Make vllm.cpp v0.0.2 builds deterministic

.github/workflows/platforms.yml

@@ -175,9 +175,16 @@ jobs:
           cmake -S vllm-cpp-sys/vllm.cpp -B target/vulkan-native \
             -G Ninja -DCMAKE_BUILD_TYPE=Release \
             -DVLLM_CPP_BUILD_TESTS=ON -DVLLM_CPP_BUILD_EXAMPLES=OFF \
-            -DVLLM_CPP_SERVER=OFF -DVLLM_CPP_CUDA=OFF \
-            -DVLLM_CPP_METAL=OFF -DVLLM_CPP_MLX=OFF \
-            -DVLLM_CPP_VULKAN=ON -DVLLM_CPP_TRITON=OFF
+            -DVLLM_CPP_SERVER=OFF -DVLLM_CPP_HIP=OFF \
+            -DVLLM_CPP_TENSTORRENT=OFF -DVLLM_CPP_LITERAL_STATIC=OFF \
+            -DVLLM_CPP_BENCH_PROFILE_CONTROL=OFF -DVLLM_CPP_NCCL=OFF \
+            -DVLLM_CPP_MARLIN=ON -DVLLM_CPP_FLASH_ATTN=ON \
+            -DVLLM_CPP_CUDA=OFF -DVLLM_CPP_METAL=OFF \
+            -DVLLM_CPP_MLX=OFF -DVLLM_CPP_VULKAN=ON \
+            -DVLLM_CPP_TRITON=OFF -DVLLM_CPP_TRITON_REGEN=OFF \
+            -DVLLM_CPP_TRITON_VENDORED_ARCH= \
+            -DVLLM_CPP_TRITON_TARGET= -DVLLM_CPP_CUTLASS_FETCH=OFF \
+            -DVLLM_CPP_SANITIZE=OFF
           cmake --build target/vulkan-native --target test_vulkan_backend test_backend_cross_device
           ctest --test-dir target/vulkan-native --output-on-failure \
             --tests-regex '^(test_vulkan_backend|test_backend_cross_device)$'

Justfile

@@ -97,10 +97,17 @@ sys: bindings-check header-check build-support-test backend-config backend-integ
 
 # Test all Linux CPU link modes and the exact shared-library exports.
 link-modes:
+    just --justfile '{{ root }}/Justfile' _link-modes \
+      '{{ root }}/vllm-cpp-sys' \
+      "${CARGO_TARGET_DIR:-{{ root }}/target/link-modes}"
+
+[private]
+_link-modes crate_root target_base:
     #!/usr/bin/env bash
     set -euo pipefail
-    repo_root={{ quote(root) }}
-    target_base=${CARGO_TARGET_DIR:-"$repo_root/target/link-modes"}
+    crate_root={{ quote(crate_root) }}
+    target_base={{ quote(target_base) }}
+    crate_root=$(cd "$crate_root" && pwd -P)
     mkdir -p "$target_base"
     target_base=$(cd "$target_base" && pwd -P)
     recipe_base="$target_base/vllm-cpp-sys-link-modes"
@@ -114,7 +121,7 @@ link-modes:
       esac
     }
     trap cleanup EXIT
-    cd "$repo_root"
+    cd "$crate_root"
     export CARGO_TERM_COLOR=never
 
     find_one() {
@@ -218,7 +225,7 @@ link-modes:
 
     prefix="$work/system-prefix"
     mkdir -p "$prefix/include" "$prefix/lib" "$prefix/lib64" "$prefix/blake3-lib"
-    cp vllm-cpp-sys/vllm.cpp/include/vllm.h "$prefix/include/"
+    cp vllm.cpp/include/vllm.h "$prefix/include/"
 
     bundled_static_lib=$(find_installed_library \
       "$bundled_static_target/release/build" libvllm.a)
@@ -395,8 +402,8 @@ package-test:
       echo 'could not read the native project version from vllm.cpp/CMakeLists.txt' >&2
       exit 1
     }
-    [[ $native_version == "$version" ]] || {
-      echo "native and crate versions differ: native=$native_version crates=$version" >&2
+    [[ $native_version == 0.0.2 ]] || {
+      echo "expected pinned native version 0.0.2, found $native_version" >&2
       exit 1
     }
     jq -e --arg version "$version" '
@@ -445,13 +452,14 @@ package-test:
     diff -u "$sys_list" <(archive_inventory "$sys_package" "vllm-cpp-sys-$version")
     diff -u "$safe_list" <(archive_inventory "$safe_package" "vllm-cpp-$safe_version")
 
