Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
22 commits
Select commit Hold shift + click to select a range
60964fa
test(KERNEL-SSM-MAMBA): RED -- the CUDA arm of the three Mamba2 SSD o…
mudler Aug 12, 2026
fcdb7d8
recover(KERNEL-SSM-MAMBA): the W2 CUDA arm, rescued byte-exact from t…
mudler Aug 12, 2026
fe06959
merge: origin/main into row/KERNEL-SSM-MAMBA-SSD-W2-FINISH (#496)
mudler Aug 13, 2026
f9915f8
WIP spec(KERNEL-SSM-MAMBA): re-author $8.3 equivalence contract, clos…
mudler Aug 13, 2026
08b6e0b
WIP spec(KERNEL-SSM-MAMBA): update $8 Now for the W2 state (#496)
mudler Aug 13, 2026
f1929de
WIP spec(KERNEL-SSM-MAMBA): record the W2 evidence -- 9/9 mutations c…
mudler Aug 13, 2026
62a4bad
spec(KERNEL-SSM-MAMBA): record the full-ctest result and the one unat…
mudler Aug 13, 2026
23bc605
spec(KERNEL-SSM-MAMBA): the ctest ran under undetected GPU contention…
mudler Aug 13, 2026
1e81914
spec(KERNEL-SSM-MAMBA): the Windows CI reds are the main baseline; th…
mudler Aug 13, 2026
f800e78
merge: origin/main into the W2 tightening branch
mudler Aug 13, 2026
05898dc
fix(KERNEL-SSM-MAMBA): the derived bound omitted nvcc FMA contraction…
mudler Aug 13, 2026
f3c943b
record(KERNEL-SSM-MAMBA): the CPU-lane full gate for the W2 tightenin…
mudler Aug 13, 2026
d12f9c8
record(KERNEL-SSM-MAMBA): §8.4's audit percentages are against the OL…
mudler Aug 13, 2026
f5c589f
docs(KERNEL-SSM-MAMBA): the new device case's own comment under-count…
mudler Aug 13, 2026
0c6dd44
spec(KERNEL-SSM-MAMBA): test_minimax_h3 ATTRIBUTED -- it is #486/#516…
mudler Aug 13, 2026
da09279
spec(KERNEL-SSM-MAMBA): all ten ctest failures reproduce STANDALONE -…
mudler Aug 13, 2026
6a418e0
spec(KERNEL-SSM-MAMBA): the gate-host evidence logs kept at ~/w2ssd-e…
mudler Aug 13, 2026
a88bab2
record(KERNEL-SSM-MAMBA): the CUDA arm COMPILES under real nvcc -- op…
mudler Aug 13, 2026
72dc6b4
merge: the W2-FINISH evidence commits into the tightening branch (#496)
mudler Aug 13, 2026
497cc63
merge: origin/main into the W2 landing branch (#496)
mudler Aug 13, 2026
5b83a69
record(KERNEL-SSM-MAMBA): all four owed W2 gates discharged by measur…
mudler Aug 13, 2026
ac7b1fc
merge: origin/main into the W2 landing branch, second pass (#496)
mudler Aug 13, 2026
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
637 changes: 626 additions & 11 deletions .agents/specs/mamba2-ssd.md

Large diffs are not rendered by default.

22 changes: 22 additions & 0 deletions src/vt/cuda/cuda_gdn.cu
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,17 @@
#include "vt/cuda/conv_update_fast.h"
#include "vt/cuda/cuda_device_caps.h"
#include "vt/cuda/cuda_gdn_internal.h"
// MAMBA2 / SSD IS NOT THE GATED DELTA RULE (.agents/specs/mamba2-ssd.md §0, §7):
// no delta-removal term, decay driven by A_log/dt, B/C shared across n_groups.
// The three kernels live in their own header and their own namespace and share
// nothing with the GDN/KDA code below; they are compiled into THIS translation
// unit because it is where the sibling SSM/linear-attention device arms already
// live, and because a new .cu has to be listed in the ROOT CMakeLists.txt, which
// check-doc-checkpoint classifies as `user_usage` and therefore charges a
// docs/USAGE.md update that a kernel exposing no command, config key or C-ABI
// entry point has nothing true to write (the same deviation W1 recorded for
// cpu_ops.cpp, mamba2-ssd.md §8.1).
#include "vt/cuda/cuda_mamba2_ssd.cuh"
#include "vt/cuda/gdn_decode_fused.h"
#include "vt/cuda/gdn_packed_decode_triton.h"
#include "vt/cuda/gdn_prefill_conv.h"
Expand Down Expand Up @@ -6641,6 +6652,17 @@ struct Registrar {
RegisterOp(OpId::kIndexCopy, DeviceType::kCUDA,
reinterpret_cast<void*>(
static_cast<IndexCopyFn>(&IndexCopyKernelCuda)));
// Mamba2 / SSD device arm (mamba2-ssd.md W2, #496) — sibling ops, never a
// parameterisation of the GDN kernels above.
RegisterOp(OpId::kMamba2ChunkScan, DeviceType::kCUDA,
reinterpret_cast<void*>(
static_cast<Mamba2ChunkScanFn>(&mamba2::Mamba2ChunkScanKernelCuda)));
RegisterOp(OpId::kMamba2StateUpdate, DeviceType::kCUDA,
reinterpret_cast<void*>(
static_cast<Mamba2StateUpdateFn>(&mamba2::Mamba2StateUpdateKernelCuda)));
RegisterOp(OpId::kRmsNormGatedGroup, DeviceType::kCUDA,
reinterpret_cast<void*>(
static_cast<RmsNormGatedGroupFn>(&mamba2::RmsNormGatedGroupKernelCuda)));
}
} registrar;

Expand Down
Loading
Loading