Repository navigation

b12x fused MoE: TrellisMX (procedural NPK) source format with an in-tree P8 runtime - #69

Open

brandonmmusic-max wants to merge 254 commits into

local-inference-lab:mainfrom

brandonmmusic-max:b12x-trellismx-p8-runtime

Open

b12x fused MoE: TrellisMX (procedural NPK) source format with an in-tree P8 runtime#69
brandonmmusic-max wants to merge 254 commits into
local-inference-lab:mainfrom
brandonmmusic-max:b12x-trellismx-p8-runtime

This pull request is big! We're only showing the most recent 250 commits

Commits on Oct 8, 2026
- Ignore inactive hybrid A4 planning controls
  lukealonso
  committed
  ae91c3f
- Document and test the packed NVFP4 scale boundary
  lukealonso
  committed
  ef044dc
- Decode packed NVFP4-CSF scales inside explicit A4 MoE launches
  lukealonso
  committed
  36046c7
- Consume prefetched packed A4 scales above the CSF stage limit
  lukealonso
  committed
  a9b4c43
- Support 64-channel intermediate widths in packed NVFP4 A4 MoE
  lukealonso
  committed
  2f5210d
- fix(triton): preserve assembler selection across imports
  lukealonso
  committed
  98fb87a
Commits on Oct 9, 2026
Commits on Oct 10, 2026
Commits on Oct 11, 2026
- Support transient b12x panels on a shared Rich console
  lukealonso
  committed
  683390d
- Pass explicit query alignment to sparse MLA metadata compilation
  lukealonso
  committed
  b779705
- [b12x] Expand compact CSF blocks into native scale layouts
  lukealonso
  committed
  97fde4b
- Support packed MXFP4 weights through the standard blockscaled GEMM API
  lukealonso
  committed
  7807f2b
- [b12x] Materialize compressed checkpoint inputs directly on CPU
  lukealonso
  committed
  fb1ee5e
- [b12x] Prepare compressed FP4 checkpoint scales on CUDA
  lukealonso
  committed
  7e5a142
- Deduplicate CSF expansion tuning across layer payload sizes
  lukealonso
  committed
  499d6bd
- Pack head all-gather copies into full-warp transfers
  lukealonso
  committed
  d032273
- Scale PCIe pull all-reduce parallelism with payload size
  lukealonso
  committed
  accc563
- Prepare QSA benchmarks with the operand invocation contract
  lukealonso
  committed
  e683461
- Use split sparse attention through 128 query rows
  lukealonso
  committed
  7da3c36
- Size sparse-attention split capacity by local KV head count
  lukealonso
  committed
  ca05e43
- Cover shared NVFP4 input packing under graph replay
  lukealonso
  committed
  b6b1f73
- Publish NVFP4 tasks without an extra resident grid barrier
  lukealonso
  committed
  bf56ccb
- Run tensor-core GEMV with K-aligned warp groups
  lukealonso
  committed
  2b71ba9
- Reduce BF16 GEMV row tiles with one block barrier
  lukealonso
  committed
  a027606
- Use four compute warps for small-row MXFP8 A16 GEMM
  lukealonso
  committed
  38a24c3
- Autotune small-row MXFP8 and NVFP4 compute warps
  lukealonso
  committed
  3a7a4db
- Allow caller-owned shared storage in b12x PCIe collectives
  lukealonso
  committed
  0ed882a
- Allow caller-owned shared storage in b12x PCIe DMA all-reduce
  
  voipmonitor
  and
  claude
  committed
  ae3a205
Commits on Oct 12, 2026
- Publish multi-row TP2 all-reduces as Simple peer pushes
  
  voipmonitor
  and
  claude
  committed
  d38edea
- Push two-rank DCP head gathers and LSE combines into peer staging
  
  voipmonitor
  and
  claude
  committed
  241d979
- Tile prefill short convolutions over tokens
  
  original-el8
  authored and
  lukealonso
  committed
  299e10e
- Score QSA representatives with tiles of query rows
  
  original-el8
  authored and
  lukealonso
  committed
  0d6cd21
- Let callers attend QSA selections themselves
  
  original-el8
  authored and
  lukealonso
  committed
  13d2774
- Let chain speculative sampling draw each row's stream independently of its batch slot
  
