Support intranode DeepEP without IMEX fabric - #7

Draft

rchalamala wants to merge 59 commits into

mainfrom

cursor/no-imex-deepep-f10b

Commits on Jul 23, 2026

Commits on Jul 24, 2026

test: cover CUDA graph dependency edge metadata

cursoragent
and
rchalamala
committed
test: link edge metadata helper to CUDA driver

cursoragent
and
rchalamala
committed
fix: preserve CUDA graph dependency edge metadata

cursoragent
and
rchalamala
committed
test: use canonical graph archive names

cursoragent
and
rchalamala
committed
test: compare edge replay output on CPU

cursoragent
and
rchalamala
committed
test: isolate legacy parallel graph fixture

cursoragent
and
rchalamala
committed
style: format edge metadata CUDA helper

cursoragent
and
rchalamala
committed
test: cover mixed CUDA graph edge metadata

cursoragent
and
rchalamala
committed
fix: batch CUDA dependencies by edge metadata

cursoragent
and
rchalamala
committed
style: format dependency batching helper

cursoragent
and
rchalamala
committed
test: cover SGLang LOAD logits shard pre-copy

cursoragent
and
rchalamala
committed
fix: pre-copy SGLang LOAD local logits shard

cursoragent
and
rchalamala
committed
test: cover SGLang regular TP logits shard pre-copy

cursoragent
and
rchalamala
committed
test: complete regular TP pre-copy coverage

cursoragent
and
rchalamala
committed
fix: pre-copy SGLang regular TP local shard

cursoragent
and
rchalamala
committed
test: complete fake SGLang gather modules

cursoragent
and
rchalamala
committed
test: cover optional SGLang gather APIs

cursoragent
and
rchalamala
committed
fix: tolerate optional SGLang gather APIs

cursoragent
and
rchalamala
committed
test: cover final no-IMEX review regressions

cursoragent
and
rchalamala
committed
fix: delegate unsupported SGLang regular-TP gather layouts

cursoragent
and
rchalamala
committed
fix: validate VMM IPC chunk-import range before interior pointer

cursoragent
and
rchalamala
committed
chore: harden experimental recipes and pin deterministic defaults

cursoragent
and
rchalamala
committed
docs: document binary dependency record widths and compatibility

cursoragent
and
rchalamala
committed
test: scope VMM IPC regression to invalid-range rejection

cursoragent
and
rchalamala
committed
test: cover cached-chunk, stale-generation, and deepep-mode rereview findings

cursoragent
and
rchalamala
committed
fix: validate cached VMM IPC chunk carve against actual mapping

cursoragent
and
rchalamala
committed
fix: verify chunk generation against live export before fd transfer

cursoragent
and
rchalamala
committed
test: run cross-process VMM IPC regressions across two GPUs

cursoragent
and
rchalamala
committed
feat: accept documented DEEPEP_MODE override in SGLang recipe

cursoragent
and
rchalamala
committed
test: assert SGLang hook doc lists _patch_attn_tp_logits_gather

cursoragent
and
rchalamala
committed
docs: document _patch_attn_tp_logits_gather in SGLang hook surface

cursoragent
and
rchalamala
committed
test: assert DeepEP dispatch routing instead of printing

cursoragent
and
rchalamala
committed
test: encode DeepEP dispatch payloads exactly in bfloat16

cursoragent
and
rchalamala
committed
test: race two importer threads on a peer chunk's first open

cursoragent
and
rchalamala
committed
test: exercise the VMM IPC insertion-race loser path under contention

cursoragent
and
rchalamala
committed
fix: factor VMM IPC chunk-request validator for both cache paths

cursoragent
and
rchalamala
committed
test: assert exact DeepEP dispatch delivery per expert

cursoragent
and
rchalamala
committed