Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

cmake: a CUDA build with no architecture option is for sm_120, and `native` builds the W4A4 unit for sm_120a (#4029, #4023)

pschaub3cd33d504c57
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2d83fad86..1cf372d97 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -174,13 +174,24 @@ elseif(STRATA_ENABLE_CUDA)
if(MSVC AND STRATA_EXPERIMENTAL_SM60 AND NOT DEFINED CMAKE_CUDA_RUNTIME_LIBRARY)
set(CMAKE_CUDA_RUNTIME_LIBRARY Shared)
endif()
+ # #4023: the default below has to be set BEFORE enable_language(CUDA): since CMake 3.18 that call fills an unset
+ # CMAKE_CUDA_ARCHITECTURES with nvcc's own default (sm_75 for CUDA 13), so a test after it never saw "undefined" and
+ # a build with no option was an sm_75 build. -DCMAKE_CUDA_ARCHITECTURES and the CUDAARCHS environment variable win.
+ set(_strata_cuda_arch_defaulted OFF)
+ if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES AND NOT DEFINED ENV{CUDAARCHS})
+ set(CMAKE_CUDA_ARCHITECTURES 120 CACHE STRING "CUDA architectures")
+ set(_strata_cuda_arch_defaulted ON)
+ endif()
enable_language(CUDA)
# `enable_language(CUDA)` gives the .cu compile rules but NOT the imported targets, so `CUDA::cudart` has to
# be brought in explicitly for the host-side tests that call cudaMalloc themselves.
find_package(CUDAToolkit REQUIRED)
option(STRATA_EXPERIMENTAL_SM60 "community build for GPUs below compute capability 7.5: Pascal sm_60, Volta sm_70 (#236)" OFF)
- if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
- set(CMAKE_CUDA_ARCHITECTURES 120 CACHE STRING "CUDA architectures" FORCE)
+ if(_strata_cuda_arch_defaulted)
+ # a recommendation, not a refusal: the release engine is built for 75;86;89;120, a hand build names its card
+ message(WARNING "Strata: CMAKE_CUDA_ARCHITECTURES was not set, so this build is for sm_120 (RTX 50) only. For another "
+ "card pass it: -DCMAKE_CUDA_ARCHITECTURES=native (the GPUs in this PC), 75 (RTX 20), 86 (RTX 30), "
+ "89 (RTX 40), or a list like \"86;120\".")
endif()
foreach(_a IN LISTS CMAKE_CUDA_ARCHITECTURES)
string(REGEX REPLACE "-.*$" "" _base "${_a}") # 120-real / 120-virtual are still 120
@@ -1916,6 +1927,12 @@ if(STRATA_NATIVE_EXPERTS)
# which exists only in the arch-specific target, so its one unit alone is built with 12x -> 12xa (the default
# binary is unchanged and runs on sm_121; a card without that image falls back to W4A8 at run time)
set(_strata_w4a4_archs ${CMAKE_CUDA_ARCHITECTURES})
+ # `native` is first replaced by the detected architectures (e.g. 120-real), so the 12x -> 12xa rewrite below applies
+ # to it too: with plain sm_120 ptxas refuses the FP4 MMA of this one unit (#4023)
+ if("native" IN_LIST _strata_w4a4_archs AND CMAKE_CUDA_ARCHITECTURES_NATIVE)
+ list(REMOVE_ITEM _strata_w4a4_archs native)
+ list(APPEND _strata_w4a4_archs ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
+ endif()
# (two steps: CMake before 3.29 rejects a \2 whose optional group did not match, "120" alone)
list(TRANSFORM _strata_w4a4_archs REPLACE "^(12[0-9])$" "\\1a")
list(TRANSFORM _strata_w4a4_archs REPLACE "^(12[0-9])-(real|virtual)$" "\\1a-\\2")
diff --git a/docs/SOURCE_INSTALL.md b/docs/SOURCE_INSTALL.md
index 1c0d7ec07..a4996c571 100644
--- a/docs/SOURCE_INSTALL.md
+++ b/docs/SOURCE_INSTALL.md
@@ -71,6 +71,9 @@ cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_CUDA=ON -DCMAKE_C
cmake --build build --target strata -j8
```
+`native` builds for the GPUs in this PC. Name the card instead for another one: `75` (RTX 20), `86` (RTX 30), `89` (RTX 40),
+`120` (RTX 50), or a list such as `"86;120"`. Without the option the build is for RTX 50 only and the configure step says so.
+
Run against the resulting binary and the matching model config. Use the build instructions for
[HIP](AMD_HIP.md), [SYCL](INTEL_ARC.md) or [Metal](MAC.md) when changing those backends.
Follow [contribution validation requirements](../CONTRIBUTING.md) before requesting review.