Ifu dev 20260828 v2.19

.github/workflows/build.yml

@@ -16,13 +16,16 @@ jobs:
     name: 'Core'
     runs-on: ubuntu-latest
     container:
-      image: nvcr.io/nvidia/cuda:12.1.0-devel-ubuntu22.04
+      image: nvcr.io/nvidia/cuda:12.9.2-devel-ubuntu22.04
       options: --user root
     steps:
       - name: 'Dependencies'
         run: |
           apt-get update
-          apt-get install -y git python3.9 pip cudnn9-cuda-12
+          apt-mark unhold libnccl2 libnccl-dev
+          apt-get install -y git python3 python3-pip cudnn9-cuda-12 \
+            libnccl2=2.30.7-1+cuda12.9 \
+            libnccl-dev=2.30.7-1+cuda12.9
           pip install cmake==3.21.0 pybind11[global] ninja "nvidia-cudnn-frontend>=1.25.0"
       - name: 'Checkout'
         uses: actions/checkout@v3
@@ -70,13 +73,16 @@ jobs:
 
       - name: Start named container
         run: |
-          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d nvcr.io/nvidia/cuda:12.8.0-devel-ubuntu22.04 sleep infinity
+          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d nvcr.io/nvidia/cuda:12.9.2-devel-ubuntu22.04 sleep infinity
 
       - name: 'Dependencies'
         run: |
           docker exec builder bash -c '\
             apt-get update && \
-            apt-get install -y git python3.9 pip cudnn9-cuda-12 && \
+            apt-mark unhold libnccl2 libnccl-dev && \
+            apt-get install -y git python3 python3-pip cudnn9-cuda-12 \
+              libnccl2=2.30.7-1+cuda12.9 \
+              libnccl-dev=2.30.7-1+cuda12.9 && \
             pip install cmake torch ninja pydantic importlib-metadata>=1.0 packaging pybind11 numpy einops onnxscript "nvidia-cudnn-frontend>=1.25.0" && \
             apt-get clean \
           '
@@ -92,7 +98,7 @@ jobs:
     name: 'JAX'
     runs-on: ubuntu-latest
     container:
-      image: ghcr.io/nvidia/jax:jax
+      image: ghcr.io/nvidia/jax:jax-2026-07-21
       options: --user root
     steps:
       - name: 'Dependencies'
@@ -143,7 +149,7 @@ jobs:
 
       - name: Start named container
         run: |
-          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d ghcr.io/nvidia/jax:jax sleep infinity
+          docker run -v $(pwd):$(pwd) -w $(pwd) --name builder -d ghcr.io/nvidia/jax:jax-2026-07-21 sleep infinity
 
       - name: 'Dependencies'
         run: |

.github/workflows/trigger-ci.yml

@@ -62,6 +62,8 @@ jobs:
            || github.actor == 'jomitchellnv'
            || github.actor == 'fheinecke'
            || github.actor == 'janekb04'
+           || github.actor == 'YangFei1990'
+           || github.actor == 'sraman-rgb'
          )
     steps:
       - name: Check if comment is issued by authorized person

.gitmodules

@@ -30,6 +30,6 @@
 [submodule "3rdparty/ck_jit"]
 	path = 3rdparty/ck_jit
 	url = https://github.com/ROCm/ck-jit.git
-[submodule "3rdparty/nccl"]
-	path = 3rdparty/nccl
-	url = https://github.com/NVIDIA/nccl.git
+[submodule "3rdparty/nccl-extensions"]
+	path = 3rdparty/nccl-extensions
+	url = https://github.com/NVIDIA/nccl-extensions.git

CONTRIBUTING.rst

@@ -13,7 +13,7 @@ Coding Guidelines
 -----------------
 
 * We follow `Google C++ Style Guide <https://google.github.io/styleguide/cppguide.html>`_. When no
-  rules can be found, follow the already occuring conventions. If there is no precedence in our
+  rules can be found, follow the already occurring conventions. If there is no precedence in our
   codebase we are open to discussion.
 * Prior to your contribution, please make sure that the code passes the linter check. We do both C++
   and Python linting. To invoke the check, please use

README.rst

@@ -510,7 +510,7 @@ System Requirements
   * Compiler: GCC 9+ or Clang 10+ with C++17 support
   * Python: 3.12 recommended
 
-* **Source Build Requirements:** CMake 3.18+, Ninja, Git 2.17+, pybind11 2.6.0+
+* **Source Build Requirements:** CMake 3.18+, Ninja, Git 2.17+, pybind11 2.6.0+, nvidia-cudnn-frontend 1.25.0+
 
 * **Notes:** FP8 features require Compute Capability 8.9+ (Ada/Hopper/Blackwell)
 

benchmarks/linear/benchmark_graph_safe_grouped_linear.py → benchmarks/linear/benchmark_graph_safe_grouped_mlp.py

@@ -13,21 +13,21 @@
 
 Example:
 
-    python benchmarks/linear/benchmark_graph_safe_grouped_linear.py
+    python benchmarks/linear/benchmark_graph_safe_grouped_mlp.py
 
 Forward-only:
 
-    python benchmarks/linear/benchmark_graph_safe_grouped_linear.py --fwd-only
+    python benchmarks/linear/benchmark_graph_safe_grouped_mlp.py --fwd-only
 
 Nsight Systems:
 
     (optionally: unset DEBUGINFOD_URLS)
 
     nsys profile \
-        --output=./benchmarks/linear/graph_safe_grouped_linear_mxfp8 \
+        --output=./benchmarks/linear/graph_safe_grouped_mlp_mxfp8 \
         --force-overwrite true \
         --trace=cuda,nvtx,cudnn,cublas \
-        python benchmarks/linear/benchmark_graph_safe_grouped_linear.py --profile
+        python benchmarks/linear/benchmark_graph_safe_grouped_mlp.py --profile
 """
 
 # Match the Qwen MXFP8 SFT launch toggles before importing TE.