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++
Gyro on the Mac: model list tests, resident wrap span without the n-gram table, trellis gate/up tile
Niko12210b736f7702aa
diff --git a/metal/src/engine/window.cpp b/metal/src/engine/window.cppindex 25a892907..db875e8b6 100644--- a/metal/src/engine/window.cpp+++ b/metal/src/engine/window.cpp@@ -459,7 +459,9 @@ bool Window::init(const strata::core::ModelGeometry& g, strata::core::SessionSta// the chunks of a model bigger than one Metal buffer overlap by the largest range bound at one offset: a tensor,// or a layer's routed experts (gate, up and down, bound together at the first of them: moe_rebase)size_t span = 0;- for (const auto& kv : f_.tensors) span = std::max<size_t>(span, kv.second.bytes);+ // (not the n-gram table: the host reads it from the SSD, the GPU never binds it - in a Gyro file it is 28.8 GB of the one file)+ for (const auto& kv : f_.tensors)+ if (kv.first != "per_layer_token_embd.weight") span = std::max<size_t>(span, kv.second.bytes);for (int l = 0; l < 48; ++l) {const std::string b = "blk." + std::to_string(l) + ".";const gguf::Tensor *g = f_.find(b + "ffn_gate_exps.weight"), *u = f_.find(b + "ffn_up_exps.weight"),@@ -896,7 +898,10 @@ void Window::layer(std::vector<mtl::Dispatch>& ds, int l, int ntok, int64_t pos0a.topk = TOPK;a.n_expert = NEXP;const bool dq2 = td.type == ref::Q2_0;- const int gl = 32, gr = 1, gs = 4, ds_ = 4;+ // the trellis gate/up: 16 lanes a row, 2 rows a lane (expert_parity --bench --sweep, M4, TQK6, 10 experts: 166.6 us a layer+ // against 174.5 for 32,2,4 and 196.8 for 32,1,4; also the fastest at 2 and 3 tokens)+ const bool tq_ = tg_.type >= 144;+ const int gl = tq_ ? 16 : 32, gr = tq_ ? 2 : 1, gs = 4, ds_ = 4;// the down tile: Q2_0 8,1; else 4,2 - but on the M4 (no M5 layout) Q5_1 4,1 and Q8_0 8,1 (4,2 spills there:// 24-36 GB/s, mac-kern 513fc615)const bool m4 = !m5_layout();diff --git a/tools/test_strata_mcp.py b/tools/test_strata_mcp.pyindex 7634c8d7b..df5be39bc 100644--- a/tools/test_strata_mcp.py+++ b/tools/test_strata_mcp.py@@ -726,18 +726,22 @@ class MacCatalog(unittest.TestCase):def test_the_model_list_is_the_macs_sizes_with_unified_memory_verdicts(self):s, hw = self.hw(32)fams = s.model_table(hw)- self.assertEqual([f["family"] for f in fams], ["qwen"], "no Unsloth, Swift or NVFP4 on a Mac in the app")+ self.assertEqual([f["family"] for f in fams], ["qwen", "gyro"], "no Unsloth, Swift or NVFP4 on a Mac in the app; Gyro after the regular sizes")by = {x["model"]: x for x in fams[0]["sizes"]}self.assertEqual(list(by), ["Q2_0", "IQ3_XXS", "IQ3_S"])+ gy = {x["model"]: x for x in fams[1]["sizes"]}+ self.assertEqual(list(gy), ["Gyro-S", "Gyro-M"])+ self.assertEqual(gy["Gyro-S"]["download_gb"], 58.5)+ self.assertTrue(all(v["on_this_pc"] and "SSD" in v["on_this_pc"] for v in gy.values()), "32 GB: both stream (said, not refused)")self.assertIn("read from the SSD", by["Q2_0"]["on_this_pc"])self.assertIn("slower", by["Q2_0"]["on_this_pc"])self.assertIn("~14 tok/s", by["Q2_0"]["on_this_pc"])self.assertTrue(by["IQ3_S"]["on_this_pc"].startswith("slow here"))self.assertTrue(all(x["images"] for x in by.values()), "the Mac engine carries the Metal picture reader: pictures are on")s, hw = self.hw(128)- self.assertEqual({x["on_this_pc"] for x in s.model_table(hw)[0]["sizes"]}, {"fits"})+ self.assertEqual({x["on_this_pc"] for f in s.model_table(hw) for x in f["sizes"]}, {"fits"})s, hw = self.hw(8)- self.assertEqual({x["on_this_pc"] for x in s.model_table(hw)[0]["sizes"]}, {"does not fit"})+ self.assertEqual({x["on_this_pc"] for f in s.model_table(hw) for x in f["sizes"]}, {"does not fit"})def test_the_install_plan_on_a_mac_has_no_card_kv_or_ram_options(self):s, hw = self.hw(32)