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++
Merge f-027-gh-eng-d (#3602 #3948 #3638 #3702 #3797 #2409 #3844: audits, fixes and tests)
Niko12219c624c75a976
diff --cc src/core/conversation_state.cppindex 4bfd47d6f,7fa40ca66..b8ce97016--- a/src/core/conversation_state.cpp+++ b/src/core/conversation_state.cpp@@@ -691,11 -397,10 +691,13 @@@ bool conversation_snapshot_validate(conreturn !draft || conversation_kv_validate(image.kv.back(), *draft, g, upto, false, error);}+ bool conversation_copy_stream_ready(int device) { return conversation_detail::copy_stream_ready(device); }+ConversationRestore conversation_snapshot_restore(const SavedConversation& image, SessionState& ss,const ModelGeometry& g, const QsaState* draft, std::string& error) {+ conversation_checkpoint_join(image.live.gdn.data());+ for (const auto& c : image.checkpoints) conversation_checkpoint_join(c.gdn.data());+ conversation_park_join_for(image); // (STRATA_PARK_VRAM_MIB: no wait when the K/V comes from its slot)if (!conversation_snapshot_validate(image, ss, g, draft, error)) return ConversationRestore::invalid;if (!sync(error)) return ConversationRestore::transfer_failed;const int64_t upto = (int64_t) image.live.ids.size();diff --cc src/core/verify.cppindex 920b55c17,193b8f5ba..1b5e18874--- a/src/core/verify.cpp+++ b/src/core/verify.cpp@@@ -4031,14 -3729,17 +4040,17 @@@ bool Verifier::instantiate_evicting(cud// #997: every layout of active slots keeps its own graph pair (20-30 MiB each on an L40S), and with an expert cache// sized to the reserve the next window graph - a batch layout or a one-request window captured on first use - can// find no VRAM left: free the least recently used layouts until it fits (they are captured again when needed)- cudaError_t ie = cudaGraphInstantiate(&ex, graph, flags);+ // (into a local: what a failed instantiate leaves in the output is not specified, and `ex` is a member that+ // capture() reads as "captured" when it is not null)+ cudaGraphExec_t got = nullptr;- cudaError_t ie = cudaGraphInstantiate(&got, graph, 0);++ cudaError_t ie = cudaGraphInstantiate(&got, graph, flags);int freed = 0;for (bool evicted = true; ie == cudaErrorMemoryAllocation && evicted;) {(void) cudaGetLastError();if (!evict_batch_graph(key, evicted, err)) return false;if (evicted) {++freed;- ie = cudaGraphInstantiate(&ex, graph, flags);- ie = cudaGraphInstantiate(&got, graph, 0);++ ie = cudaGraphInstantiate(&got, graph, flags);}}if (freed > 0)@@@ -4051,7 -3752,7 +4063,7 @@@if (!evict_window_graph(keep_t, evicted, err)) return false;if (evicted) {++freed_sizes;- ie = cudaGraphInstantiate(&ex, graph, flags);- ie = cudaGraphInstantiate(&got, graph, 0);++ ie = cudaGraphInstantiate(&got, graph, flags);}}if (freed_sizes > 0)diff --cc src/kernels/cuda/fused_gr.cuindex fe8672989,3811e7824..12c74a5dc--- a/src/kernels/cuda/fused_gr.cu+++ b/src/kernels/cuda/fused_gr.cu@@@ -2,7 -2,7 +2,8 @@@#include "strata/core/emulate.hpp"#include "strata/core/sm75_series.hpp"#include "strata/core/sm120_series.hpp"+#include "strata/core/defaults_026.hpp"+ #define STRATA_Q81_FASTMATH_DIV 1 // #3948: the q8_1 tails divide as native_mmvq.cu (--use_fast_math) does, see q8_1_finite.hpp#include "strata/kernels/q8_1_finite.hpp" // #606: q8_1_ds#include "strata/core/geometry.hpp"#include "strata/kernels/fused_gr.hpp"