Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
fece87e
[None][feat] Add Mistral Large 4 support
yizhang-nv Oct 29, 2026
942c709
[TRTLLM-17383][infra] Download Boost from AWS Artifactory (#21133)
sharabeshj-nvda Oct 28, 2026
001128a
[None][infra] PLC Pipeline update (#19557)
yuanjingx87 Oct 26, 2026
5583189
[TRTLLMINF-550][infra] Pull CUDA 13 PyTorch wheels from the cu132 mir…
sharabeshj-nvda Oct 28, 2026
cf7ae79
[None][chore] Set the Mistral Large 4 side-branch version to 1.4.0a0 …
yizhang-nv Oct 30, 2026
ed21308
[None][fix] Enforce container wheel version and source commit (#20564)
chzblych Oct 20, 2026
33d21e9
[None][fix] match nightly release BOLT behavior to postmerge (#20612)
tburt-nv Oct 21, 2026
331ae0c
[TRTLLMINF-342][fix] Unify Docker image orchestration and nightly whe…
niukuo Oct 22, 2026
9df1a00
[None][fix] Keep requirements installable on Python 3.10
EmmaQiaoCh Oct 31, 2026
5f00308
[None][doc] Mistral Large 4: install from the release container or wh…
yizhang-nv Nov 2, 2026
bed3750
[None][refactor] Mistral Large 4: own the checkpoint weight loader an…
yizhang-nv Nov 4, 2026
3fb4be0
[None][chore] Mistral Large 4: remove defensive checks and leftovers …
yizhang-nv Nov 4, 2026
3030ff5
[None][feat] Add VR200 support for Mistral Large 4 modeling_v2 (#21604)
yt0428 Nov 6, 2026
1295b9d
[None][test] Add Mistral Large 4 GB300 perf CI coverage (#22056)
yt0428 Nov 11, 2026
061ec01
[None][perf] Overlap the Mistral Large 4 SM107 decode shared expert w…
yt0428 Nov 16, 2026
cf9dd3f
[None][perf] Add a global split to sparse MLA decode for low-row deco…
yt0428 Nov 16, 2026
3dfa750
[None][perf] Use a direct BF16 GEMM for Mistral Large 4 narrow decode…
yt0428 Nov 16, 2026
3976d54
[None][perf] Use a split radix Top-K for Mistral Large 4 DSpark verif…
yt0428 Nov 16, 2026
416b15a
[None][perf] Fuse FP8 quantization and GEMM for Mistral Large 4 low-M…
yt0428 Nov 16, 2026
2eaa79c
[None][test] Route the native DSpark indexer test through the target'…
yt0428 Nov 17, 2026
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
3 changes: 2 additions & 1 deletion 3rdparty/fetch_content.json
Original file line number Diff line number Diff line change
Expand Up @@ -59,7 +59,8 @@
"git_repository": "https://github.com/deepseek-ai/FlashMLA.git",
"git_tag": "15f13e5030374295491c5ce31b02d7e63a7772c6",
"git_submodules_recurse": true,
"source_subdir": "dont-add-this-project-with-add-subdirectory"
"source_subdir": "dont-add-this-project-with-add-subdirectory",
"patch_file": "patches/flashmla_release_gil.patch"
},
{
"name": "fmt",
Expand Down
17 changes: 17 additions & 0 deletions 3rdparty/patches/flashmla_release_gil.patch
Original file line number Diff line number Diff line change
@@ -0,0 +1,17 @@
--- a/csrc/api/api.cpp
+++ b/csrc/api/api.cpp
@@ -7,9 +7,9 @@

