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++

SYCL engine: strata --version answers before any GPU is touched; no SYCL device is built before main

Niko1221424b364a88a6
diff --git a/sycl/include/dpct/blas_utils.hpp b/sycl/include/dpct/blas_utils.hpp
index 4a58aad33..60afdfb26 100644
--- a/sycl/include/dpct/blas_utils.hpp
+++ b/sycl/include/dpct/blas_utils.hpp
@@ -303,14 +303,14 @@ public:
_saved_queue_ptr = q_ptr;
}
static inline sycl::queue &get_saved_queue() noexcept {
+ if (!_saved_queue_ptr) _saved_queue_ptr = &::dpct::cs::get_default_queue();
return *_saved_queue_ptr;
}
private:
::dpct::cs::queue_ptr _queue_ptr = &::dpct::cs::get_default_queue();
math_mode _mm = math_mode::mm_default;
- static inline ::dpct::cs::queue_ptr _saved_queue_ptr =
- &::dpct::cs::get_default_queue();
+ static inline ::dpct::cs::queue_ptr _saved_queue_ptr = nullptr;
};
using descriptor_ptr = descriptor *;
diff --git a/sycl/include/dpct/detail/memory_detail.hpp b/sycl/include/dpct/detail/memory_detail.hpp
index e0bad4389..1c4c5d9ee 100644
--- a/sycl/include/dpct/detail/memory_detail.hpp
+++ b/sycl/include/dpct/detail/memory_detail.hpp
@@ -822,14 +822,11 @@ public:
/// Constructor with range
device_memory(const sycl::range<Dimension> &range_in)
: _size(range_in.size() * sizeof(T)), _range(range_in), _reference(false),
- _host_ptr(nullptr), _device_ptrs(get_ptrs_size(), nullptr) {
+ _host_ptr(nullptr) {
static_assert((Memory == global) || (Memory == constant) ||
(Memory == shared),
"device memory region should be global, constant or shared");
- // Make sure that singleton class mem_mgr and dev_mgr will destruct later
- // than this.
- detail::mem_mgr::instance();
- dev_mgr::instance();
+ // Strata: the pointer table, mem_mgr and dev_mgr are made at the first get_ptr_impl() (sycl/tools/fixups.py)
}
/// Constructor with range
@@ -927,6 +924,11 @@ private:
}
value_t *get_ptr_impl(sycl::queue &q) {
+ if (_device_ptrs.empty()) {
+ detail::mem_mgr::instance();
+ dev_mgr::instance();
+ _device_ptrs.assign(get_ptrs_size(), nullptr);
+ }
#ifdef DPCT_USM_LEVEL_NONE
auto &ptr = _device_ptrs.front();
#else
@@ -953,7 +955,7 @@ private:
bool _reference;
value_t *_host_ptr;
std::vector<value_t *> _device_ptrs;
- sycl::queue _q;
+ std::optional<sycl::queue> _q; // Strata: a default-built sycl::queue picks a device (sycl/tools/fixups.py)
};
template <class T, memory_region Memory>
class device_memory<T, Memory, 0> : public device_memory<T, Memory, 1> {
diff --git a/sycl/src/program/generate.cpp b/sycl/src/program/generate.cpp
index 5a4d9f0aa..2a4db4d16 100644
--- a/sycl/src/program/generate.cpp
+++ b/sycl/src/program/generate.cpp
@@ -1723,6 +1723,15 @@ int main(int argc, char **argv) try {
// pipe or a file is block-buffered, so a program that dies loses every line it had already printed - which
// turns "it crashed at step 7" into "it crashed somewhere", and the difference is a debugging session.
std::setvbuf(stdout, nullptr, _IONBF, 0);
+ // `strata --version` / `-V` (also `strata generate --version`): the version and nothing else, before any GPU is touched, as the CUDA
+ // and HIP engines have answered it. `strata update` runs it as the engine's smoke test, and the Intel package build runs it with
+ // an empty environment.
