[NVBUG-6448152][test] TEST ONLY native midpoint async-consensus discriminator

.claude/skills/exec-local-compile/SKILL.md

@@ -45,7 +45,7 @@ git checkout main && git pull
 Run the build command (**incremental by default** — omit `-c`/`--clean` unless explicitly requested or the incremental build fails):
 
 ```bash
-./scripts/build_wheel.py --use_ccache -a "<arch>" -f --nvtx
+./scripts/build_wheel.py --trt_root /usr/local/tensorrt --use_ccache -a "<arch>" -f --nvtx
 ```
 
 Replace `<arch>` with the target GPU architecture (see Architecture Reference below). If not specified by the user, auto-detect from `nvidia-smi`.
@@ -66,6 +66,7 @@ python3 -c "import tensorrt_llm; print(tensorrt_llm.__version__)"
 
 | Flag | Description |
 |------|-------------|
+| `--trt_root /usr/local/tensorrt` | TensorRT installation path (standard in NVIDIA containers) |
 | `-a "<arch>"` | Target GPU architecture(s) |
 | `--nvtx` | Enable NVTX markers for profiling |
 | `--use_ccache` | Use ccache for faster recompilation |

.claude/skills/exec-slurm-compile/SKILL.md

@@ -204,6 +204,7 @@ A successful build ends with a message like `Successfully built tensorrt_llm` or
 
 | Flag | Description |
 |------|-------------|
+| `--trt_root /usr/local/tensorrt` | TensorRT installation path (standard in NVIDIA containers) |
 | `-a "100-real"` | Target architecture — `100` for Blackwell, `90` for Hopper, etc. |
 | `--nvtx` | Enable NVTX markers for profiling |
 | `--no-venv` | Skip virtual environment creation |
@@ -226,6 +227,7 @@ Common architecture values:
 | `sbatch: error: invalid partition` | Verify partition name with `sinfo -s` |
 | `sbatch: error: invalid account` | Check available accounts with `sacctmgr show assoc user=$USER` |
 | Container image not found | Verify the `.sqsh` path exists and is readable |
+| Build fails with missing TensorRT | Ensure `--trt_root` points to the correct path inside the container |
 | Build OOM (out of memory) | Reduce parallelism with `-j <N>` flag to `build_wheel.py` |
 | `srun: error: Unable to create step` | The node may lack enroot/pyxis — check with cluster admin |
 | Job stuck in `PD` state | Check `squeue -j <id> -o %R` for the reason (e.g., resource limits, priority) |

.claude/skills/exec-slurm-compile/scripts/compile.sh

@@ -19,7 +19,7 @@
 # Usage: compile.sh <repo_dir> [build_wheel_args...]
 #
 # Default build_wheel.py flags:
-#   -a "100-real" --nvtx --no-venv
+#   --trt_root /usr/local/tensorrt -a "100-real" --nvtx --no-venv
 # Any extra arguments after repo_dir are forwarded to build_wheel.py,
 # overriding the defaults above.
 
@@ -36,6 +36,7 @@ if [[ $# -gt 0 ]]; then
 else
     echo "[compile.sh] Running default build command"
     python3 ./scripts/build_wheel.py \
+        --trt_root /usr/local/tensorrt \
         -a "100-real" \
         --nvtx
 fi

