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++
DeepSeek engine (AMD, diagnostic): the io threads' busy time in the prompt stream line
Niko12215b969c1f7a41
diff --git a/src/dsv41/ds4_model.cu b/src/dsv41/ds4_model.cuindex ea6531386..e4a0ba7e4 100644--- a/src/dsv41/ds4_model.cu+++ b/src/dsv41/ds4_model.cu@@ -598,6 +598,7 @@ struct Chunk {double dbg_stage_ms = 0, dbg_dma_ms = 0, dbg_evt_ms = 0;uint64_t it_file = 0, it_ram = 0, it_spec = 0, it_spec_unused = 0;uint64_t io_d0 = 0, io_b0 = 0; // io's unbuffered / buffered bytes when the run started+ double io_busy0 = 0; // io's busy time when the run started (DS4_PROF)// layer-major (a run of several chunks, every layer on this device): the layers one after another, each through// all the chunks; a chunk's state between its layers (h | pre | topk | cand, chunk k at k chunk_hand_bytes()) in// host memory, and the layer cache - RAM slots where the experts a layer's chunk reads from the files are kept for@@ -7739,6 +7740,7 @@ struct Model::Impl {c.lc_first = c.lc_chk = 0;c.io_d0 = c.io->direct_bytes();c.io_b0 = c.io->buffered_bytes();+ c.io_busy0 = c.io->busy_ms();if (!c.fl_layers.empty()) {flow_spec(c, 0, c.fl_layers[0], run.T[0]);if (!flow_pump(c, err)) return false;@@ -8502,10 +8504,11 @@ struct Model::Impl {if (c.lm)std::snprintf(lcs, sizeof lcs, ", %llu again from the layer cache (layer-major, %d chunks, %.1f GB ""of RAM)", (unsigned long long) c.it_lc, (int) run.T.size(), (double) c.lc_bytes / 1e9);- char dbgs[160] = "";+ char dbgs[200] = "";if (pm.on)- std::snprintf(dbgs, sizeof dbgs, "; copy_ahead: staging %.0f ms, DMA issues %.0f ms, slot waits %.0f ms",- c.dbg_stage_ms, c.dbg_dma_ms, c.dbg_evt_ms);+ std::snprintf(dbgs, sizeof dbgs,+ "; io threads busy %.0f ms, copy_ahead staging %.0f ms, DMA issues %.0f ms, slot waits %.0f ms",+ c.io->busy_ms() - c.io_busy0, c.dbg_stage_ms, c.dbg_dma_ms, c.dbg_evt_ms);std::fprintf(stderr,"ds4 prompt stream, CUDA%d: %d tokens in %.0f ms; %llu experts streamed: %llu from the files ""(%.1f GB, %.2f GB/s: %.1f GB from the disk unbuffered, %.1f GB through the page cache), %llu "