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++
dsv41 Windows: queued GPU work submitted at once on NVIDIA too (win_submit), after every flag hand-off; DS4_WIN_SUBMIT=0 turns it off
Niko12215ec845f6bcb4
diff --git a/src/dsv41/ds4_kernels.cu b/src/dsv41/ds4_kernels.cuindex b04e9eec9..58df7c19b 100644--- a/src/dsv41/ds4_kernels.cu+++ b/src/dsv41/ds4_kernels.cu@@ -2123,6 +2123,23 @@ void xfer_put(const Xfer& x, unsigned* flag, unsigned seq, unsigned* cnt, cudaStfor (int i = 0; i < x.n; ++i) total += x.bytes[i] / 16;const int blocks = (int) std::max<size_t>(1, std::min<size_t>(64, (total + kThreads - 1) / kThreads));xfer_put_kernel<<<blocks, kThreads, 0, s>>>(x, flag, seq, cnt);+ // (another GPU, or the host, waits for this flag: on Windows the put must not sit in the queue - #303: two RTX 3080s+ // hung at the first hand-off, the receiver spinning on a flag whose writer had never been submitted)+ win_submit(s);+}+void win_submit(cudaStream_t s) {+#if defined(_WIN32)+ static const bool on = [] {+ const char* v = std::getenv("DS4_WIN_SUBMIT");+ return v == nullptr || std::atoi(v) != 0;+ }();+ if (on) {+ (void) cudaStreamQuery(s);+ (void) cudaGetLastError();+ }+#else+ (void) s;+#endif}void xfer_get(const Xfer& x, const unsigned* flag, unsigned seq, cudaStream_t s) {size_t total = 0;@@ -2133,7 +2150,10 @@ void xfer_get(const Xfer& x, const unsigned* flag, unsigned seq, cudaStream_t s)void wait_flags(const unsigned* flags, int n, int stride, unsigned seq, cudaStream_t s) {if (n > 0) wait_flags_kernel<<<1, 32, 0, s>>>(flags, n, stride, seq);}-void set_flag(unsigned* flag, unsigned v, cudaStream_t s) { set_flag_kernel<<<1, 1, 0, s>>>(flag, v); }+void set_flag(unsigned* flag, unsigned v, cudaStream_t s) {+ set_flag_kernel<<<1, 1, 0, s>>>(flag, v);+ win_submit(s); // (the writer on another GPU waits for this acknowledgement before its next put - #384)+}void allreduce(float* x, int n, const ArGroup& a, int mi, unsigned seq, unsigned* cnt, cudaStream_t s) {const int slot = (int) (seq % (unsigned) a.slots);if (a.nm > 2 && a.sums != nullptr && a.bflags != nullptr) { // (the reduce-scatter: ar_rs_kernel)diff --git a/src/dsv41/ds4_kernels.cuh b/src/dsv41/ds4_kernels.cuhindex 72488e9b3..9bc16233d 100644--- a/src/dsv41/ds4_kernels.cuh+++ b/src/dsv41/ds4_kernels.cuh@@ -279,6 +279,9 @@ struct Xfer {int n = 0;};void xfer_put(const Xfer& x, unsigned* flag, unsigned seq, unsigned* cnt, cudaStream_t s);+// Windows (WDDM, CUDA and HIP alike) holds launches back until the host waits or queries the stream: the work queued on+// `s` submitted now (cudaStreamQuery, which does not wait). Elsewhere nothing. DS4_WIN_SUBMIT=0: off (A/B tests)+void win_submit(cudaStream_t s);void xfer_get(const Xfer& x, const unsigned* flag, unsigned seq, cudaStream_t s);/// Waits (one thread) until flags[i * stride] >= seq for i < n (a pipelined chunk's readers have taken a buffer);/// set_flag: *flag = v after a system fence (a reader's acknowledgement).diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cuindex a9e370441..1eaeccebc 100644--- a/src/dsv41/ds4_model.cu+++ b/src/dsv41/ds4_model.cu@@ -3047,16 +3047,10 @@ struct Model::Impl {}// ---- the flow (see Chunk)- // Windows' HIP runtime submits queued work only when the host waits on the GPU: the flow's copies and the compute- // it queued between its waits are submitted at once (cudaStreamQuery does that without waiting); elsewhere nothing- static void kick(cudaStream_t s) {-#if defined(_WIN32) && defined(STRATA_USE_HIP)- (void) cudaStreamQuery(s);- (void) cudaGetLastError();-#else- (void) s;-#endif- }+ // Windows (WDDM: HIP's runtime and CUDA's driver alike) submits queued work only when the host waits on the GPU:+ // the flow's copies and the compute it queued between its waits are submitted at once (win_submit); elsewhere+ // nothing. (CUDA too since #303: two RTX 3080s hung before the first layer, one GPU spinning on the other's flow)+ static void kick(cudaStream_t s) { win_submit(s); }size_t flow_region(int r) const { // a part's room in a staging slot (read from its offset rounded down)return ((flow_part[r] + os::kDirectAlign - 1) & ~(size_t) (os::kDirectAlign - 1)) + os::kDirectAlign;}