+    native_exclude_pattern='^vllm\.cpp/cmake/(CudaArchFeaturesTest|CudaSourceGencodeTest|DumpTritonAOTContract|InSourceGuardTest|TritonAOTDefaultTest|TritonAOTMultiArchTest|VerifyExports)\.cmake$|^vllm\.cpp/include/vt/rocm/(rocm_gelu_mul_sep|rocm_gemma4_expert_geglu|rocm_matmul_batch|rocm_rmsnorm_plus_add)\.h$|^vllm\.cpp/src/vllm/entrypoints/openai/(api_server|server_main)\.cpp$|^vllm\.cpp/src/vllm/platforms/tenstorrent\.cpp$|^vllm\.cpp/src/vt/cuda/marlin/.*/generate_kernels\.py$|^vllm\.cpp/src/vt/rocm/[^/]+\.hip$|^vllm\.cpp/src/vt/tenstorrent/|^vllm\.cpp/src/vt/vulkan/shaders/'
     native_inventory() {
       local base=$1
       local member
       shift
       for member in "$@"; do
         find "$base/$member" -type f -print
-      done | sed "s#^$base/##" | LC_ALL=C sort
+      done | sed "s#^$base/##" | grep -Ev "$native_exclude_pattern" | LC_ALL=C sort
     }
     native_members=(
       vllm.cpp/CMakeLists.txt
@@ -461,9 +469,11 @@ package-test:
       vllm.cpp/include
       vllm.cpp/src
       vllm.cpp/scripts/triton-aot-compile.py
+      vllm.cpp/tests/vt/test_rocm_backend.cpp
       vllm.cpp/triton_kernels
       vllm.cpp/third_party/README.md
       vllm.cpp/third_party/blake3
+      vllm.cpp/third_party/doctest/doctest.h
       vllm.cpp/third_party/minja
       vllm.cpp/third_party/nlohmann
       vllm.cpp/third_party/vulkan
@@ -481,6 +491,7 @@ package-test:
         README.md \
         THIRD_PARTY.md \
         build.rs \
+        licenses/DOCTEST-MIT.txt \
         licenses/FLASH-ATTENTION-BSD-3-CLAUSE.txt \
         licenses/FLASH-LINEAR-ATTENTION-MIT.txt \
         src/bindings.rs \
@@ -500,7 +511,7 @@ package-test:
       <(native_inventory "$repo_root/vllm-cpp-sys" "${native_members[@]}") \
       <(native_inventory "$sys_root" "${native_members[@]}")
     diff -u \
-      <(printf '%s\n' CMakeLists.txt LICENSE NOTICE cmake include scripts src third_party triton_kernels | LC_ALL=C sort) \
+      <(printf '%s\n' CMakeLists.txt LICENSE NOTICE cmake include scripts src tests third_party triton_kernels | LC_ALL=C sort) \
       <(find "$sys_root/vllm.cpp" -mindepth 1 -maxdepth 1 -printf '%f\n' | LC_ALL=C sort)
 