.coderabbit.yaml

@@ -22,26 +22,7 @@ reviews:
   auto_title_placeholder: '@coderabbitai title'
   auto_title_instructions: 'Format: "[<category>] <title>". Category must be one of: fix, feat, doc, infra, style, refactor, perf, test, chore, revert. Enclose the category in square brackets. Title should be concise (<= 60 chars). Example: "[feat] Add logit_bias support".'
   commit_status: false
-  high_level_summary_instructions: |
-    Always produce two review sections in the summary:
-
-    **Dev Engineer Review**
-    Review all changes for correctness and consistency, including:
-    - Code changes: correctness, performance, API consistency (CODING_GUIDELINES.md), error handling, regressions.
-    - Config files: valid values, no typos, consistency with related configs, no unintended scope changes.
-    - Test list files (test-db/, qa/, waives.txt): correct format, valid test paths, appropriate bug references, no duplicates.
-
-    **QA Engineer Review**
-    Always include this section when any files under tests/ are touched.
-    For test-list-only changes (only tests/integration/test_lists/ files):
-      - List which test-db/ or qa/ files were modified and what entries were added or removed.
-      - Verdict: "needs follow-up" if CBTS coverage data is unavailable, otherwise "sufficient" or "insufficient".
-    For test-code changes (files outside tests/integration/test_lists/):
-      - List test functions added, modified, or removed.
-      - State whether each is covered in tests/integration/test_lists/ (test-db/ for CI, qa/ for manual QA).
-      - Verdict: sufficient, insufficient, or needs follow-up.
-    If no test files are touched, write "No test changes."
-  collapse_walkthrough: false
+  collapse_walkthrough: true
   assess_linked_issues: true
   related_issues: true
   related_prs: true
@@ -50,27 +31,15 @@ reviews:
   poem: false
   review_status: false
   auto_review:
-    auto_incremental_review: true
+    auto_incremental_review: false
     drafts: false
     base_branches: ["main", "release/.+"]
   path_instructions:
     - path: "tests/**"
       instructions: |
         Act as a QA engineer reviewing test changes and coverage for TensorRT-LLM.
-        Always produce a test coverage summary, even if no issues are found.
-
-        If the change touches ONLY files under tests/integration/test_lists/ (no test-code changes):
-        - Report which test-db/ or qa/ list files were modified and what entries were added or removed.
-        - Do NOT require changed test functions for this path.
-        - Use verdict "needs follow-up" when cbts_touchmap.sqlite or a CBTS coverage report is unavailable
-          to confirm the impacted test scope; otherwise use "sufficient" or "insufficient".
-
-        If the change includes test-code files (outside tests/integration/test_lists/), the summary must include:
-        1. Which test functions were added, modified, or removed.
-        2. Whether each changed test is listed in the appropriate test list files under
-           tests/integration/test_lists/ (test-db/ for CI, qa/ for manual QA).
-        3. A coverage verdict: sufficient, insufficient, or needs follow-up.
-        Keep feedback actionable: reference concrete list file names when suggesting additions.
+        Keep feedback actionable: suggest concrete list file names and whether
+        coverage is sufficient, insufficient, or needs follow-up outside the PR.
     - path: "tests/integration/test_lists/qa/**"
       instructions: |
         Files here are manually-triggered QA perf/regression lists, maintained

.gitattributes

@@ -16,5 +16,6 @@ docs/source/blogs/media/tech_blog10_full_strategy_performance.png filter=lfs dif
 docs/source/blogs/media/tech_blog10_context_wait_performance.png  filter=lfs diff=lfs merge=lfs -text
 cpp/tensorrt_llm/kernels/trtllmGenKernels/fmha/cubin/kernelMetaInfo_cubin.cpp filter=lfs diff=lfs merge=lfs -text
 cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/cubin/xqa_kernel_cubin.cpp filter=lfs diff=lfs merge=lfs -text
+tensorrt_llm/_torch/visual_gen/cute_dsl_kernels/blackwell/attention/cubins/*/*/*.so filter=lfs diff=lfs merge=lfs -text
 docs/source/blogs/media/tech_blog26_deepseek_v4_hybrid_attention.png filter=lfs diff=lfs merge=lfs -text
 docs/source/blogs/media/tech_blog26_deepseek_v4_mhc_moe.png filter=lfs diff=lfs merge=lfs -text

.github/CODEOWNERS

@@ -437,7 +437,6 @@
 /cpp/tensorrt_llm/batch_manager/allocateKvCache.cpp @NVIDIA/trt-llm-kv-cache-manager-devs
 /cpp/tests/unit_tests/batch_manager/kvCacheManagerTest.cpp @NVIDIA/trt-llm-kv-cache-manager-devs
 /cpp/tests/unit_tests/batch_manager/kvCacheUtilsTest.cpp @NVIDIA/trt-llm-kv-cache-manager-devs
-/tensorrt_llm/_torch/attention_backend/sparse/*/cache_manager.py @NVIDIA/trt-llm-kv-cache-manager-devs
 /tensorrt_llm/_torch/pyexecutor/kv_cache_manager_v2.py @NVIDIA/trt-llm-kv-cache-manager-devs
 /tensorrt_llm/_torch/pyexecutor/resource_manager.py @NVIDIA/trt-llm-kv-cache-manager-devs
 /cpp/tensorrt_llm/nanobind/batch_manager/kvCacheManager.h @NVIDIA/trt-llm-kv-cache-manager-devs