  original-el8
  authored and
  lukealonso
  committed
  cbb56c0
- Key A16 dense GEMM programs by their compute-warp count
  
  original-el8
  authored and
  lukealonso
  committed
  6a460f6
- Run speculative-decode short convolutions in one launch
  
  original-el8
  authored and
  lukealonso
  committed
  ed8005a
Commits on Oct 13, 2026
- Support DeepSeek V4 cache writes and 64-head retrieval
  lukealonso
  committed
  d4f2289
- Pipeline large RoCEnante all-reduces as chunked two-shot reductions
  
  original-el8
  authored and
  lukealonso
  committed
  425fb6b
- Let RoCEnante run over a caller's pinned region and describe itself for native runtimes
  
  original-el8
  authored and
  lukealonso
  committed
  64916e8
Commits on Oct 14, 2026
- Let RoCEnante gather contributions whose lengths the device decides
  
  original-el8
  authored and
  lukealonso
  committed
  b2fe573
- Expect a variable-length gather's failure from its own launch as well
  
  original-el8
  authored and
  lukealonso
  committed
  396d37d
- List the variable-length all-gather among RoCEnante's surfaces
  
  original-el8
  authored and
  lukealonso
  committed
  8f76075
- Check the variable-length benchmark's rows against regenerated contributions
  
  original-el8
  authored and
  lukealonso
  committed
  98ad7ee
- Format RoCEnante's variable-length all-gather with the repository's hooks
  
  original-el8
  authored and
  lukealonso
  committed
  b19720a
- Derive uncapped DSA selection width from page-table metadata
  lukealonso
  committed
  6d7f238
- Support Windows checkpoint loading and JIT preparation
  lukealonso
  committed
  9873d7a
Commits on Oct 15, 2026
Commits on Oct 16, 2026
- b12x: fix tiled short-convolution row alignment
  lukealonso
  committed
  4b964e8
- b12x: support noncausal dense MLA
  lukealonso
  committed
  4b4589c
- b12x: add CPU encoder for CSF expert scales
  lukealonso
  committed
  5e2e9f6
- Add Step-5 kernel support and checkpoint MoE benchmark to b12x
  lukealonso
  committed
  6a61066
- Fuse per-head sigmoid gating into selected attention output
  lukealonso
  committed
  9eac39b
- b12x: fuse BF16 projection and hyperconnection stream gating
  
  original-el8
  authored and
  lukealonso
  committed
  e072445
- Support FP8 circular attention history with BF16 candidates
  lukealonso
  committed
  a706c3b
- Merge remote-tracking branch 'lil/main' into dev/lil-main
  lukealonso
  committed
  b1f081e
- fix(b12x): wait for two-shot sends before reusing pipe_send
  
  original-el8
  authored and
  lukealonso
  committed
  f3bbd1c
- feat(b12x): add experimental neighbor-ring RoCE transport
  
  original-el8
  authored and
  lukealonso
  committed
  14ba719
Commits on Oct 17, 2026
- b12x: spread dense MLA chunks over every launched split
  
  voipmonitor
  and
  claude
  committed
  40c7a24
- b12x: retune dense MLA splits after spreading chunks over every split
  
  voipmonitor
  and
  claude
  committed
  468feed
- b12x: let island reduce-scatter and vocabulary argmax use a caller IPC allocator
  
  authored and
  voipmonitor
  committed
  c6eca58
- b12x: memoize the dense MLA wave-balanced split search
  
  authored and
  voipmonitor
  committed
  02cb642
- b12x: load every KDA verification token before the recurrence
  
  voipmonitor
  and
  claude
  committed
  7e20380
- Add prepared Mamba SSD and Bolt routing support to b12x
  lukealonso
  committed
  3372e2d
- b12x: autotune inline CSF scales across routing mixtures
  lukealonso
  committed
  0213747
- b12x: align untuned CSF MoE defaults with routing regimes
  lukealonso
  committed
  d3afbb7
- b12x: accelerate compact MXFP4 decode and prefill
  lukealonso
  committed
  a6b096f
- b12x: apply SiTU in the micro kernel's gated W4A16 activation
  
  authored and
  voipmonitor
  committed
  582bf2c
- Use transitive source dependencies for b12x compiler and tuning caches
  lukealonso
  committed
  c46a0b5
- Fix Trellis lookup allocation for non-Trellis W4A16 MoE
  
