feat(attention): cuDNN FROST attention backend for head_dim in (256, 512], with context parallelism - #3527

Open

nvegesna-netizen wants to merge 32 commits into

NVIDIA:mainfrom

nvegesna-netizen:nvegesna/te-frost-d512-cp

Commits on Sep 16, 2026

feat(attention): cuDNN FROST kernel wrapper for head_dim in (256, 512]

nvegesna-netizen
and
claude
committed
cdab664
feat(attention): select and dispatch FrostAttention from DotProductAttention

nvegesna-netizen
and
claude
committed
cf97f19
feat(attention): context parallelism for FROST across p2p, all_gather and a2a

nvegesna-netizen
and
claude
committed
9b79991
test(attention): CP coverage for FrostAttention at head_dim 512

nvegesna-netizen
and
claude
committed
24713e1
[pre-commit.ci] auto fixes from pre-commit.com hooks

pre-commit-ci[bot]
authored and
nvegesna-netizen
committed
bc15d3a
test(attention): run the FrostAttention CP configs from pytest

nvegesna-netizen
and
claude
committed
6631fcd
fix(attention): gate FROST on the cuDNN Frontend version and key plans by device

nvegesna-netizen
and
claude
committed
4381310
fix(attention): repair the FROST plan-cache key arity and harden the version gate

nvegesna-netizen
and
claude
committed
4069bfa
fix(attention): reject a v that does not match k, and refine the version probe

nvegesna-netizen
and
claude
committed
3706931
fix(attention): update the mixed-THD backend unpack for the new return value

nvegesna-netizen
and
claude
committed
02ac9c7
fix(attention): bind a cuDNN stream and close the remaining silent-wrong-answer paths

nvegesna-netizen
and
claude
committed
7755a05
fix(attention): JIT-compile FROST plans under the device their handle names

nvegesna-netizen
and
claude
committed
69d1741
test(attention): anchor FROST numerics to an fp32 reference, not to itself

nvegesna-netizen
and
claude
committed
012a692
feat(attention): honour deterministic on the FROST path, and document the backend

nvegesna-netizen
and
claude
committed
f36a021
docs(attention): scope the FROST exclusivity claim to context parallelism

nvegesna-netizen
and
claude
committed
c47a0e5
fix(attention): drop a duplicate deterministic parameter on the fused p2p wrapper

nvegesna-netizen
and
claude
committed
a82a74b
fix(attention): decline FROST when determinism is required

nvegesna-netizen
and
claude
committed
f5f9593
test(attention): make the FROST oracle float64, since an fp32 one is not a reference

nvegesna-netizen
and
claude
committed
685d190
feat(attention): express FROST masking as a diagonal band, adding sliding window

nvegesna-netizen
and
claude
committed
0ef2fe6
fix(attention): carry the sliding window through a2a, and decline it for p2p

nvegesna-netizen
and
claude
committed
839326c
test(attention): cover the sliding window in backward, at its boundary, and per cp_comm_type

nvegesna-netizen
and
claude
committed
99c0c1e
fix(attention): let the CP sliding-window asserts know FROST exists

nvegesna-netizen
and
claude
committed
8a0b2e2
docs(attention): correct the claimed cuDNN import-ordering hazard

nvegesna-netizen
and
claude
committed
f7de29a
test(attention): apply the window for every mask type in the reference

nvegesna-netizen
and
claude
committed
74b31a9
docs(attention): justify the p2p sliding-window decline from the ring, not precedent

nvegesna-netizen
and
claude
committed
0133263
fix(attention): bind the FROST flag on the ONNX path, decline what was silently ignored

nvegesna-netizen
and
claude
committed
e7af980
fix(attention): read qkv_type from attention_params, not the rebound local

nvegesna-netizen
and
claude
committed
d94220f
test(attention): cover the ONNX-export branch on hardware that can run it

nvegesna-netizen
and
claude
committed
e5071de

Commits on Sep 17, 2026

feat(attention): allow FrostAttention with cp_comm_type=a2a+p2p

nvegesna-netizen
and
claude
committed
bcbad7c
fix(attention): handle a list-valued cp_group in FrostAttention.forward

nvegesna-netizen
and
claude
committed
edd82de
test(attention): cover fp16 in the backward and under context parallelism

nvegesna-netizen
and
claude
committed
3429a6f
docs(attention): narrow the FrostAttention availability claim

nvegesna-netizen
and
claude
committed
9bb63c3

Navigation Menu

feat(attention): cuDNN FROST attention backend for head_dim in (256, 512], with context parallelism - #3527

feat(attention): cuDNN FROST attention backend for head_dim in (256, 512], with context parallelism#3527
nvegesna-netizen wants to merge 32 commits into
NVIDIA:mainfrom
nvegesna-netizen:nvegesna/te-frost-d512-cp

Commits on Sep 16, 2026

feat(attention): cuDNN FROST kernel wrapper for head_dim in (256, 512]

feat(attention): select and dispatch FrostAttention from DotProductAttention

feat(attention): context parallelism for FROST across p2p, all_gather and a2a

test(attention): CP coverage for FrostAttention at head_dim 512

[pre-commit.ci] auto fixes from pre-commit.com hooks

test(attention): run the FrostAttention CP configs from pytest

fix(attention): gate FROST on the cuDNN Frontend version and key plans by device

fix(attention): repair the FROST plan-cache key arity and harden the version gate

fix(attention): reject a v that does not match k, and refine the version probe

fix(attention): update the mixed-THD backend unpack for the new return value

fix(attention): bind a cuDNN stream and close the remaining silent-wrong-answer paths

fix(attention): JIT-compile FROST plans under the device their handle names

test(attention): anchor FROST numerics to an fp32 reference, not to itself

feat(attention): honour deterministic on the FROST path, and document the backend

docs(attention): scope the FROST exclusivity claim to context parallelism

fix(attention): drop a duplicate deterministic parameter on the fused p2p wrapper

fix(attention): decline FROST when determinism is required

test(attention): make the FROST oracle float64, since an fp32 one is not a reference

feat(attention): express FROST masking as a diagonal band, adding sliding window

fix(attention): carry the sliding window through a2a, and decline it for p2p

test(attention): cover the sliding window in backward, at its boundary, and per cp_comm_type

fix(attention): let the CP sliding-window asserts know FROST exists

docs(attention): correct the claimed cuDNN import-ordering hazard

test(attention): apply the window for every mask type in the reference

docs(attention): justify the p2p sliding-window decline from the ring, not precedent

fix(attention): bind the FROST flag on the ONNX path, decline what was silently ignored

fix(attention): read qkv_type from attention_params, not the rebound local

test(attention): cover the ONNX-export branch on hardware that can run it

Commits on Sep 17, 2026

feat(attention): allow FrostAttention with cp_comm_type=a2a+p2p

fix(attention): handle a list-valued cp_group in FrostAttention.forward

test(attention): cover fp16 in the backward and under context parallelism

docs(attention): narrow the FrostAttention availability claim