Repository navigation

[3/3] perf(dsv4): add hardware-aware adaptive verification - #71

Open

calvarado2004 wants to merge 28 commits into

FlashML-org:mainfrom

calvarado2004:pr/dsv4-dspark-adaptive

Open

[3/3] perf(dsv4): add hardware-aware adaptive verification#71
calvarado2004 wants to merge 28 commits into
FlashML-org:mainfrom
calvarado2004:pr/dsv4-dspark-adaptive

Commits on Sep 16, 2026
- feat(dsv4): add tensor-parallel DeepSeek-V4 runtime
  calvarado2004
  committed
  6f6b5ca
- fix(server): align TP GPU mapping after rebase
  
  nhannguyen2709
  authored and
  calvarado2004
  committed
  29ae93c
- test(dsv4): verify TP expert bank contents
  
  nhannguyen2709
  authored and
  calvarado2004
  committed
  754d209
- fix(dsv4): reject TP1 FTW layout under TP
  
  nhannguyen2709
  authored and
  calvarado2004
  committed
  ca52181
- feat(moe): allow pageable CPU-layer banks
  
  nhannguyen2709
  authored and
  calvarado2004
  committed
  4380389
- test(dsv4): pin the ds_fp4 expert bank TP contract before the kernel change
  calvarado2004
  committed
  e07436b
- feat(dsv4): shard the ds_fp4 expert banks across TP ranks
  calvarado2004
  committed
  6b5187e
- test(dsv4): slice the fused gate|up rows as two blocks, not one
  calvarado2004
  committed
  de17781
- test(dsv4): accumulate the TP expert partials in fp32
  calvarado2004
  committed
  9117793
- test(dsv4): retarget the TP contract test at the shared layer classes
  calvarado2004
  committed
  cbda938
- test(dsv4): assert the shrinking axis, and restore the dim parametrization
  calvarado2004
  committed
  0dd900d
- test(engine): give the dsv4 config stub the tp_info EngineConfig always has
  calvarado2004
  committed
  e7b3d36
- feat(moe): bound parallel expert prefetch
  
  nhannguyen2709
  authored and
  calvarado2004
  committed
  4185099
- fix(dsv4): release each expert shard's page cache as the serial reader leaves it
  
  calvarado2004
  and
  claude
  committed
  5dc0803
- fix(dsv4): unmap each expert shard before dropping its page cache
  calvarado2004
  committed
  e206163
- feat(dsv4): implement exact DSpark speculative decoding
  calvarado2004
  committed
  33f409f
- fix(dsv4): make hybrid DSpark serving operational
  calvarado2004
  committed
  f0f3f07
- perf(dsv4): add TP verification and adaptive DSpark scheduling
  calvarado2004
  committed
  31eb9a5
- perf(dsv4): accelerate probabilistic DSpark verification
  calvarado2004
  committed
  6c0a0c0
- fix(dspark): detokenize accepted blocks exactly once
  calvarado2004
  committed
  a082e46
- perf(dsv4): fuse the HC pre-norm reduction into one kernel
  
  authored and
  calvarado2004
  committed
  6214026
- fix(dsv4): harden long prefills with fused HC norm
  calvarado2004
  committed
  2cacfc4
- perf(nccl): all-reduce TP tensors in place
  calvarado2004
  committed
  affe163
- feat(engine): report final NUMA-aware model placement
  calvarado2004
  committed
  3cb150d
- fix(engine): make placement report self-contained
  calvarado2004
  committed
  e6957bd
- perf(moe): prioritize repeated speculative routes
  calvarado2004
  committed
  a9a3806
- perf(dspark): bypass persistently weak drafts
  calvarado2004
  committed
  9b3eda7
- feat(scripts): enable the dspark acceptance fallback by default
  calvarado2004
  committed
  5c4cbd6

Commits

Commits on Sep 16, 2026