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
m.doc() = "FlashMLA";
- m.def("sparse_decode_fwd", &sparse_attn_decode_interface);
- m.def("dense_decode_fwd", &dense_attn_decode_interface);
- m.def("sparse_prefill_fwd", &sparse_attn_prefill_interface);
- m.def("dense_prefill_fwd", &FMHACutlassSM100FwdRun);
- m.def("dense_prefill_bwd", &FMHACutlassSM100BwdRun);
+ m.def("sparse_decode_fwd", &sparse_attn_decode_interface, pybind11::call_guard<pybind11::gil_scoped_release>());
+ m.def("dense_decode_fwd", &dense_attn_decode_interface, pybind11::call_guard<pybind11::gil_scoped_release>());
+ m.def("sparse_prefill_fwd", &sparse_attn_prefill_interface, pybind11::call_guard<pybind11::gil_scoped_release>());
+ m.def("dense_prefill_fwd", &FMHACutlassSM100FwdRun, pybind11::call_guard<pybind11::gil_scoped_release>());
+ m.def("dense_prefill_bwd", &FMHACutlassSM100BwdRun, pybind11::call_guard<pybind11::gil_scoped_release>());
}
2 changes: 1 addition & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -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-13.4.1-green)](https://developer.nvidia.com/cuda-downloads)
[![torch](https://img.shields.io/badge/torch-2.14.0-green)](https://pytorch.org)
[![version](https://img.shields.io/badge/release-1.4.0rc1-green)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/version.py)
[![version](https://img.shields.io/badge/release-1.4.0a0-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)&nbsp;&nbsp;&nbsp;|&nbsp;&nbsp;&nbsp;[Performance](https://nvidia.github.io/TensorRT-LLM/developer-guide/perf-overview.html)&nbsp;&nbsp;&nbsp;|&nbsp;&nbsp;&nbsp;[Examples](https://nvidia.github.io/TensorRT-LLM/quick-start-guide.html)&nbsp;&nbsp;&nbsp;|&nbsp;&nbsp;&nbsp;[Documentation](https://nvidia.github.io/TensorRT-LLM/)&nbsp;&nbsp;&nbsp;|&nbsp;&nbsp;&nbsp;[Roadmap](https://github.com/NVIDIA/TensorRT-LLM/issues?q=is%3Aissue%20state%3Aopen%20label%3Aroadmap)
Expand Down
14 changes: 13 additions & 1 deletion docker/Dockerfile.multi
Original file line number Diff line number Diff line change
Expand Up @@ -171,6 +171,7 @@ COPY setup.py requirements.txt requirements-dev.txt requirements-grpc-smg.txt co
ENV CCACHE_DIR=/root/.cache/ccache
# Build the TRT-LLM wheel
ARG GIT_COMMIT
ARG TRT_LLM_VER
ARG GITHUB_MIRROR=""
ARG BUILD_WHEEL_ARGS="--clean"
ARG BUILD_WHEEL_SCRIPT="scripts/build_wheel.py"
Expand All @@ -182,9 +183,13 @@ RUN --mount=type=cache,target=/root/.cache/pip --mount=type=cache,target=${CCACH
# Authenticates `git submodule update` and the CMake FetchContent clones
# (3rdparty/fetch_content.json) that this build pulls from github.com.
. docker/common/github_auth.sh && \
set -- ${BUILD_WHEEL_ARGS} && \
if [ "${BUILD_WHEEL_SCRIPT}" = "scripts/build_wheel.py" ] && [ -n "${TRT_LLM_VER}" ]; then \
set -- "$@" --version-override "${TRT_LLM_VER}"; \
fi && \
TRTLLM_BUILD_SOURCE_COMMIT=${GIT_COMMIT} \
GITHUB_MIRROR=${GITHUB_MIRROR} \
python3 ${BUILD_WHEEL_SCRIPT} ${BUILD_WHEEL_ARGS}
python3 ${BUILD_WHEEL_SCRIPT} "$@"

FROM ${DEVEL_IMAGE} AS release

Expand Down Expand Up @@ -236,6 +241,13 @@ ARG TARGETARCH
ENV TRT_LLM_GIT_COMMIT=${GIT_COMMIT} \
TRT_LLM_VERSION=${TRT_LLM_VER}

