Repository navigation

cuda: add experimental native Flash Next family and local tensor parallelism - #952

Closed

akol1 wants to merge 137 commits into

ashhart:mainfrom

akol1:codex/flashnext-cuda-native

Closed

cuda: add experimental native Flash Next family and local tensor parallelism#952
akol1 wants to merge 137 commits into
ashhart:mainfrom
akol1:codex/flashnext-cuda-native

Commits on Oct 14, 2026
- lane core and native CUDA host for two-rank families with a vision tower: batched prompt passes (prefill_many, addStreams), a backend's refusals for a memory guard, follow() for a tensor-parallel r…
  BobClawblaw
  committed
  39de039
Commits on Oct 15, 2026
Commits on Oct 16, 2026
- Add lossless NVFP4 compressed scale planes and TP slices
  akol1
  committed
  3f069e3
- Integrate shared native TP host APIs with Flash Next CUDA
  akol1
  committed
  6837fae
- core: admit ModelOpt NVFP4 checkpoints from config.json's quantization block
  
  Bizuayeu
  authored and
  akol1
  committed
  33d6494
- core: read ModelOpt NVFP4 linears as stored, with an f32 reference row
  
  Bizuayeu
  authored and
  akol1
  committed
  9d4a6fa
- core: the ModelOpt NVFP4 linear's test decodes with core/quant's reference, not an f32 row of its own
  
  Bizuayeu
  authored and
  akol1
  committed
  798d38f
- Add lossless NVFP4 compressed scale planes and TP slices
  akol1
  committed
  d06009d
- cuda: refuse tensor parallelism before loading unranked families
  akol1
  committed
  2868a9a
- core: read exact rank-local NVFP4 tensors with plain or compressed scales
  akol1
  committed
  d631b15
- core: read exact rank-local NVFP4 tensors with plain or compressed scales
  akol1
  committed
  0a0bb86
- cuda: serve losslessly packed MXFP8 projections through native qlinear
  akol1
  committed
  f1f7f9a
- cuda: serve losslessly packed MXFP8 projections through native qlinear
  akol1
  committed
  17ce054
- Expose FP32 lane projection outputs before residual reduction
  akol1
  committed
  f6a0055
- Read native NVFP4 n-gram rows directly from mapped checkpoint shards
  akol1
  committed
  1f17be1
- Expose FP32 lane projection outputs before residual reduction
  akol1
  committed
  3d56b2a
- Preserve dense connection weights and FP32 normalization in native CUDA passes
  akol1
  committed
  32be3e5
- Support exact BF16 token rows and dense vocabulary projections
  akol1
  committed
  2187111
- Load mixed dense and microscaled projections without requantization
  akol1
  committed
  97715a2
- Load exact rank-local NVFP4 routed expert partitions
  akol1
  committed
  2661d28
- Run the mixed-format shared expert separately from routed partitions
  akol1
  committed
  041cc0a
- Bound CUDA rank control and NCCL collective failures
  akol1
  committed
  359eef3
- Add bounded Flash Next rank replay and checkpoint norm conversion
  akol1
  committed
  567be14
- Serve mixed-format Flash Next ranks with bounded admission and health
  akol1
  committed
  aa4015e
- Report unavailable engine health and preserve admission HTTP 503
  akol1
  committed
  a801b9a
- perf: queue TP work and disable delayed control frames
  akol1
  committed
  09830ff
- feat: add bounded opt-in tensor-rank decode graph replay
  akol1
  committed
  e2ebb0d
- Load native ModelOpt MTP projections and exact draft vocabulary rows
  akol1
  committed
  64bd35b
- Mirror MTP verification windows and prefix commits across CUDA ranks
  akol1
  committed
  1967f90
- Add opt-in affine proposal head with independent arithmetic checks
  akol1
  committed
  69e335e
- Allow bounded draft chains through the full verification window
  akol1
  committed
  a7af69b
- Expose bounded NVFP4 expert CTA geometry without changing tile arithmetic
  akol1
  committed
  c325d76
- Export fixed-history full score rows for independent model fidelity checks
  akol1
  committed
  e71e400
- Experiment with exact vocabulary-column projection across CUDA ranks
  akol1
  committed
  961e1d3
- Add single-rank score export to isolate full-model fidelity
  akol1
  committed
  8f014e1
- Disable Nagle on accepted HTTP streaming connections
  akol1
  committed
  c611ade
- Disable Nagle on accepted HTTP streaming connections
  akol1
  committed
  7b0cd5a
- Prefer shared plain rounds when admitting concurrent decode streams
  akol1
  committed
  94cefdd
- Merge remote-tracking branch 'origin/main' into codex/http-stream-nodelay
  akol1
  committed
  4e31bd7
- Merge upstream main and preserve distinct engine readiness and stale-state failures
  akol1
  committed
  91f9f4b
- Merge upstream main and add MXFP8 numerical contract coverage
  akol1
  committed
  d40bc9e
- Merge current main and retain FP32 projections across packed formats
  akol1
  committed
  cb23d3d
- Merge refreshed MXFP8 dependency and numerical contract harness
  akol1
  committed
  ff97f72
- Exercise FP32 packed projections through numerical contract gates
  akol1
  committed
  50c4f56
- Cover served dense and attention projection dimensions
  akol1
  committed
  37578da
- Merge current main for host-validated CSF scale decoding
  akol1
  committed
  8a41d66
- Merge refreshed lossless CSF dependency and current main
  akol1
  committed
  c224f30
- test(cuda): register MXFP8 launch resource limits
  akol1
  committed
  a008b5c
- Merge commit 'a008b5cd' into codex/cuda-f32-projections
  akol1
  committed
  b3d4713
- test(cuda): register FP32 projection launch resources
  akol1
  committed
  5bb7392
- Merge lossless NVFP4 reader support for the native family backend
  akol1
  committed
  3eecc6e
- Merge bounded serving health contracts for native rank owners
  akol1
  committed
  4f11337
- Merge accepted-connection streaming latency control
  akol1
  committed
  1695e16
- cuda: refresh compiler-derived launch resource records
  akol1
  committed
  cb2c749
- cuda: integrate native Flash Next family with current engine
  akol1
  committed
  d9ce107
- test(cuda): expose writeback double rounding against FP64
  akol1
  committed
  fe461bb
- fix(cuda): fuse Flash Next hidden-state writeback before BF16 rounding
  akol1
  committed
  8f7af95
- cuda: refresh compiler-derived launch resource records
  akol1
  committed
  2e59316
- style: format native family integration with pinned Zig
  akol1
  committed
  b0753fa
- Merge compiler resource records into MXFP8 qualification
  akol1
  committed
  839a5a4
- Merge qualified resource records through MXFP8 dependency
  akol1
  committed
  787064f
- test(cuda): declare native Flash Next launch resource inventory
  akol1
  committed
  f141cd5
- test(cuda): qualify hyper-connection writeback through numerical gates
  akol1
  committed
  af32562
- test(cuda): contract ordered slice sums and history copies
  akol1
  committed
  79480c0
- test(cuda): check utility guards at every shape
  akol1
  committed
  e4efbb2
- test(cuda): run native history and slice utility contracts
  akol1
  committed
  0e78b5f
- test(cuda): independently contract sparse attention selection
  akol1
  committed
  5049af9
- test(cuda): contract proposal maximum and log-sum-exp independently
  akol1
  committed
  63000ad
- test(cuda): run independent sparse selection contracts
  akol1
  committed
  d467307
- test(cuda): record compiled native family launch resources
  akol1
  committed
  8874feb
- fix(test): retain CUDA contract hooks for the runner lifetime
  akol1
  committed
  4683601
- Merge remote-tracking branch 'origin/main' into codex/engine-unavailable-health
  akol1
  committed
  dd11a22
- Merge remote-tracking branch 'origin/main' into codex/http-stream-nodelay
  akol1
  committed
  fdbf481
- test: bound accepted TCP stream checks
  akol1
  committed
  7b94b5e
- flashnext cuda: refuse rank snapshot reuse without draft state
  