+ for (int i = 1; i < argc && i < 3; ++i) {
+ if (std::strcmp(argv[i], "--version") == 0 || std::strcmp(argv[i], "-V") == 0) {
+ std::printf("strata %s\n", STRATA_VERSION);
+ return 0;
+ }
+ }
#if (defined(STRATA_USE_HIP) || defined(STRATA_HIP_GFX906)) && !defined(_WIN32)
// AMD, a file-backed arena (STRATA_ARENA_MMAP): ROCclr copies a pageable source of 1 MiB or more by locking its
// pages in place (a GPU userptr), and keeps them - so every expert the VRAM fill copied from the mapped
diff --git a/sycl/tools/fixups.py b/sycl/tools/fixups.py
index 053d2810d..b26e2bc9e 100644
--- a/sycl/tools/fixups.py
+++ b/sycl/tools/fixups.py
@@ -340,3 +340,51 @@ edit("src/kernels/cuda/native_router.dp.cpp", lambda s: s.replace(
edit("include/strata/core/mtp.hpp", lambda s: s.replace(
" dpct::queue_ptr cs_ = &dpct::get_in_order_queue();\n",
" dpct::queue_ptr cs_ = nullptr; // load() creates it; not the device's queue, which ~MtpDrafter would destroy\n"))
+
+# `strata --version` must print before any GPU is touched (the engine smoke test of `strata update` runs it, and a user outside the
+# `render` group has no usable device). Two things in dpct's helper headers looked up the SYCL devices before main:
+# 1. blas::descriptor's `static inline queue_ptr _saved_queue_ptr = &get_default_queue();` (an inline variable): the default is
+# looked up in get_saved_queue() instead.
+def lazy_saved_queue(s):
+ s = s.replace(" static inline ::dpct::cs::queue_ptr _saved_queue_ptr =\n &::dpct::cs::get_default_queue();",
+ " static inline ::dpct::cs::queue_ptr _saved_queue_ptr = nullptr;")
+ return s.replace(" static inline sycl::queue &get_saved_queue() noexcept {\n return *_saved_queue_ptr;",
+ " static inline sycl::queue &get_saved_queue() noexcept {\n"
+ " if (!_saved_queue_ptr) _saved_queue_ptr = &::dpct::cs::get_default_queue();\n"
+ " return *_saved_queue_ptr;")
+edit("include/dpct/blas_utils.hpp", lazy_saved_queue)
+
+# 2. dpct::device_memory (global_memory / constant_memory) sized its per-device pointer table, and so built mem_mgr and dev_mgr, in its
+# constructor. The engine's namespace-scope tables (iq4nl_values, c_codes) are heap objects built before main. Both now happen at the
+# first get_ptr_impl(); the destruction order the original comment protected does not arise for objects that are never freed.
+def lazy_device_memory(s):
+ s = s.replace(" _host_ptr(nullptr), _device_ptrs(get_ptrs_size(), nullptr) {\n", " _host_ptr(nullptr) {\n")
+ s = s.replace(" // Make sure that singleton class mem_mgr and dev_mgr will destruct later\n // than this.\n"
+ " detail::mem_mgr::instance();\n dev_mgr::instance();\n }\n",
+ " // Strata: the pointer table, mem_mgr and dev_mgr are made at the first get_ptr_impl() (sycl/tools/fixups.py)\n }\n")
+ s = s.replace(" std::vector<value_t *> _device_ptrs;\n sycl::queue _q;\n};", # a default-built sycl::queue picks a device
+ " std::vector<value_t *> _device_ptrs;\n std::optional<sycl::queue> _q; // Strata: a default-built sycl::queue picks a device (sycl/tools/fixups.py)\n};")
+ return s.replace(" value_t *get_ptr_impl(sycl::queue &q) {\n#ifdef DPCT_USM_LEVEL_NONE\n",
+ " value_t *get_ptr_impl(sycl::queue &q) {\n"
+ " if (_device_ptrs.empty()) {\n detail::mem_mgr::instance();\n dev_mgr::instance();\n"
+ " _device_ptrs.assign(get_ptrs_size(), nullptr);\n }\n#ifdef DPCT_USM_LEVEL_NONE\n")
+edit("include/dpct/detail/memory_detail.hpp", lazy_device_memory)
+
+# 3. the migrated generate.cpp has no `--version` (src/program/generate.cpp answers it before any GPU is touched): the engine printed its
+# usage and exited 2. Same lines as the CUDA tree, at the top of main().
+def version_flag(s):
+ if '"--version"' in s:
+ return s
+ return s.replace(" std::setvbuf(stdout, nullptr, _IONBF, 0);\n#if (defined(STRATA_USE_HIP)",
+ " std::setvbuf(stdout, nullptr, _IONBF, 0);\n"
+ " // `strata --version` / `-V` (also `strata generate --version`): the version and nothing else, before any GPU is touched, as the CUDA\n"
+ " // and HIP engines have answered it. `strata update` runs it as the engine's smoke test, and the Intel package build runs it with\n"
+ " // an empty environment.\n"
+ " for (int i = 1; i < argc && i < 3; ++i) {\n"
+ " if (std::strcmp(argv[i], \"--version\") == 0 || std::strcmp(argv[i], \"-V\") == 0) {\n"
+ " std::printf(\"strata %s\\n\", STRATA_VERSION);\n"
+ " return 0;\n"
+ " }\n"
+ " }\n"
+ "#if (defined(STRATA_USE_HIP)", 1)
+edit("src/program/generate.cpp", version_flag)