     safe_expected="$temp/vllm-cpp.expected"
@@ -535,6 +546,7 @@ package-test:
     required_sys_members=(
       README.md
       THIRD_PARTY.md
+      licenses/DOCTEST-MIT.txt
       licenses/FLASH-ATTENTION-BSD-3-CLAUSE.txt
       licenses/FLASH-LINEAR-ATTENTION-MIT.txt
       vllm.cpp/include/vllm.h
@@ -546,13 +558,24 @@ package-test:
       vllm.cpp/src/vt/cuda/cuda_matmul_fp8_cutlass.cu
       vllm.cpp/src/vt/cuda/flash_attn/src/flash.h
       vllm.cpp/src/vt/cuda/marlin/core/scalar_type.hpp
+      vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_80/MANIFEST
+      vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_86/MANIFEST
+      vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_89/MANIFEST
+      vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_90a/MANIFEST
+      vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_100a/MANIFEST
       vllm.cpp/src/vt/cuda/triton_aot_vendored/sm_121a/MANIFEST
+      vllm.cpp/src/vt/metal/metal_mlx_provider.mm
+      vllm.cpp/src/vt/vulkan/vulkan_spirv.h
+      vllm.cpp/src/vllm/platforms/rocm.cpp
+      vllm.cpp/include/vt/rocm/rocm_arch.h
+      vllm.cpp/include/vt/rocm/rocm_runtime.h
+      vllm.cpp/tests/vt/test_rocm_backend.cpp
       vllm.cpp/scripts/triton-aot-compile.py
       vllm.cpp/triton_kernels/chunk_delta_h.py
-      vllm.cpp/src/vt/vulkan/vulkan_spirv.h
       vllm.cpp/third_party/README.md
       vllm.cpp/third_party/blake3/LICENSE_A2
       vllm.cpp/third_party/blake3/LICENSE_CC0
+      vllm.cpp/third_party/doctest/doctest.h
       vllm.cpp/third_party/minja/LICENSE
       vllm.cpp/third_party/nlohmann/json.hpp
       vllm.cpp/third_party/vulkan/vulkan_core.h
@@ -580,17 +603,37 @@ package-test:
       vllm.cpp/benchmarks
       vllm.cpp/docs
       vllm.cpp/examples
-      vllm.cpp/tests
       vllm.cpp/tools
-      vllm.cpp/third_party/doctest
       vllm.cpp/third_party/httplib
+      vllm.cpp/cmake/CudaArchFeaturesTest.cmake
+      vllm.cpp/cmake/CudaSourceGencodeTest.cmake
+      vllm.cpp/cmake/DumpTritonAOTContract.cmake
+      vllm.cpp/cmake/InSourceGuardTest.cmake
+      vllm.cpp/cmake/TritonAOTDefaultTest.cmake
+      vllm.cpp/cmake/TritonAOTMultiArchTest.cmake
+      vllm.cpp/cmake/VerifyExports.cmake
+      vllm.cpp/src/vllm/entrypoints/openai/api_server.cpp
+      vllm.cpp/src/vllm/entrypoints/openai/server_main.cpp
+      vllm.cpp/src/vllm/platforms/tenstorrent.cpp
+      vllm.cpp/src/vt/cuda/marlin/libtorch_stable/moe/marlin_moe_wna16/generate_kernels.py
+      vllm.cpp/src/vt/rocm
+      vllm.cpp/src/vt/tenstorrent
+      vllm.cpp/src/vt/vulkan/shaders
     )
     for member in "${denied_sys_members[@]}"; do
       [[ ! -e $sys_root/$member ]] || {
         echo "denied upstream tree or record leaked into sys package: $member" >&2
         exit 1
       }
     done
+    diff -u \
+      <(printf '%s\n' vllm.cpp/tests/vt/test_rocm_backend.cpp) \
+      <(find "$sys_root/vllm.cpp/tests" -type f -printf '%P\n' \
+        | sed 's#^#vllm.cpp/tests/#' | LC_ALL=C sort)
+    diff -u \
+      <(printf '%s\n' vllm.cpp/third_party/doctest/doctest.h) \
+      <(find "$sys_root/vllm.cpp/third_party/doctest" -type f -printf '%P\n' \
+        | sed 's#^#vllm.cpp/third_party/doctest/#' | LC_ALL=C sort)
 
     forbidden_pattern='(^|/)(target|stuff|\.git|\.github|__pycache__|\.cache|cache|fixtures?|downloads?|_deps|sdk)(/|$)|(^|/)(cutlass)(/|$)|(^|/)(model\.safetensors|tokenizer\.json|tokenizer_config\.json)$|\.(o|obj|a|so|dylib|dll|pyc|safetensors|gguf|pt|pth)$'
     for listing in "$sys_list" "$safe_list"; do
@@ -633,6 +676,7 @@ package-test:
         LICENSE-MIT \
         NOTICE \
         THIRD_PARTY.md \
+        licenses/DOCTEST-MIT.txt \
         licenses/FLASH-ATTENTION-BSD-3-CLAUSE.txt \
         licenses/FLASH-LINEAR-ATTENTION-MIT.txt \
         vllm.cpp/LICENSE \
@@ -671,13 +715,16 @@ package-test:
 