  cwschroeder
  authored and
  akol1
  committed
  a56a005
- Reject non-finite errors in CUDA reference comparisons
  
  cwschroeder
  authored and
  akol1
  committed
  df44f51
- Merge remote-tracking branch 'origin/main' into codex/mxfp8-checker-review
  akol1
  committed
  5a94b2f
- Merge branch 'codex/mxfp8-checker-review' into codex/projection-fixture-review
  akol1
  committed
  6b78da2
- test: poison each FP32 projection dispatch and reduction scratch
  
  cwschroeder
  authored and
  akol1
  committed
  37359a1
- test(cuda): reject no-op projection dispatches after stale prefixes
  akol1
  committed
  4544c5c
- test(server): cover streamed admission and drained health regressions
  
  cwschroeder
  authored and
  akol1
  committed
  f859ed4
- fix(native): report drained Flash Next hosts unavailable
  
  cwschroeder
  authored and
  akol1
  committed
  e8f50da
- Merge projection checker and dispatch fixture review fixes
  akol1
  committed
  e832734
- Merge commit '7b94b5e6d20a4a18b7d8d4707e6c640049044958' into codex/flashnext-review-integration
  akol1
  committed
  e649136
- fix(server): admit streams before opening and wire family readiness
  akol1
  committed
  8f87428
- Merge admission and readiness review fixes into native family
  akol1
  committed
  f9a0585
- native: keep CUDA context configuration in a focused module
  akol1
  committed
  66c9e37
- test(native): retain context-window coverage in the host root
  akol1
  committed
  beec1aa
Commits on Oct 18, 2026
- test(cuda): clear experts-rows buffers on the kernels' stream
  
  balancedscorpion
  authored and
  akol1
  committed
  ab7f7ba
- docs: measure the ModelOpt FP4-Local checkpoint on one GB10
  
  balancedscorpion
  authored and
  akol1
  committed
  b6fd0d9

Commits

Commits on Oct 14, 2026

Commits on Oct 15, 2026

Commits on Oct 16, 2026

Commits on Oct 18, 2026