Repository navigation
Conversation
This was referenced Nov 10, 2026
pggPL
force-pushed
the
jax_moe_fused_situ
branch
from
November 12, 2026 13:11
2f155f6 to
4c7ac44
Compare
8 of 13 tasks
Select the shared cuDNN grouped GEMM SiTU forward/backward APIs when their JAX bridge exposes both soft-cap parameters. Preserve fallback for incompatible frontends and devices, and pass device expert offsets through the existing fused path. [skip ci] Signed-off-by: Pawel Gadzinski <pgadzinski@nvidia.com>
pggPL
force-pushed
the
jax_moe_fused_situ
branch
from
November 12, 2026 13:49
4c7ac44 to
ab2ae5e
Compare
5 tasks done
pggPL
marked this pull request as ready for review
November 14, 2026 14:08
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Build on #75 (stack: #73 → #74 → #75, with shared vectorized GLU templates in #74) to run SiTU-GLU through cuDNN's existing fused grouped GEMM/MXFP8 forward and backward kernels. Device expert offsets bound active receive rows, including overallocated buffers. Requires the cuDNN Frontend JAX bridge change merged into
developon 2026-11-14: NVIDIA/cudnn-frontend#2540 SiTU fusion needs cuDNN Frontend 1.33.0 or adevelopbuild at or after 30201756 (first nightly:1.33.0.dev82169718); the general fusion floor stays 1.32.0. The eligibility check inspects the installed JAX entry signatures forsitu_beta1/situ_beta2, so older 1.33.0 dev builds fall back to the unfused path instead of failing.Validated the rebuilt stack on four GB200 GPUs: 54 standalone grouped SiTU tests, 36 resource/API tests, 40 FSDP/layout cases, and 26 QB router tests. The four-rank MoE run passed 38 cases per rank, including 12 complete fused forward/backward configurations covering SwiGLU controls, SiTU 4/25 and 2/8, standard/native weight layouts, and early/late routing weights. Twelve cases were skipped because the unfused fixture intentionally disables fusion. The native grouped fallback now reuses #74's shared GLU templates; its common regression sweep passed 212 C++ cases, plus 22 PyTorch and 22 existing JAX activation cases. Existing numerical tolerances are unchanged. Pre-commit, source license checks, C++ lint and the full JAX Python lint suite passed (10.00/10). No performance improvement is claimed; SM103 runtime and full Kimi-K3 training remain unverified.
Type of change
Changes
Checklist: