[CI/Build][Docker] Bump nvidia-cutlass-dsl to 4.6.0 and drop packaging workarounds (#47442)

Signed-off-by: Artem Perevedentsev <aperevedents@nvidia.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Artem Perevedentsev
2026-07-16 13:51:41 +00:00
committed by GitHub
co-authored by Matthew Bonanni
parent b8168e33e0
commit 7cd1d57b74
7 changed files with 8 additions and 120 deletions
@@ -2,7 +2,6 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Inference-only Qwen3-Next/Qwen3.5 model."""
import functools
from typing import Literal
import torch
@@ -83,70 +82,6 @@ if GDN_AITER_TRITON_AVAILABLE:
logger = init_logger(__name__)
# TODO(arpera): remove ``_is_libs_cu13_install_intact`` and its caller in
# ``_resolve_gdn_prefill_backend`` once the upstream packaging bug is
# fixed and the broken wheels are yanked / superseded on PyPI:
# https://github.com/NVIDIA/cutlass/issues/3170
# https://github.com/NVIDIA/cutlass/issues/3259
@functools.cache
def _is_libs_cu13_install_intact() -> bool:
"""Return True if every file installed by ``nvidia-cutlass-dsl-libs-cu13``
matches the SHA-256 declared in its wheel ``RECORD``.
``nvidia-cutlass-dsl-libs-base`` and ``nvidia-cutlass-dsl-libs-cu13``
both ship into the shared ``nvidia_cutlass_dsl/`` namespace and
write many of the same on-disk paths (the runtime ``.so``, the MLIR
Python bindings, cuTe-DSL Python sources, ...) with different
content. Whichever wheel extracts last wins; with a parallel
installer (e.g. ``uv``) the order is racy and the resulting venv
can end up with a mix of files from both variants. The
``-libs-base`` variant fails MLIR legalization when JIT-compiling
the FlashInfer Blackwell GDN prefill kernel, and any other
cuTe-DSL-based kernel can break too if on-disk files diverge from
what ``-libs-cu13``'s wheel expects. Tracked upstream at:
* https://github.com/NVIDIA/cutlass/issues/3170
* https://github.com/NVIDIA/cutlass/issues/3259
This helper re-hashes every file the ``-libs-cu13`` wheel claims to
own and compares against its declared SHA-256. Returns False on any
error (uninstalled, missing RECORD, missing file, hash mismatch).
Result is cached per-process.
"""
import hashlib
import importlib.metadata
import pybase64 as base64
try:
dist = importlib.metadata.distribution("nvidia-cutlass-dsl-libs-cu13")
except importlib.metadata.PackageNotFoundError:
return False
files = dist.files
if not files:
return False
for pkg_path in files:
file_hash = pkg_path.hash
# Skip RECORD rows without a hash (RECORD itself, generated
# ``.pyc`` files, ...) and any non-SHA-256 hash modes.
if file_hash is None or not file_hash.value:
continue
if file_hash.mode != "sha256":
continue
try:
with open(pkg_path.locate(), "rb") as f:
digest = hashlib.sha256(f.read()).digest()
except OSError:
return False
actual = base64.urlsafe_b64encode(digest).decode().rstrip("=")
if actual != file_hash.value:
return False
return True
def _resolve_gdn_prefill_backend(
vllm_config: VllmConfig,
) -> tuple[str, Literal["triton", "flashinfer", "cutedsl"]]:
@@ -157,9 +92,7 @@ def _resolve_gdn_prefill_backend(
* ``platform == cuda``;
* one of the following:
- Hopper (SM90) — no further constraints;
- Blackwell (SM10.x) with ``head_k_dim == 128``, ``cuda_runtime >= 13``,
and an intact ``nvidia-cutlass-dsl-libs-cu13`` install on disk
(see :func:`_is_libs_cu13_install_intact`).
- Blackwell (SM10.x) with ``head_k_dim == 128``, ``cuda_runtime >= 13``.
In-tree CuteDSL GDN prefill kernel is chosen when:
* "cutedsl" is requested; (opt-in only)
@@ -190,19 +123,8 @@ def _resolve_gdn_prefill_backend(
and head_k_dim == 128
and current_platform.get_cuda_runtime_major() >= 13
):
supports_flashinfer = _is_libs_cu13_install_intact()
supports_flashinfer = True
supports_cutedsl = True
if not supports_flashinfer:
logger.warning_once(
"FlashInfer Blackwell GDN requires an intact nvidia-cutlass-dsl"
"-libs-cu13 install, but some on-disk files do not match the "
"SHA-256 declared in its RECORD (install-order race in "
"nvidia-cutlass-dsl packaging -- see "
"https://github.com/NVIDIA/cutlass/issues/3170 and "
"https://github.com/NVIDIA/cutlass/issues/3259). Falling back "
"to Triton/FLA. Repair with: pip install --force-reinstall "
"--no-deps nvidia-cutlass-dsl-libs-cu13"
)
if backend in ["flashinfer", "auto"] and supports_flashinfer:
return backend, "flashinfer"