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++
tests: three suite gaps a live PC showed (#3236)
Norm5b14fe23f7ea
diff --git a/tools/test_conversation_cache_disabled.py b/tools/test_conversation_cache_disabled.pyindex d389ac9f9..59c3e6851 100644--- a/tools/test_conversation_cache_disabled.py+++ b/tools/test_conversation_cache_disabled.py@@ -8,6 +8,8 @@ import sysimport tempfileimport unittest+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_disabled import NAMES, SETTINGS, STATE_KEYS, common_args, state_hashes, verifydiff --git a/tools/test_conversation_cache_growth.py b/tools/test_conversation_cache_growth.pyindex 364234491..f89f2be82 100644--- a/tools/test_conversation_cache_growth.py+++ b/tools/test_conversation_cache_growth.py@@ -8,6 +8,8 @@ import sysimport tempfileimport unittest+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_growth import NAMES, SECRET, SETTINGS, STATE_KEYS, verifydiff --git a/tools/test_conversation_cache_http_smoke.py b/tools/test_conversation_cache_http_smoke.pyindex fd0a79d2d..50211a87d 100644--- a/tools/test_conversation_cache_http_smoke.py+++ b/tools/test_conversation_cache_http_smoke.py@@ -9,6 +9,8 @@ import tempfileimport unittestfrom unittest.mock import Mock+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_http_smoke import rundiff --git a/tools/test_conversation_cache_isolation.py b/tools/test_conversation_cache_isolation.pyindex 7205d4d4d..d87e9287c 100644--- a/tools/test_conversation_cache_isolation.py+++ b/tools/test_conversation_cache_isolation.py@@ -7,6 +7,8 @@ import sysimport tempfileimport unittest+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_isolation import STATE_KEYS, fixtures, verifyfrom tools.gguf_reader import GGUFFilediff --git a/tools/test_conversation_cache_parity.py b/tools/test_conversation_cache_parity.pyindex 3f950ebd9..b9a790134 100644--- a/tools/test_conversation_cache_parity.py+++ b/tools/test_conversation_cache_parity.py@@ -6,6 +6,8 @@ import sysimport tempfileimport unittest+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_parity import STATE_KEYS, engine_args, pressure_budget_hint, verify_admission, verify_exchange, verify_pressure, verify_resultsdiff --git a/tools/test_conversation_cache_soak.py b/tools/test_conversation_cache_soak.pyindex 7d815890b..7ac12a6d7 100644--- a/tools/test_conversation_cache_soak.py+++ b/tools/test_conversation_cache_soak.py@@ -6,6 +6,8 @@ import sysimport tempfileimport unittest+sys.path.insert(0, str(Path(__file__).resolve().parents[1])) # runnable as tools/test_x.py from the repo root too+from tools.conversation_cache_soak import STATE_KEYS, answer_text, verifydiff --git a/tools/test_iq_pack.py b/tools/test_iq_pack.pyindex 0e0f5e8b2..f1faeabe4 100644--- a/tools/test_iq_pack.py+++ b/tools/test_iq_pack.py@@ -14,9 +14,25 @@ import numpy as npsys.path.insert(0, str(Path(__file__).resolve().parent)) # runnable from the repo root toofrom _paths import add_gguf_py-add_gguf_py()-from gguf import GGUFWriter, GGMLQuantizationType as Q, quants-import iq_pack+import iq_pack # needs numpy + gguf_reader only, not gguf-py: import it unguarded++try:+ add_gguf_py() # llama.cpp's gguf-py: setup.py clones it; gguf_py() exits when absent+ from gguf import GGUFWriter, GGMLQuantizationType as Q, quants+except SystemExit:+ GGUF_PY = None+else:+ GGUF_PY = "found"+++def require_gguf_py(test):+ """setUpClass for the classes that build GGUF fixtures: skip (not fail) on a checkout without llama.cpp."""