  eugr
  authored and
  lukealonso
  committed
  969eaf9
- Bound RoCE generations and model source-buffer ownership
  
  original-el8
  authored and
  lukealonso
  committed
  fd29310
- Enumerate bounded ring ownership with equivalent QP helpers
  
  original-el8
  authored and
  lukealonso
  committed
  c637412
Commits on Oct 18, 2026
Commits on Oct 19, 2026
- Rename compressed MLA arithmetic constraint to compute_mode
  lukealonso
  committed
  ab8f0b9
- Memoize b12x callable source discovery during preparation
  lukealonso
  committed
  a38aacd
- Accept disjoint mHC views in shared scratch storage
  lukealonso
  committed
  dbd4635
- Avoid redundant copies in in-place top-k masking
  voipmonitor
  committed
  622cec1
- perf(b12x): keep eight relayed packets in flight per sender and rail
  
  authored and
  lukealonso
  committed
  d51d1b0
- formal(b12x): model the relay window in the ring ownership checks
  
  authored and
  lukealonso
  committed
  4092b28
Commits on Oct 21, 2026
Commits on Oct 22, 2026
- feat(b12x): add graduated autotune effort and correct candidate pruning
  lukealonso
  committed
  7364364
- Parallelize candidate program discovery during preparation
  lukealonso
  committed
  dc7493c
- Reject unsliceable mHC TF32 projection layouts at every tuning effort
  lukealonso
  committed
  c18261d
- Reclaim losing autotune executables between race batches
  lukealonso
  committed
  4eebbe1
- Persist b12x tuning progress and device-independent program discovery
  lukealonso
  committed
  16e0240
- Preserve singleton weight rows in BF16 prefill TMA descriptors
  lukealonso
  committed
  3602ebe
- Read F8_E8M0 scale storage in the CSF MXFP4 encoder
  lukealonso
  committed
  0765eaf
- fix(b12x): forward relayed ring packets in serial order
  
  FujitsuPolycom
  authored and
  lukealonso
  committed
  cb41087
- feat(b12x): open-chain mode for the neighbor ring transport
  
  FujitsuPolycom
  authored and
  lukealonso
  committed
  ed2a68a
- Defer cyclic collection until autotune job cleanup
  lukealonso
  committed
  ff6f0e8
- Break CuTe executor self-references during runtime initialization
  lukealonso
  committed
  3563f33
Commits on Oct 23, 2026
Commits on Oct 24, 2026
Commits on Oct 25, 2026
Commits on Oct 26, 2026
Commits on Oct 27, 2026
Commits on Oct 28, 2026
Commits on Oct 29, 2026
- Prevent dense split-K output offsets from overflowing 32 bits
  lukealonso
  committed
  f5f5341
Commits on Oct 31, 2026
- b12x fused MoE: TrellisMX (procedural NPK) source format with an in-tree P8 runtime
  brandonmmusic-max
  committed
  c37d2d9
Commits on Nov 1, 2026
- Support device-selected PCIe column gathers
  lukealonso
  committed
  a7f94e2
- Measure and recommend PCIe gather profiles
  lukealonso
  committed
  361e45f
Commits on Nov 3, 2026
Commits on Nov 4, 2026
- Declare calibrated A4 prefill for packed NVFP4 experts
  lukealonso
  committed
  dc37404
- Fix moe sfb double reshape (#71)
  mark-yong
  authored
  3c2729e
- Share storage across PCIe startup tuning races (#72)
  voipmonitor
  authored
  e977747
Commits on Nov 5, 2026
Commits on Nov 6, 2026
- [b12x] Register bounce slots with io_uring before CUDA (#76)
  original-el8
  authored
  057e6d1
- Merge local-inference-lab/flashinfer main into b12x-trellismx-p8-runtime
  brandonmmusic-max
  committed
  7b0d0f2

Commits

Commits on Oct 8, 2026

Commits on Oct 9, 2026

Commits on Oct 10, 2026

Commits on Oct 11, 2026

Commits on Oct 12, 2026

Commits on Oct 13, 2026

Commits on Oct 14, 2026

Commits on Oct 15, 2026

Commits on Oct 16, 2026

Commits on Oct 17, 2026

Commits on Oct 18, 2026

Commits on Oct 19, 2026