Skip to content

Add public GELU-MLP and a ModelOpt-anchored Qwen-Image NVFP4 proxy #695

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
YangXu1990uiuc wants to merge 5 commits into
base: develop
Choose a base branch
from
Open

Add public GELU-MLP and a ModelOpt-anchored Qwen-Image NVFP4 proxy #695

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
dd6d2c8
gemm: add a dense GELU MLP torch op
YangXu1990uiuc Aug 21, 2026
2d65aef
benchmark: add a Qwen-Image MLP and SDPA matrix
YangXu1990uiuc Aug 21, 2026
15665c1
benchmark: add a ModelOpt-anchored Qwen-Image NVFP4 path
YangXu1990uiuc Aug 21, 2026
f5e5db8
Merge current develop and address GELU/NVFP4 review feedback
YangXu1990uiuc Sep 9, 2026
50b5342
fix(bench): separate sampling protocols from GPU product policy
YangXu1990uiuc Sep 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions .gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ cover/
# Local end-to-end benchmark artifacts
qwen3.8-factorial-results/
qwen-image-bf16-results/
qwen-image-nvfp4-results/

# Translations
*.mo
Expand Down
1 change: 1 addition & 0 deletions AGENTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,6 +82,7 @@ First invocation builds the hook environments and can take >5 minutes; later run

## Conventions

- Keep benchmark collection policy separate from runtime support checks: a request to measure on a particular full-SM GPU constrains the agent's measurement/reporting, not user-facing scripts. Gate on actual kernel capabilities, record device/SM-count metadata, and cover a supported non-matching GPU name and SM count in CPU-only device-selection tests.
- `include/` is header-only: no `.cpp` files, no new required dependencies. Vendored third-party code lives in `include/cudnn_frontend/thirdparty/`.
- Every new frontend-only Python API needs: `APIBase` subclass + wrapper, lazy export in `python/cudnn/__init__.py`, docs under `docs/fe-oss-apis/`, and pytest coverage under `test/python/fe_api/`. Full recipe: [python/cudnn/AGENTS.md](python/cudnn/AGENTS.md) and the `cutedsl-kernel-integration` skill.
- Frontend-only OSS APIs are experimental; keep the lazy-import boundary intact (no eager `torch`/`cutlass` imports at `cudnn` import time). CuTeDSL is a required dependency now, but a tensor framework is not, and `import cudnn` still has to stay cheap.
Expand Down
18 changes: 18 additions & 0 deletions THIRD_PARTY_LICENSES.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -302,3 +302,21 @@ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

-------------------------------------------------------------------------------
11. FlashInfer
https://github.com/flashinfer-ai/flashinfer

The benchmark-private SM100 NVFP4 activation quantizer under
python/cudnn/gemm/ops/csrc/ is derived from FlashInfer's
include/flashinfer/gemm/nvfp4_smooth_quantize_sm100.cuh at commit
f212ec8230486e3615502b8af75fe7022c60b2f3. The upstream source in turn
credits NVIDIA TensorRT-LLM's nvfp4SmoothQuantize kernel and quantization
helpers. The derived source retains its NVIDIA copyright and provenance.
Local changes include provenance, formatting annotations, and host argument validation.

FlashInfer
Copyright 2025-2026 NVIDIA
Copyright 2023-2026 FlashInfer community (https://flashinfer.ai/)

License: Apache License 2.0 (same text as this repository's LICENSE.txt)
Loading
Loading