     sys_package_size=$(stat -c '%s' "$sys_package")
     sys_unpacked_size=$(du -sb "$sys_root" | cut -f1)
+    sys_regular_file_bytes=$(find "$sys_root" -type f -printf '%s\n' \
+      | awk '{ total += $1 } END { print total + 0 }')
+    sys_package_sha256=$(sha256sum "$sys_package" | awk '{ print $1 }')
     sys_entry_count=$(wc -l < "$sys_list")
     safe_package_size=$(stat -c '%s' "$safe_package")
     safe_unpacked_size=$(du -sb "$safe_root" | cut -f1)
     safe_entry_count=$(wc -l < "$safe_list")
     ((sys_package_size <= 6 * 1024 * 1024))
-    ((sys_unpacked_size <= 36 * 1024 * 1024))
-    ((sys_entry_count <= 1300))
+    ((sys_unpacked_size <= 40 * 1024 * 1024))
+    ((sys_entry_count <= 1400))
     ((safe_package_size <= 128 * 1024))
     ((safe_unpacked_size <= 256 * 1024))
     ((safe_entry_count <= 40))
@@ -693,11 +740,8 @@ package-test:
         and .[0].features.default == ["bundled"]
     ' <(cargo metadata --manifest-path "$sys_root/Cargo.toml" \
       --locked --offline --no-deps --format-version 1) >/dev/null
-    (
-      cd "$sys_root"
-      CARGO_NET_OFFLINE=true CARGO_TARGET_DIR="$temp_target" \
-        cargo test --locked --release --tests --offline
-    )
+    just --justfile "$repo_root/Justfile" _link-modes \
+      "$sys_root" "$temp/extracted-link-modes"
 
     sys_consumer="$temp/sys-consumer"
     mkdir -p "$sys_consumer/src"
@@ -819,8 +863,9 @@ package-test:
         cargo run --locked --release --offline
     )
 
-    printf 'sys package: %d entries, %d bytes unpacked, %d bytes compressed\n' \
-      "$sys_entry_count" "$sys_unpacked_size" "$sys_package_size"
+    printf 'sys package: %d entries, %d regular-file bytes, %d du bytes, %d compressed bytes, sha256 %s\n' \
+      "$sys_entry_count" "$sys_regular_file_bytes" "$sys_unpacked_size" \
+      "$sys_package_size" "$sys_package_sha256"
     printf 'safe package: %d entries, %d bytes unpacked, %d bytes compressed\n' \
       "$safe_entry_count" "$safe_unpacked_size" "$safe_package_size"
 

vllm-cpp-sys/Cargo.toml

@@ -26,11 +26,31 @@ include = [
     "/licenses/**",
     "/vllm.cpp/CMakeLists.txt",
     "/vllm.cpp/cmake/**",
+    "!/vllm.cpp/cmake/CudaArchFeaturesTest.cmake",
+    "!/vllm.cpp/cmake/CudaSourceGencodeTest.cmake",
+    "!/vllm.cpp/cmake/DumpTritonAOTContract.cmake",
+    "!/vllm.cpp/cmake/InSourceGuardTest.cmake",
+    "!/vllm.cpp/cmake/TritonAOTDefaultTest.cmake",
+    "!/vllm.cpp/cmake/TritonAOTMultiArchTest.cmake",
+    "!/vllm.cpp/cmake/VerifyExports.cmake",
     "/vllm.cpp/include/**",
+    "!/vllm.cpp/include/vt/rocm/rocm_gelu_mul_sep.h",
+    "!/vllm.cpp/include/vt/rocm/rocm_gemma4_expert_geglu.h",
+    "!/vllm.cpp/include/vt/rocm/rocm_matmul_batch.h",
+    "!/vllm.cpp/include/vt/rocm/rocm_rmsnorm_plus_add.h",
     "/vllm.cpp/src/**",
+    "!/vllm.cpp/src/vllm/entrypoints/openai/api_server.cpp",
+    "!/vllm.cpp/src/vllm/entrypoints/openai/server_main.cpp",
+    "!/vllm.cpp/src/vllm/platforms/tenstorrent.cpp",
+    "!/vllm.cpp/src/vt/cuda/marlin/**/generate_kernels.py",
+    "!/vllm.cpp/src/vt/rocm/*.hip",
+    "!/vllm.cpp/src/vt/tenstorrent/**",
+    "!/vllm.cpp/src/vt/vulkan/shaders/**",
+    "/vllm.cpp/tests/vt/test_rocm_backend.cpp",
     "/vllm.cpp/triton_kernels/**",
     "/vllm.cpp/scripts/triton-aot-compile.py",
     "/vllm.cpp/third_party/blake3/**",
+    "/vllm.cpp/third_party/doctest/doctest.h",
     "/vllm.cpp/third_party/minja/**",
     "/vllm.cpp/third_party/nlohmann/**",
     "/vllm.cpp/third_party/vulkan/**",