.github/workflows/bot-command.yml

@@ -52,14 +52,14 @@ jobs:
             "`--disable-reuse-test ` *(OPTIONAL)* : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.\n\n" +
             "`--disable-fail-fast ` *(OPTIONAL)* : Disable fail fast on build/tests/infra failures.\n\n" +
             "`--skip-test ` *(OPTIONAL)* : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does **NOT** update GitHub check status.\n\n" +
-            "`--stage-list \"A10-PyTorch-1, xxx\"` *(OPTIONAL)* : Only run the specified test stages. Supports wildcard `*` for pattern matching (e.g., `\"*PerfSanity*\"` matches all stages containing PerfSanity). Examples: \"A10-PyTorch-1, xxx\", \"*PerfSanity*\". The patterns `\"*\"`, `\"*Post-Merge*\"`, and `\"*PerfSanity*\"`, including equivalent escaped or repeated-star forms and their use in comma-separated lists, require the `ci: post-merge approved` PR label. Note: Does **NOT** update GitHub check status.\n\n" +
+            "`--stage-list \"A10-PyTorch-1, xxx\"` *(OPTIONAL)* : Only run the specified test stages. Supports wildcard `*` for pattern matching (e.g., `\"*PerfSanity*\"` matches all stages containing PerfSanity). Examples: \"A10-PyTorch-1, xxx\", \"*PerfSanity*\". The broad patterns `\"*\"` and `\"*Post-Merge*\"`, including equivalent escaped or repeated-star forms and their use in comma-separated lists, require the `ci: post-merge approved` PR label. Note: Does **NOT** update GitHub check status.\n\n" +
             "`--gpu-type \"A30, H100_PCIe\"` *(OPTIONAL)* : Only run the test stages on the specified GPU types. Examples: \"A30, H100_PCIe\". Note: Does **NOT** update GitHub check status.\n\n" +
             "`--test-backend \"pytorch, cpp\"` *(OPTIONAL)* : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: \"pytorch, cpp\" (does not run test stages with tensorrt or triton backend). Note: Does **NOT** update GitHub pipeline status.\n\n" +
             "`--only-multi-gpu-test ` *(OPTIONAL)* : Only run the multi-GPU tests. Note: Does **NOT** update GitHub check status.\n\n" +
             "`--disable-multi-gpu-test ` *(OPTIONAL)* : Disable the multi-GPU tests. Note: Does **NOT** update GitHub check status.\n\n" +
             "`--add-multi-gpu-test ` *(OPTIONAL)* : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.\n\n" +
             "`--post-merge ` *(OPTIONAL)* : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline. Requires the `ci: post-merge approved` PR label applied by an active member of `NVIDIA/trt-llm-ci-approvers`. The approval label remains in place when new commits are pushed.\n\n" +
-            "`--extra-stage \"H100_PCIe-TensorRT-Post-Merge-1, xxx\"` *(OPTIONAL)* : Run the ordinary L0 pre-merge pipeline and specified test stages. Supports wildcard `*` for pattern matching. Examples: --extra-stage \"H100_PCIe-TensorRT-Post-Merge-1, xxx\", --extra-stage \"*Post-Merge*\". The patterns `\"*\"`, `\"*Post-Merge*\"`, and `\"*PerfSanity*\"`, including equivalent escaped or repeated-star forms and their use in comma-separated lists, require the `ci: post-merge approved` PR label.\n\n" +
+            "`--extra-stage \"H100_PCIe-TensorRT-Post-Merge-1, xxx\"` *(OPTIONAL)* : Run the ordinary L0 pre-merge pipeline and specified test stages. Supports wildcard `*` for pattern matching. Examples: --extra-stage \"H100_PCIe-TensorRT-Post-Merge-1, xxx\", --extra-stage \"*Post-Merge*\". The broad patterns `\"*\"` and `\"*Post-Merge*\"`, including equivalent escaped or repeated-star forms and their use in comma-separated lists, require the `ci: post-merge approved` PR label.\n\n" +
             "`--detailed-log ` *(OPTIONAL)* : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.\n\n" +
             "`--debug ` *(OPTIONAL)* : **Experimental feature**. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in the `stage-list` parameter to access the appropriate container environment. Note: Does **NOT** update GitHub check status.\n\n" +
             "`--high-priority ` *(OPTIONAL)* : Run the pipeline with high priority. This option is restricted to authorized users only and will route the job to a high-priority queue.\n\n" +