# Check installed metadata without importing CUDA-linked modules.
RUN --mount=type=bind,source=docker/common/validate_wheel_metadata.py,target=/tmp/validate_wheel_metadata.py \
--mount=type=bind,source=tensorrt_llm/version.py,target=/tmp/trtllm_source_version.py \
python3 /tmp/validate_wheel_metadata.py \
--expected-version "${TRT_LLM_VER}" --expected-commit "${GIT_COMMIT}" \
--source-version-file /tmp/trtllm_source_version.py

# Generate OSS attribution file for release image
RUN --mount=type=bind,source=scripts/generate_container_oss_attribution.sh,target=/mnt/gen_attribution.sh \
bash /mnt/gen_attribution.sh "release" "${TRT_LLM_VER}" "${TARGETARCH}"
Expand Down
2 changes: 1 addition & 1 deletion docker/common/install_base.sh
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ if [ -n "${GITHUB_MIRROR}" ]; then
fi

if [ -n "${GITHUB_MIRROR}" ]; then
BOOST_URL="https://urm.nvidia.com/artifactory/sw-dl-triton-generic-local/triton/ci-cd/binaries/boost/1.80.0/boost_1_80_0.tar.gz"
BOOST_URL="https://artifactory.nvidia.com/artifactory/sw-dl-triton-generic-local/triton/ci-cd/binaries/boost/1.80.0/boost_1_80_0.tar.gz"
else
BOOST_URL="https://archives.boost.io/release/1.80.0/source/boost_1_80_0.tar.gz"
fi
Expand Down
60 changes: 60 additions & 0 deletions docker/common/validate_wheel_metadata.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
"""Check installed wheel provenance without importing CUDA-linked modules."""

import re
from argparse import ArgumentParser
from importlib.metadata import distribution
from pathlib import Path

from packaging.version import Version


def validate_wheel_metadata(
expected_version: str, expected_commit: str, source_version_file: Path | None = None
) -> None:
"""Check installed metadata, resolving relative versions from the build source file."""
if not expected_version and not expected_commit:
return
if expected_version.startswith((".", "+")):
if source_version_file is None:
raise SystemExit("Relative versions require --source-version-file")
match = re.search(r'(?m)^__version__ = "([^"]+)"$', source_version_file.read_text())
if not match:
raise SystemExit(f"Unable to read __version__ from {source_version_file}")
base_version = match.group(1)
expected_version = (
base_version
if base_version.endswith(expected_version)
else base_version + expected_version
)
wheel = distribution("tensorrt_llm")
if expected_version and Version(wheel.version) != Version(expected_version):
raise SystemExit(
f"TRT-LLM version mismatch: installed={wheel.version}, expected={expected_version}"
)
if expected_commit:
if not re.fullmatch(r"[0-9a-fA-F]{40}", expected_commit):
raise SystemExit(f"Expected a full 40-character source commit SHA: {expected_commit}")
source_urls = []
for entry in wheel.metadata.get_all("Project-URL", []):
label, _, url = entry.partition(",")
if label.strip() == "Source Commit":
source_urls.append(url.strip().lower())
expected_url = f"https://github.com/NVIDIA/TensorRT-LLM/commit/{expected_commit}"
if source_urls != [expected_url.lower()]:
raise SystemExit(
f"TRT-LLM source commit mismatch: installed={source_urls or '<missing>'}, "
f"expected={expected_url}"
)


if __name__ == "__main__":
parser = ArgumentParser(description=__doc__)
parser.add_argument("--expected-version", default="")
parser.add_argument("--expected-commit", default="")
parser.add_argument(
"--source-version-file", type=Path, help="Build-context version.py for relative overrides"
)
args = parser.parse_args()
validate_wheel_metadata(args.expected_version, args.expected_commit, args.source_version_file)
Loading
Loading