vllm-cpp-sys/README.md

@@ -4,7 +4,7 @@ Raw Rust bindings and native linking for the stable C API of [vllm.cpp](https://
 
 This crate exposes checked-in generated unsafe declarations for the 35 exported C symbols in ABI version 17. Callers are responsible for pointer validity, lifetimes, callback threading, status/error handling, and matching every native allocation with its documented free function. Applications should prefer the current safe [`vllm-cpp`](https://docs.rs/vllm-cpp) crate unless they require direct ABI access. Ordinary consumers do not need Just, bindgen, or libclang.
 
-The package contains Rust declarations and conformance tests, native build/link integration, the pinned native source inputs required by the supported feature set, and their licenses/notices. It excludes upstream tests, fixtures, models, examples, benchmarks, agent records, fetched SDKs, external CUTLASS trees, and build output. See the repository [changelog](https://github.com/querymt/vllm-cpp-rs/blob/main/CHANGELOG.md) and [release process](https://github.com/querymt/vllm-cpp-rs/blob/main/RELEASING.md) for the coordinated crate boundary.
+The package contains Rust declarations and conformance tests, native build/link integration, the pinned native source inputs required by the supported feature set, and their licenses/notices. It excludes upstream test trees except for one unconditional ROCm syntax-check translation unit, plus fixtures, models, examples, benchmarks, agent records, fetched SDKs, external CUTLASS trees, and build output. See the repository [changelog](https://github.com/querymt/vllm-cpp-rs/blob/main/CHANGELOG.md) and [release process](https://github.com/querymt/vllm-cpp-rs/blob/main/RELEASING.md) for the coordinated crate boundary.
 
 ## Link Modes
 
@@ -28,12 +28,14 @@ Feature selection is build configuration, not runtime or hardware evidence. Acce
 |---|---|---|
 | `cuda` | Linux x86_64/aarch64; requires exact `VLLM_CPP_CUDA_ARCHITECTURES` | Static mode links CMake's exact `CUDA_CUDART` and `CUDA_cublasLt_LIBRARY` results; dynamic mode relies on `libvllm.so` dependencies |
 | `cuda-cutlass` | Implies CUDA; requires canonical `VLLM_CPP_CUTLASS_DIR` with CUTLASS >=4.5.0; fetch is OFF; `103a`/`110` rejected | Header-only external input |
-| `triton-aot` | Implies CUDA; checked-in single-architecture artifacts only; regeneration is OFF | Static mode also links CMake's exact `CUDA_cuda_driver_LIBRARY`; dynamic mode relies on `libvllm.so` |
+| `triton-aot` | Implies CUDA; embeds all six checked-in AOT trees; regeneration is OFF | Static mode also links CMake's exact `CUDA_cuda_driver_LIBRARY`; dynamic mode relies on `libvllm.so` |
 | `vulkan` | Linux x86_64/aarch64; packaged Khronos headers and checked-in SPIR-V | No Vulkan SDK link; runtime loader uses `dlopen` |
 | `metal` | Exact `aarch64-apple-darwin`; native MSL backend | Static mode links `c++`, then `Metal` and `Foundation` frameworks |
 | `mlx` | Implies Metal; requires canonical external `MLX_ROOT` with `include/mlx/array.h`, `lib/libmlx.dylib`, and `lib/mlx.metallib` | Static mode adds `$MLX_ROOT/lib` before `dylib=mlx`; no rpath or packaged MLX payload |
 