.github/workflows/lfs-sync.yml

@@ -44,12 +44,9 @@ jobs:
 
     steps:
       - name: Checkout merge commit without LFS smudge
-        uses: actions/checkout@v6
+        uses: actions/checkout@v4
         with:
           ref: ${{ github.event.pull_request.merge_commit_sha }}
-          # This job only handles merged PRs and does not execute code from the
-          # checked-out tree, so explicitly allow the required merge checkout.
-          allow-unsafe-pr-checkout: true
           lfs: false
           fetch-depth: 2
           token: ${{ secrets.GITHUB_TOKEN }}
@@ -203,7 +200,7 @@ jobs:
 
       - name: Comment on PR — already in storage
         if: steps.detect-lfs.outputs.lfs_status == 'present'
-        uses: actions/github-script@v8
+        uses: actions/github-script@v7
         with:
           script: |
             const files = require('fs').readFileSync('/tmp/lfs_files.txt', 'utf8').trim().split('\n');
@@ -224,7 +221,7 @@ jobs:
 
       - name: Comment on PR — sync succeeded
         if: steps.detect-lfs.outputs.lfs_status == 'pending' && steps.verify.outcome == 'success'
-        uses: actions/github-script@v8
+        uses: actions/github-script@v7
         with:
           script: |
             const files = require('fs').readFileSync('/tmp/lfs_files.txt', 'utf8').trim().split('\n');
@@ -245,7 +242,7 @@ jobs:
 
       - name: Comment on PR — sync failed
         if: always() && steps.detect-lfs.outputs.lfs_status == 'pending' && (steps.fetch-lfs.outcome != 'success' || steps.push-lfs.outcome != 'success' || steps.verify.outcome != 'success')
-        uses: actions/github-script@v8
+        uses: actions/github-script@v7
         with:
           script: |
             const forkRepo = context.payload.pull_request.head.repo.full_name;

.gitignore

@@ -57,6 +57,7 @@ tensorrt_llm/flash_mla/
 tensorrt_llm/flash_mla_cpp_tllm.*.so
 tensorrt_llm/flash_mla_cpp_tllm.pyi
 tensorrt_llm/runtime/kv_cache_manager_v2/**/*.so
+!tensorrt_llm/_torch/visual_gen/cute_dsl_kernels/blackwell/attention/cubins/**/*.so
 **/*__mypyc*.so
 tensorrt_llm/scripts
 *docs/cpp_docs*

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "3rdparty/MSA"]
-	path = 3rdparty/MSA
-	url = https://gitlab.com/nvidia/tensorrt-llm/oss-components/msa.git

ATTRIBUTIONS-Python.md

@@ -5261,7 +5261,7 @@ For more information, please refer to <http://unlicense.org>
   - `Tracker`: https://github.com/tox-dev/py-filelock/issues
 
 
-## flashinfer-python (0.6.15)
+## flashinfer-python (0.6.14)
 
 ### Licenses
 License: `Apache-2.0`

LICENSE

@@ -19,13 +19,6 @@ Original Source: https://github.com/Dao-AILab/causal-conv1d
 Copyright (c) 2024, Tri Dao.
 Licensed under the BSD 3-Clause License
 
---------------------------------------------------------------------------------
-CUTLASS
---------------------------------------------------------------------------------
-Original Source: https://github.com/NVIDIA/cutlass
-Copyright (c) 2017 - 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-Licensed under the BSD 3-Clause License
-
 --------------------------------------------------------------------------------
 flash-attention
 --------------------------------------------------------------------------------
@@ -41,14 +34,6 @@ Original Source: https://github.com/fla-org/flash-linear-attention
 Copyright (c) 2023-2025 Songlin Yang
 Licensed under the MIT License
 
---------------------------------------------------------------------------------
-FlashInfer
---------------------------------------------------------------------------------
-Original Source: https://github.com/flashinfer-ai/flashinfer
-Copyright 2025-2026 NVIDIA
-Copyright 2023-2026 FlashInfer community (https://flashinfer.ai/)
-Licensed under the Apache License 2.0
-
 --------------------------------------------------------------------------------
 InstructEval
 --------------------------------------------------------------------------------
@@ -74,14 +59,6 @@ Original Source: https://github.com/state-spaces/mamba
 Copyright 2023 Tri Dao, Albert Gu
 Licensed under the Apache License 2.0
 
