Repository navigation

[PyTorch] Add causal MXFP8 attention without future-token leakage through V scales - #3865

Draft

romitjain wants to merge 58 commits into

NVIDIA:mainfrom

romitjain:mxfp8-completed-block-causal

Draft

[PyTorch] Add causal MXFP8 attention without future-token leakage through V scales#3865
romitjain wants to merge 58 commits into
NVIDIA:mainfrom
romitjain:mxfp8-completed-block-causal

Commits on Oct 11, 2026
Commits on Oct 15, 2026
Commits on Oct 16, 2026
- Merge main into py-all
  vcherepanov-nv
  committed
  56606d2
Commits on Oct 19, 2026
- Fix cuDNN attention support guards and output initialization
  vcherepanov-nv
  committed
  13141e3
Commits on Oct 20, 2026
- Merge branch 'main' into py-all
  vcherepanov-nv
  committed
  4a70c52
- Fix build tools imports in JAX dependency tests
  vcherepanov-nv
  committed
  e9e9b8c
- Skip deterministic cuDNN execution tests below SM90
  vcherepanov-nv
  committed
  9e0f6e7
- Share MXFP8 attention preparation rules across frontends
  vcherepanov-nv
  committed
  01ffce2
- Add dense MXFP8 attention to JAX
  vcherepanov-nv
  committed
  da5c011
- Extend JAX MXFP8 attention layouts and feature coverage
  vcherepanov-nv
  committed
  a32c9fa
- Reuse prequantized MXFP8 operands in JAX attention
  vcherepanov-nv
  committed
  950cf8d
- Enable packed and prequantized MXFP8 context parallelism
  vcherepanov-nv
  committed
  f5efdfc
Commits on Oct 21, 2026
- Skip unsupported JAX cuDNN attention test configurations
  vcherepanov-nv
  committed
  88fd9d2
Commits on Oct 27, 2026
- Require precompiled cuDNN plans during JAX graph deserialization
  vcherepanov-nv
  committed
  6bb6482
- Temporarily disable FROST plans in JAX cuDNN attention
  vcherepanov-nv
  committed
  710e419
Commits on Oct 29, 2026
- Fix PyTorch fused attention precision and inference dropout
  vcherepanov-nv
  committed
  070c008
- Fix SBHD paged-cache dimensions in PyTorch attention
  vcherepanov-nv
  committed
  5ab71e0
Commits on Oct 30, 2026
- Merge main into py-all
  vcherepanov-nv
  committed
  569e9d3
- Fall back from fused attention when the cuDNN frontend is unavailable
  
  vcherepanov-nv
  and
  claude
  committed
  cad3188
- Cache the fused attention cache-diagnostics level
  
  vcherepanov-nv
  and
  claude
  committed
  ff6d4a0
- Propagate fused attention setup errors again
  
  vcherepanov-nv
  and
  claude
  committed
  8ee7f42
Commits on Nov 5, 2026
- Merge branch 'main' into py-all
  vcherepanov-nv
  committed
  4db3a2a
- Fix FP8 attention test backward mode isolation
  vcherepanov-nv
  committed
  5b5cf2d
- Fix host scalar lifetime in JAX cuDNN graph execution
  vcherepanov-nv
  committed
  6737567
Commits on Nov 6, 2026
- Fix JAX attention documentation formatting
  vcherepanov-nv
  committed
  a2596ce
- Merge branch 'main' into py-all
  vcherepanov-nv
  committed
  0054179
Commits on Nov 10, 2026
- Merge main into py-all
  vcherepanov-nv
  committed
  b175ee4
- Simplify cuDNN attention documentation
  vcherepanov-nv
  committed
  8598bcd
Commits on Nov 11, 2026
- Match common attention batch bucketing
  vcherepanov-nv
  committed
  a93a55f
- Remove JAX attention cache size limits
  vcherepanov-nv
  committed
  dd4457a
- Address PyTorch cuDNN attention review feedback
  vcherepanov-nv
  committed
  c0890e2
- Retire the legacy PyTorch fused attention wrapper
  vcherepanov-nv
  committed
  5213d19
- Address straightforward PyTorch attention review feedback
  vcherepanov-nv
  committed
  5c21002
- Remove redundant attention metadata tests
  vcherepanov-nv
  committed
  5344e71
- Fix JAX cuDNN support checks and reuse common handles
  vcherepanov-nv
  committed
  6727540
- Replace the JAX backend binding with a Python enum
  vcherepanov-nv
  committed
  303d618
- Support both O precisions in JAX FP8 backward
  vcherepanov-nv
  committed
  e396500
- Match main cuDNN attention heuristic selection
  vcherepanov-nv
  committed
  d5fdde2
- Group cuDNN attention dispatch and rename MXFP8 helpers
  vcherepanov-nv
  committed
  509322f
- Match fused attention bias query and runtime descriptors
  vcherepanov-nv
  committed
  c290457
Commits on Nov 12, 2026
- Remove MXFP8 support-surface and scale-byte tests
  vcherepanov-nv
  committed
  b65fd23
- Remove the empty ragged-capacity descriptor test
  vcherepanov-nv
  committed
  9231140
- Group JAX distributed attention tests
  vcherepanov-nv
  committed
  b52a40c
- Share cuDNN attention caches across SDPA paths
  vcherepanov-nv
  committed
  06536d6
- Consolidate JAX FP8 attention numerical tests
  vcherepanov-nv
  committed
  d1249bf
- Merge main into py-all
  vcherepanov-nv
  committed
  bd81acc
- Skip the bias graph cache test in deterministic mode
  vcherepanov-nv
  committed
  9080bdd
- Fix architecture assumptions in PyTorch attention tests
  vcherepanov-nv
  committed
  303ab18
- Remove JAX build dependency metadata tests
  vcherepanov-nv
  committed
  b9315cf
- Reuse shared graph attributes in JAX Flex Attention
  vcherepanov-nv
  committed
  3bdbc83
- Add Flex Attention cache diagnostics
  vcherepanov-nv
  committed
  b8b8b6b
- Merge branch 'main' into py-all
  vcherepanov-nv
  committed
  c689b37
Commits on Nov 13, 2026
- [PyTorch] Add causal MXFP8 attention without future-token leakage through V scales
  
  committed
  58582b2

Commits

Commits on Oct 11, 2026

Commits on Oct 15, 2026

Commits on Oct 16, 2026

Commits on Oct 19, 2026

Commits on Oct 20, 2026

Commits on Oct 21, 2026

Commits on Oct 27, 2026

Commits on Oct 29, 2026

Commits on Oct 30, 2026

Commits on Nov 5, 2026

Commits on Nov 6, 2026

Commits on Nov 10, 2026

Commits on Nov 11, 2026

Commits on Nov 12, 2026

Commits on Nov 13, 2026