-CUDA architectures are exactly `80`, `86`, `87`, `89`, `90a`, `100a`, `103a`, `110`, `120a`, `121a`, or `120a;121a`. Triton accepts only `80`, `86`, `89`, `90a`, `100a`, or `121a`. Plain CUDA passes a nonexistent CUTLASS root to CMake so ambient source cannot silently change the build. Use separate `CARGO_TARGET_DIR` values for each backend/link combination.
+CUDA architectures are exactly `80`, `86`, `87`, `89`, `90a`, `100a`, `103a`, `110`, `120a`, `121a`, or `120a;121a`. `triton-aot` accepts every one of those values, including the fat `120a;121a` build, and embeds the complete `sm_80`, `sm_86`, `sm_89`, `sm_90a`, `sm_100a`, and `sm_121a` vendored set. Runtime dispatch selects only an exact-SM cubin; architectures without an exact tree use the portable CUDA path rather than a neighboring cubin. Plain CUDA passes a nonexistent CUTLASS root to CMake so ambient source cannot silently change the build. Use separate `CARGO_TARGET_DIR` values for each backend/link combination.
+
+Cargo builds explicitly disable HIP/ROCm, Tenstorrent, NCCL, native tests/examples, and the HTTP server. CUTLASS fetching and Triton regeneration remain OFF, so supported bundled builds never download native inputs. Marlin and FlashAttention remain explicitly enabled when their CUDA architecture and external CUTLASS prerequisites are satisfied.
 
 Metal and MLX remain bundled-only and do not imply `bundled`; `--no-default-features` callers must explicitly select `bundled,metal` or `bundled,mlx`. Setting `MLX_ROOT` without `mlx` is rejected. MLX is never downloaded, copied, or packaged, and applications are responsible for making `libmlx.dylib` and its metallib deployment visible at runtime.
 

vllm-cpp-sys/THIRD_PARTY.md

@@ -12,7 +12,8 @@
 | Marlin / GPTQ-Marlin slice | `vllm.cpp/src/vt/cuda/marlin/**` | vLLM commit `e24d1b24`, with retained file notices and vllm.cpp adapter changes | Apache-2.0 in `vllm.cpp/LICENSE` and retained source notices |
 | Flash Linear Attention Triton kernels | `vllm.cpp/triton_kernels/**`, generated AOT files under `vllm.cpp/src/vt/cuda/triton_aot_vendored/**` | FLA source ported through vLLM 0.24.0; exact source and artifact hashes are pinned in each AOT `MANIFEST` | `licenses/FLASH-LINEAR-ATTENTION-MIT.txt` |
 | nlohmann/json | `vllm.cpp/third_party/nlohmann/**` | nlohmann/json 3.12.0 | MIT notice retained in `json.hpp` |
+| doctest | `vllm.cpp/third_party/doctest/doctest.h` | doctest 2.5.2, required by the packaged ROCm syntax-check source | `licenses/DOCTEST-MIT.txt` |
 
-The package excludes upstream tests, fixtures, benchmarks, models, external SDKs, and build output. It contains only native source/build inputs and required licenses/notices.
+The package excludes upstream tests, fixtures, benchmarks, models, external SDKs, and build output. It contains only native source/build inputs and required licenses/notices, including the one ROCm test translation unit compiled unconditionally as a syntax check.
 
 vllm.cpp is an independent community project and is not affiliated with or endorsed by the vLLM project, the PyTorch Foundation, or the Linux Foundation.

vllm-cpp-sys/build.rs

@@ -45,9 +45,11 @@ fn main() {
         "vllm.cpp/cmake",
         "vllm.cpp/include/vllm.h",
         "vllm.cpp/src",
+        "vllm.cpp/tests/vt/test_rocm_backend.cpp",
         "vllm.cpp/triton_kernels",
         "vllm.cpp/scripts/triton-aot-compile.py",
         "vllm.cpp/third_party/blake3",
+        "vllm.cpp/third_party/doctest/doctest.h",
         "vllm.cpp/third_party/minja",
         "vllm.cpp/third_party/nlohmann",
         "vllm.cpp/third_party/vulkan",