[Docs] Add compute capability 11.0 to the support matrix

[0z5a]

Description

docs/support_matrix.rst does not list compute capability 11.0 (Jetson AGX Thor). Footnote 3 also says MXFP8 is limited to 10.x, but _compute_mxfp8_support() in transformer_engine/pytorch/quantization.py enables it from 10.0 up to, but not including, 12.0, which includes 11.0.

This follows up on #2556, where I posted a full L0 run on Jetson AGX Thor: #2556 (comment)

Depends on #3706. The 11.0 NVFP4 cell says "Yes" because stochastic rounding in the RHT cast-fusion kernels works on 11.0 only with that fix. Without it, the default NVFP4 recipe returns all-zero wgrad on 11.0. If #3706 does not land, that cell should be "Partial" with a footnote like the one for 12.x.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Compute capability table: add 11.0, Blackwell (Jetson), Jetson AGX Thor.
• Format table: add 11.0 with BF16, FP8 (per tensor), FP8 block scaling, MXFP8 and NVFP4 all "Yes".
• Footnote 3: MXFP8 is supported on 10.0 through 11.x.
• Footnote 4: the full NVFP4 recipe is supported on 10.x and 11.0.
• Default recipe table: 11.0 uses MXFP8BlockScaling.

Basis

Jetson AGX Thor, JetPack 7.2, CUDA 13.2, cuDNN 9.20, TE built from source with NVTE_CUDA_ARCHS=110.

• is_fp8_available(), is_fp8_block_scaling_available(), is_mxfp8_available() and is_nvfp4_available() all return True, and get_default_fp8_recipe() returns MXFP8BlockScaling.
• In the L0 PyTorch run linked above, the FP8, FP8 block-scaling and MXFP8 quantization and GEMM suites pass (for example test_mxfp8: 447 passed, 0 failed). Other MXFP8-related failures fall into two groups. Four are attention tests: MXFP8 FusedAttention needs cuDNN 9.21 or newer on any architecture, and JetPack 7.2 ships 9.20. Two are test_grouped_mlp cases for the CuTeDSL fused grouped MLP. TE enables that path only on 10.x, and its cuDNN-frontend kernel uses redux.f32, which sm_110a lacks (CuTeDSL grouped GEMM wrappers accept SM110 (Jetson Thor) but fail in ptxas: redux.f32 not supported on sm_110a cudnn-frontend#1686). Neither affects the recipe-level support this table describes.
• tests/pytorch/nvfp4: 9314 passed, 0 failed with [Common] Use software stochastic rounding in the RHT cast fusion kernels #3706; 15 failed without it.

I did not build the Sphinx docs. The change adds table rows in the existing format and edits two footnotes.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas (n/a)
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works (n/a, docs only)
• New and existing unit tests pass locally with my changes (n/a, docs only)

🤖 Generated with Claude Code