---------------------------------------------------------------------------------
-MSA (MiniMax Sparse Attention)
---------------------------------------------------------------------------------
-Original Source: https://github.com/MiniMax-AI/MSA
-Copyright (c) 2026 MiniMax
-Licensed under the MIT License
-
-
 --------------------------------------------------------------------------------
 Quack
 --------------------------------------------------------------------------------

README.md

@@ -10,7 +10,7 @@ TensorRT LLM
 [![python](https://img.shields.io/badge/python-3.10-green)](https://www.python.org/downloads/release/python-31012/)
 [![cuda](https://img.shields.io/badge/cuda-13.2.1-green)](https://developer.nvidia.com/cuda-downloads)
 [![torch](https://img.shields.io/badge/torch-2.11.0-green)](https://pytorch.org)
-[![version](https://img.shields.io/badge/release-1.3.0rc23-green)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/version.py)
+[![version](https://img.shields.io/badge/release-1.3.0rc21-green)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/version.py)
 [![license](https://img.shields.io/badge/license-Apache%202-blue)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/LICENSE)
 
 [Architecture](https://nvidia.github.io/TensorRT-LLM/developer-guide/overview.html)   |   [Performance](https://nvidia.github.io/TensorRT-LLM/developer-guide/perf-overview.html)   |   [Examples](https://nvidia.github.io/TensorRT-LLM/quick-start-guide.html)   |   [Documentation](https://nvidia.github.io/TensorRT-LLM/)   |   [Roadmap](https://github.com/NVIDIA/TensorRT-LLM/issues?q=is%3Aissue%20state%3Aopen%20label%3Aroadmap)

cpp/include/tensorrt_llm/batch_manager/cacheTransceiver.h

@@ -27,6 +27,7 @@
 #include "tensorrt_llm/executor/dataTransceiverState.h"
 #include "tensorrt_llm/runtime/utils/mpiUtils.h"
 #include "tensorrt_llm/runtime/utils/pgUtils.h"
+#include <cstddef>
 #include <future>
 #include <memory>
 #include <mutex>
@@ -55,6 +56,7 @@ class BaseKVCacheManager;
 
 class CacheSender;
 class CacheReceiver;
+class ContextTransferCoordinator;
 
 class CacheTransceiverComm
 {
@@ -149,6 +151,25 @@ class CacheTransceiverComm
         TLLM_THROW("Input arguments only supported in mpi");
     }
 
+    [[nodiscard]] std::unique_ptr<mpi::MpiRequest> sendAsync(
+        void const* buffer, std::size_t size, mpi::MpiType dtype, int dest, mpi::MpiTag tag) const
+    {
+        TLLM_CHECK_WITH_INFO(isMpi(), "Point-to-point cache-transceiver status messages require MPI.");
+        return mMpiComm->sendAsync(buffer, size, dtype, dest, tag);
+    }
+
+    [[nodiscard]] bool iprobe(int source, mpi::MpiTag tag, MPI_Status* status) const
+    {
+        TLLM_CHECK_WITH_INFO(isMpi(), "Point-to-point cache-transceiver status messages require MPI.");
+        return mMpiComm->iprobe(source, tag, status);
+    }
+
+    void recv(void* buffer, std::size_t size, mpi::MpiType dtype, int source, mpi::MpiTag tag) const
+    {
+        TLLM_CHECK_WITH_INFO(isMpi(), "Point-to-point cache-transceiver status messages require MPI.");
+        static_cast<void>(mMpiComm->recv(buffer, size, dtype, source, tag));
+    }
+
     CacheTransceiverComm split(int color, int key)
     {
         if (isMpi())
@@ -307,6 +328,7 @@ class CacheTransceiver : public BaseCacheTransceiver
 
     std::shared_ptr<CacheTransceiverComm> mGroupComm;
     std::shared_ptr<CacheTransceiverComm> mGroupTensorParaComm, mGroupPipeParaComm, mGroupDataComm, mGroupTPInDPComm;
+    std::unique_ptr<ContextTransferCoordinator> mContextTransferCoordinator;
 
     executor::kv_cache::CommState const* mCommState;
     std::unique_ptr<executor::kv_cache::CacheState> mCacheState;