Back to milestones

Frontend 1.33.0

Open

No due date

•Last updated Nov 1, 2026

70% complete

QSA backward: one two-pass engine for SM100, SM103 and SM107
Unread
#1825 opened Oct 24, 2026 · Updated Oct 26, 2026
7
build: require CuTeDSL 4.7.1 and JAX 0.6.0
Unread
#1819 opened Oct 23, 2026 · Updated Oct 26, 2026
3
Add JAX LayerNorm and RMSNorm with native gradients
Unread
#1812 opened Oct 23, 2026 · Updated Oct 31, 2026
19
fix(kda): reduce rounding error in Kimi CP state updates
Unread
#1804 opened Oct 23, 2026 · Updated Oct 30, 2026
6
fix(pygraph): tensor_like declares the dtype of non-torch buffers
Unread
#1772 opened Oct 21, 2026 · Updated Oct 26, 2026
14
perf(dsa): FP16-derived coarse key for wide BF16 rows in indexer_top_k
Unread
#1764 opened Oct 21, 2026 · Updated Oct 26, 2026Draft
2
Preserve head-major GDP K/V gradients for channel views
Unread
#1748 opened Oct 20, 2026 · Updated Oct 26, 2026Draft
2
Accept raw uint8 FP8 operands with dtype metadata in the block-scaled grouped GEMM wrappers
Unread
#1729 opened Oct 19, 2026 · Updated Oct 26, 2026
5
Relax SM restrictions in TensorIR MemBound Fusion samples
Unread
#1687 opened Oct 16, 2026 · Updated Oct 26, 2026
4
Optimize blk64 BSA with independent Q/KV pairing and split TMA loads
Unread
#1681 opened Oct 16, 2026 · Updated Oct 26, 2026Draft
6
NSA selection attention: mask keys past the sequence end (non-causal THD)
Unread
#1624 opened Oct 13, 2026 · Updated Oct 26, 2026
12
NSA top-k reduction: mask the compressed key one past the sequence
Unread
#1623 opened Oct 13, 2026 · Updated Oct 26, 2026
13
Add GNN MHA Simple Python and PyTorch API
Unread
#1588 opened Oct 12, 2026 · Updated Oct 30, 2026
6
[Docs] Fix remaining document-check wording from #1576
Unread
#1586 opened Oct 12, 2026 · Updated Oct 26, 2026
1
Fix typos in default kernel cache section of graph-api docs
Unread
#1583 opened Oct 12, 2026 · Updated Oct 26, 2026
2
test_mhas_v2 on cc 10.7 outgrew the SM107-test_mhas job's 2400 s timeout (13252 -> 17007 items)
NVIDIA/cudnn-frontend#1526 · opened on Oct 9, 2026 · Frontend 1.33.0
Update the cutdsl based Megamoe kernels
Unread
#1511 opened Oct 9, 2026 · Updated Oct 26, 2026
10
Match grouped SReLU forward to saved activation precision
Unread
#1501 opened Oct 8, 2026 · Updated Oct 26, 2026
7
Read ALiBi slopes and other constant tensors from device memory
Unread
#1419 opened Oct 6, 2026 · Updated Oct 26, 2026
4
[Bug]: cuDNN 9.27 BF16 SDPA forward nondeterminism on GB10 (COMPOSITE and UNIFIED; fixed-input repro)
NVIDIA/cudnn-frontend#1412 · opened on Oct 6, 2026 · Frontend 1.33.0
3
[DSA] Add prepared BF16 variable-length Top-K for GB300 and Rubin
Unread
#1298 opened Sep 30, 2026 · Updated Oct 26, 2026
2
Rule 8 (gated_attention_block): quant words in a workspace slot, per-device graph handle with locked re-stream
Unread
#1186 opened Sep 22, 2026 · Updated Oct 27, 2026
95
Rule 8 (HSTU): caller-owned workspace for block-sparse metadata and the bwd fp32 accumulator, R5 declines, compile-time fakes
Unread
#1185 opened Sep 22, 2026 · Updated Oct 27, 2026
107
Rule 8 (flex_attention): compile() from fake tensors, no device memory at compile
Unread
#1184 opened Sep 22, 2026 · Updated Oct 27, 2026
111
Rule 8 (NSA/CSA): plan-time host ints, SWA caller-owned workspace and THD ragged offsets; wrappers unchanged
Unread
#1183 opened Sep 22, 2026 · Updated Oct 27, 2026
115