+ @classmethod+ def setUpClass(cls):+ if GGUF_PY is None:+ raise unittest.SkipTest("llama.cpp's gguf-py not found: run setup.py (it clones llama.cpp into "+ "third_party/) or set STRATA_GGUF_PY")+ return setUpClassdef write_gguf(path, tensors, split=None, arch=True):@@ -37,6 +53,7 @@ def write_gguf(path, tensors, split=None, arch=True):class CompatibilityTests(unittest.TestCase):+ setUpClass = require_gguf_py(unittest.TestCase)def test_bf16_halfway_rounds_to_even(self):values = np.array([0x3F808000, 0x3F818000, 0xBF808000, 0xBF818000], dtype=np.uint32)got = np.frombuffer(iq_pack.bf16_bytes(values.view(np.uint8), "F32"), dtype=np.uint16)@@ -253,6 +270,8 @@ class ConversionTests(unittest.TestCase):"""CS2: FORM decides the engine's form; exact F32->BF16 by default, rounding only with --compat-bf16, the PLEconv F32 -> F16 by the loader (index kind 3), non-finite values refused, every conversion in conversions.json."""+ setUpClass = require_gguf_py(unittest.TestCase)+def test_exact_f32_to_bf16_by_default(self):with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmp:x = bits(np.arange(64, dtype=np.uint32).reshape(2, 32) << 16 | 0x3F000000) # BF16 values held as F32@@ -362,8 +381,9 @@ def expert(seed, shape):return np.random.default_rng(seed).standard_normal((N_EXPERT, *shape)).astype(np.float32)-def split_model(root, name="model", split_layer=True, dup=False, split_no=None, gate1=Q.Q8_0, seed=0):+def split_model(root, name="model", split_layer=True, dup=False, split_no=None, gate1=None, seed=0):"""Writes the 4 shards; returns their paths."""+ gate1 = Q.Q8_0 if gate1 is None else gate1 # not a default arg: the module imports without gguf-pygu, dn = (32, 32), (32, 32)router = [(f"blk.{l}.ffn_gate_inp.weight", np.full((N_EXPERT, 32), 0.5, np.float32), Q.F32) for l in (0, 1)]l0 = [(f"blk.0.ffn_{r}_exps.weight", expert(seed + 10 + i, gu if r != "down" else dn), Q.Q8_0)@@ -406,6 +426,7 @@ def expected_experts(model):class SplitArtifactTests(unittest.TestCase):+ setUpClass = require_gguf_py(unittest.TestCase)def test_four_shards_and_a_layer_split_per_role(self):with tempfile.TemporaryDirectory(ignore_cleanup_errors=True) as tmp:root = Path(tmp)@@ -689,6 +710,7 @@ class HadamardTests(unittest.TestCase):class NativeExpertsV5Tests(unittest.TestCase):"""The v5 mark (Hadamard-folded experts) rewrites only the tables the packer writes, v3 and v4."""+ setUpClass = require_gguf_py(unittest.TestCase) # test_packer_refuses_before_writing builds GGUF fixturesROWS = "0 145 146 0 100 1 2 3\n1 145 146 100 100 4 5 6\n"def table(self, version, tail="(n_expert 2, total 200)"):diff --git a/tools/test_strata_mcp.py b/tools/test_strata_mcp.pyindex 0fdb78bc8..fab2199f8 100644--- a/tools/test_strata_mcp.py+++ b/tools/test_strata_mcp.py@@ -164,8 +164,12 @@ class FakeRoot(unittest.TestCase):self.strata.disk_free = lambda p: 1000.0self.tools = M.Tools(self.strata)self.extra = []+ # the tests must never see a Strata server that happens to run on the real default port on this PC+ self._default_port = M.DEFAULT_PORT+ M.DEFAULT_PORT = free_port()def tearDown(self):+ M.DEFAULT_PORT = self._default_portfor p in self.extra:if p.poll() is None:p.kill()