Ifu dev 20260803 v2.18 merge

.clang-format

@@ -261,7 +261,7 @@ SpacesInParensOptions:
   InEmptyParentheses: false
   Other:           false
 SpacesInSquareBrackets: false
-Standard:        Auto
+Standard:        c++17
 StatementAttributeLikeMacros:
   - Q_EMIT
 StatementMacros:
@@ -277,4 +277,3 @@ WhitespaceSensitiveMacros:
   - PP_STRINGIZE
   - STRINGIZE
 ...
-

.github/workflows/build.yml

@@ -23,7 +23,7 @@ jobs:
         run: |
           apt-get update
           apt-get install -y git python3.9 pip cudnn9-cuda-12
-          pip install cmake==3.21.0 pybind11[global] ninja
+          pip install cmake==3.21.0 pybind11[global] ninja "nvidia-cudnn-frontend>=1.25.0"
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
@@ -77,7 +77,7 @@ jobs:
           docker exec builder bash -c '\
             apt-get update && \
             apt-get install -y git python3.9 pip cudnn9-cuda-12 && \
-            pip install cmake torch ninja pydantic importlib-metadata>=1.0 packaging pybind11 numpy einops onnxscript && \
+            pip install cmake torch ninja pydantic importlib-metadata>=1.0 packaging pybind11 numpy einops onnxscript "nvidia-cudnn-frontend>=1.25.0" && \
             apt-get clean \
           '
 
@@ -96,7 +96,7 @@ jobs:
       options: --user root
     steps:
       - name: 'Dependencies'
-        run: pip install cmake==3.21.0 pybind11[global]
+        run: pip install cmake==3.21.0 pybind11[global] "nvidia-cudnn-frontend>=1.25.0"
       - name: 'Checkout'
         uses: actions/checkout@v3
         with:
@@ -148,7 +148,7 @@ jobs:
       - name: 'Dependencies'
         run: |
           docker exec builder bash -c '\
-            pip install cmake==3.21.0 pybind11[global] einops onnxscript && \
+            pip install cmake==3.21.0 pybind11[global] einops onnxscript "nvidia-cudnn-frontend>=1.25.0" && \
             pip install torch --no-cache-dir --index-url https://download.pytorch.org/whl/cu130
           '
       - name: 'Build'

.github/workflows/trigger-ci.yml

@@ -60,6 +60,8 @@ jobs:
            || github.actor == 'kainzhong'
            || github.actor == 'cspades'
            || github.actor == 'jomitchellnv'
+           || github.actor == 'fheinecke'
+           || github.actor == 'janekb04'
          )
     steps:
       - name: Check if comment is issued by authorized person

.gitmodules

@@ -8,9 +8,6 @@
 	path = examples/pytorch/minGPT
 	url = https://github.com/HaiShaw/minGPT.git
 	branch = transformer-engine
-[submodule "3rdparty/cudnn-frontend"]
-	path = 3rdparty/cudnn-frontend
-	url = https://github.com/NVIDIA/cudnn-frontend.git
 [submodule "3rdparty/aotriton"]
 	path = 3rdparty/aotriton
 	url = https://github.com/ROCm/aotriton.git

MANIFEST.in

@@ -1,2 +1,3 @@
 recursive-include transformer_engine/common/include *.*
+recursive-include transformer_engine/common/cmake *.cmake
 recursive-include build_tools *.py *.txt

README.rst

@@ -347,6 +347,11 @@ Transformer Engine
 Latest News
 ===========
 
+* [06/2026] `Boosting MoE Training Throughput with Advanced Fusion Kernels <https://developer.nvidia.com/blog/boosting-moe-training-throughput-with-advanced-fusion-kernels/>`_
+* [06/2026] `Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning <https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf>`_
+* [06/2026] `Train Models Faster with JAX and MaxText Using NVFP4 on NVIDIA Blackwell <https://developer.nvidia.com/blog/train-models-faster-with-jax-and-maxtext-using-nvfp4-on-nvidia-blackwell/>`_
+* [04/2026] `Run High-Throughput Reinforcement Learning Training with End-to-End FP8 Precision <https://developer.nvidia.com/blog/run-high-throughput-reinforcement-learning-training-with-end-to-end-fp8-precision/>`_
+* [02/2026] `Using NVFP4 Low-Precision Model Training for Higher Throughput Without Losing Accuracy <https://developer.nvidia.com/blog/using-nvfp4-low-precision-model-training-for-higher-throughput-without-losing-accuracy/>`_
 * [12/2025] `NVIDIA Nemotron 3: Efficient and Open Intelligence <https://arxiv.org/abs/2512.20856>`_ - trained with NVFP4 on Transformer Engine
 * [11/2025] `NVIDIA Blackwell Architecture Sweeps MLPerf Training v5.1 Benchmarks <https://developer.nvidia.com/blog/nvidia-blackwell-architecture-sweeps-mlperf-training-v5-1-benchmarks/>`_
 * [11/2025] `Scale Biology Transformer Models with PyTorch and NVIDIA BioNeMo Recipes <https://developer.nvidia.com/blog/scale-biology-transformer-models-with-pytorch-and-nvidia-bionemo-recipes/>`_

benchmarks/attention/benchmark_attention.py

@@ -23,8 +23,6 @@
 num_iters = 3
 # checkpointing
 ckpt_attn = False
-# workspace optimization path for cuDNN attention
-workspace_opt = True
 # QKV memory layout
 qkv_layout = "bshd_bshd_bshd"
 # padding between sequences for qkv_format=thd
@@ -59,7 +57,6 @@ def benchmark_dot_product_attention(model, fused_attn_supported, flash_attn_supp
                 "FusedAttention",
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )
@@ -70,7 +67,6 @@ def benchmark_dot_product_attention(model, fused_attn_supported, flash_attn_supp
                 "FlashAttention",
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )
@@ -90,7 +86,6 @@ def benchmark_dot_product_attention(model, fused_attn_supported, flash_attn_supp
                 "FusedAttention",
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )
@@ -107,7 +102,6 @@ def benchmark_dot_product_attention(model, fused_attn_supported, flash_attn_supp
                 "FlashAttention",
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )

benchmarks/attention/benchmark_attention_rocm.py

@@ -35,8 +35,6 @@
 num_iters = 10
 # Checkpointing attention
 ckpt_attn = False
-# Workspace optimization for attention
-workspace_opt = True
 # QKV memory layout
 qkv_layout = "bshd_bshd_bshd"
 # Padding between sequences for qkv_format=thd
@@ -211,7 +209,6 @@ def benchmark_dot_product_attention(model, attention, column_name, dirname):
                 attention,
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )
@@ -232,7 +229,6 @@ def benchmark_dot_product_attention_profiler(model, attention, column_name):
                 attention,
                 ckpt_attn,
                 qkv_layout,
-                workspace_opt,
                 pad_between_seqs,
                 is_training,
             )