPM and DFS_WEIGHT sorts
fzyzcjy
committed
Extend pause(retract) to waiting chunked-resume reqs
fzyzcjy
committed
Reset disagg send-side state on chunked-resume retract
fzyzcjy
committed
Count chunked-resume tail in runtime mem check (page_size > 1)
fzyzcjy
committed
Document filter_batch(exclude_chunked_req=True) at every call site
fzyzcjy
committed
Include PP microbatch reqs in abort_request batch_rids dedup
fzyzcjy
committed
Document why Stage A chunk-stash runs at iter boundary instead of end-of-prior-iter
fzyzcjy
committed
Filter chunked-resume reqs from split_prefill_batch before pdmux merge
fzyzcjy
committed
Apply black-jupyter formatting (CI lint fixup)
fzyzcjy
committed
Remove v1 SWA chunked-req stash gate test (gate was deleted in v2)
fzyzcjy
committed
Drop v1 has_chunked_req kwarg + delete v1 add_chunked_req SWA tests
fzyzcjy
committed
Guard _handle_finished_req against PP cross-microbatch double-finalize
fzyzcjy
committed
Seed has_pending_chunk/is_dllm/host_hit_length on test_prefill_adder mocks
fzyzcjy
committed
Re-add ScheduleBatch.chunked_req marker for PP cross-mb filter exclusion
fzyzcjy
committed
Revert "Re-add ScheduleBatch.chunked_req marker for PP cross-mb filter exclusion"
fzyzcjy
committed
Bump pending_middle_outputs for last-chunk admits + decrement-first output proc
fzyzcjy
committed
Revert "Bump pending_middle_outputs for last-chunk admits + decrement-first output proc"
fzyzcjy
committed
Exclude in-flight other-mb reqs in filter_batch (PP chunked-resume race)
fzyzcjy
committed
Revert "Exclude in-flight other-mb reqs in filter_batch (PP chunked-resume race)"
fzyzcjy
committed
Conditionally exclude in-flight other-mb chunked-resume reqs (PP, max_new_tokens > 1)
fzyzcjy
committed

Commits on May 20, 2026

Merge branch 'main' into feat/stateless_scheduler_b
fzyzcjy
authored

Commits on May 25, 2026

Merge branch 'main' into feat/stateless_scheduler_b
fzyzcjy
committed