All pull requests

New pull request

[FROST] Preserve Int64 THD output strides through descriptor setup
#1224YangXu1990uiuc opened Sep 25, 2026Collaborator Frontend 1.31.0
4
frost(sdpa_fwd_sm80): serve dense GQA/MQA natively, drop the per-execute K/V head expansion
#1223harshithkantamneni opened Sep 25, 2026
2
Reduce Python graph execution overhead with ordered tensor bindings
#1222YangXu1990uiuc opened Sep 25, 2026Collaborator Frontend 1.31.0
24
fix(bsa): stabilize SM100 forward softmax for large logits
#1221Butterfingrz opened Sep 24, 2026Contributor
1
Update Benchmarking Artifacts - backend 9.27.0.21, frontend ab43be0
#1219brandonfzhang opened Sep 24, 2026Collaborator
1
Reduce FROST THD host overhead with native argument binding
#1217YangXu1990uiuc opened Sep 24, 2026Collaborator Frontend 1.31.0
9
[FROST] Prepare SM100 D128 FP8 forward host launches
#1216YangXu1990uiuc opened Sep 24, 2026Collaborator Frontend 1.31.0
26
frost(sdpa): paged KV on the SM100 MXFP8 flavors (F8_128x4 descale pools, page_size % 128)
#1214Adnios opened Sep 24, 2026Collaborator Frontend 1.30.0
13
BSA: optimize SM90 blk64 forward performance
#1213Butterfingrz opened Sep 24, 2026Contributor
11
frost(sdpa/bwd): sm107 (Rubin) d=256 backward -- bf16/fp16 and per-tensor FP8 kernels ported from the reference CTM pipelines, engine rows, tests, perf
#1212RomanAnders90 opened Sep 24, 2026Contributor
2
test(sdpa/frost): run the Stats assertion of test_fp8_thd_sliding_window before _check (follow-up to #1209)
#1211RomanAnders90 opened Sep 24, 2026Contributor Frontend 1.30.0
1
Add C++ and Python APIs for custom numerical weight dequantization
#1210scottyokim opened Sep 24, 2026
1
Add bounded causal GQA backward for SM107
#1207layalir opened Sep 23, 2026
7
[DSv4.1] Reduce Engram backward register pressure on SM100
#1204YangXu1990uiuc opened Sep 23, 2026Collaborator Frontend 1.31.0
4
FROST GEMM heuristic: decline split-K on block-scale chains; keep block-scale MMA-K at 32 bytes
#1201brandonfzhang opened Sep 23, 2026Collaborator Frontend 1.31.0
3
[DSv4.1] Support native N-major BF16 expert input gradients
#1199YangXu1990uiuc opened Sep 23, 2026Collaborator Frontend 1.31.0
7
Add aligned HCA backward for GB300 and Rubin
#1198layalir opened Sep 22, 2026 Frontend 1.31.0
16
Rule 8 (gated_attention_block): quant words in a workspace slot, per-device graph handle with locked re-stream
#1186YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
64
Rule 8 (HSTU): caller-owned workspace for block-sparse metadata and the bwd fp32 accumulator, R5 declines, compile-time fakes
#1185YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
83
Rule 8 (flex_attention): compile() from fake tensors, no device memory at compile
#1184YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
72
Rule 8 (NSA/CSA): plan-time host ints, SWA caller-owned workspace and THD ragged offsets; wrappers unchanged
#1183YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
87
Rule 8 (DSA): plan-style classes own no device memory -- required outputs, caller-owned workspaces, plan-time host ints; wrappers unchanged
#1182YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
68
Rule 8 core: caller-owned workspaces and no host blocking in the graph API, FROST MoE GEMM and SDPA, grouped GEMM and Hopper KDA; R9 detector and recipes R10/R11
#1181YangXu1990uiuc opened Sep 22, 2026Collaborator Frontend 1.31.0
88
Specialize top-k row stride for SM100 DSA H16/H32/H96
#1166icavan opened Sep 21, 2026Contributor Frontend 1.31.0
6
BSA: add native SM120 blk128 Sage FP8 forward
#1163tiffany940107 opened Sep 21, 2026Contributor Frontend 1.31.0
3