Skip to content

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++

Fix #3469, #3437: UPDATE.bat and setup fetch the engine from the signed update feed when the release has no flat zip, and the minimum engine is 0.1.46

Niko12215cdb59cfa987
diff --git a/docs/AMD_HIP.md b/docs/AMD_HIP.md
index 9330f0cd7..7fac6fdc0 100644
--- a/docs/AMD_HIP.md
+++ b/docs/AMD_HIP.md
@@ -145,7 +145,7 @@ On a PC with no NVIDIA card Strata can use, the AMD card is chosen by itself; wi
- **Detection:** setup reads the display adapters Windows lists (their PCI ids; the VRAM size from the display
driver's registry entry). An integrated Radeon is listed as not supported, except the Ryzen AI Max "Strix Halo" (Radeon 8060S / 8050S / 8040S, gfx1151: [STRIX_HALO.md](STRIX_HALO.md)); a Strix Point (890M / 880M, gfx1150), Krackan (860M / 840M, gfx1152) or Phoenix / Hawk Point (780M / 760M, gfx1103) Radeon is named as what it is and not supported (on Linux the 780M / 760M has an opt-in build: [RADEON_780M.md](RADEON_780M.md)).
- **Engine (0.2.0 and later):** the installers and `strata update` use the engine packages in [Ready-made packages](#ready-made-packages-020); the zip described next is the START-HERE route of 0.1.x.
-- **Engine (START-HERE, 0.1.x):** the ready-made `strata-windows-x64-hip.zip` from the release (built by `tools\hip\build_windows.bat`
+- **Engine (START-HERE, 0.1.x):** the ready-made engine package for your card, fetched from the signed update feed (releases from 0.2 carry no flat `strata-windows-x64-hip.zip`; that zip is what `build_windows.bat` makes for a self-build) (built by `tools\hip\build_windows.bat`
for gfx1100, gfx1101, gfx1102, gfx1200, gfx1201, gfx1030 and, from 0.1.40, gfx1151 (Strix Halo, unvalidated on Windows)) goes into `engine\`. It carries the ROCm libraries the
engine loads (`engine\rocm\bin`: the HIP runtime, hipBLAS / rocBLAS / hipBLASLt with their kernels for these cards,
amd_comgr and the Microsoft C++ runtime; ROCm 10.2.0a20260930 from AMD's TheRock builds, licenses in
diff --git a/setup.py b/setup.py
index 047b00316..dad6c060a 100644
--- a/setup.py
+++ b/setup.py
@@ -241,7 +241,7 @@ CUDA12_WHEELS = ["nvidia-cublas-cu12==12.9.1.4", "nvidia-cuda-runtime-cu12==12.9
# toolkit the CUDA 12 zip is built with (cuBLAS 12.9.1.4, runtime 12.9.79). Not tested on such an old driver here.
CUDA12_MIN_DRIVER = 528 if WIN else 525
ENGINE12_DIR = "engine-cuda12"
-MIN_ENGINE = (0, 1, 45) # versions compare all numbers; v0.1.45: Qwen3.6 parallel requests and pictures, fixes #2346 #2372 #2396; v0.1.44.1: CPUs without AVX2 start again (#2324: an illegal instruction before main in the NVFP4 CPU rows), the CUDA 12.0/12.1 toolkits build (#2322), a crashed engine start is named instead of retried as a GPU runtime start, Mac M2 (#2323); v0.1.44: UD-Q4/Q5/Q6_K_XL and Q8_0 Unsloth files in setup (K-quant experts, MMQ prompt library strata_kq), NVFP4 packs, Qwen3.6-35B-A3B, Intel Arc A-series kernels, the pipelined-decode and latency work; v0.1.43.1: the adaptive overlap no longer collapses decode on pinned-host desktops (#1964 #1975), Windows PLE mmap stall (#1984), pinned prompt uploads (#1996), the I/O watchdog allowance; v0.1.43: adaptive-tier copies overlap on CUDA, P100 (sm_60) kernel tables, RDNA4 (gfx1200/gfx1201) defaults (route tail skip, measured PCIe share, 16K prompt chunks, fused IQ3 prompt experts); v0.1.42.1: the GGUF-in-place file tier on Windows unmaps every shard, dual-socket starts judge each NUMA node and STRATA_DROP_CACHE_ON_EXIT, tools/strata_bench.py; v0.1.42: STRATA_ROUTE_TAIL_SKIP=7 default on CUDA (+13-16% decode), the AMD 2-GPU checkpoint fix (H13), Phase 2 fixes and opt-ins; v0.1.41.1 (folded into 0.1.42): the #1576 card order, #1616 stage size, #1595 CPU share, #1630/#1607 Windows HIP chunk cap and commit-limit warning, #1603 parallel-request bookkeeping; v0.1.41: CPU prefill share on by default (NVIDIA, chunks < 1024), --batch-groups auto, the file tier read in batches on Windows, Pascal decode fix, per-arch tables; v0.1.40.3: the #1357 MTP router guard, the #1376 Windows HIP cache floor, #461 runtime DLLs, Intel A750 first-request fix; v0.1.40.2: F4 verify windows, the Linux file tier (#1194), the stager wait (#1057), #1264/#1201/#1139 fixes, opt-in CPU share (#1282), Intel Arc; v0.1.40: --resident-experts on a layer split with the split+resident variant (#848), --kv k8v4 with KV streaming (#711); v0.1.39: the #577 file-tier regression fixed, the OpenAI Responses API (#451, Codex), a reply stuck on one token ended (#606), the head before the arena (#620), effort_position (#458), --vram-reserve hot resize opt-in (#533), PR batch; v0.1.38: prompts faster (one gather per expert group #372, the first chunk's PLE rows beside layer 0 #374, DeltaNet three heads per thread #413), --kv q4_0 prompts on tensor cores (#452), Q5_0 experts on the GPU (#473), IQ4_XS on AVX-2 (#415), unbuffered expert loading on Windows (#357 #362), --peer-device (#531), a 6 GB card starts (#496), PR batch; v0.1.37: a silent engine is restarted (#481), Windows AMD counts the desktop's VRAM (#380 #377 #497), a steadier PCIe probe (#485), fixes #496 #495 #498 #505 #493; v0.1.36: a cancelled prompt logged as read so far (#471), the draft-head hint (#474), UPDATE.bat (#475), --expert-profile-save (#477); v0.1.35: Windows AMD uses its bundled HIP runtime (#468 #461), the low-RAM resident mode on Windows 32 GB (#467), fixes #460 #459 #446 #447 #457 #448 #444; v0.1.34: AMD on Windows (a ready-made HIP engine), an MCP server for AI assistants (tools/strata_mcp.py), a shorter README; v0.1.33: a portable image encoder again (#411 #412), setup recommends instead of forcing (#406 #403 #364 #384), fixes #352 #365 #369 #371 #375 #393 #408 #414; v0.1.32: split prompts faster (#340), AMD router +12%, Unsloth Q4 in setup, faster Q4 prompts, #326/#327/#342/#344 fixes, PR batch; v0.1.31: Unsloth UD-Q4_K_XL (experimental), GGUF-in-place low-RAM mode, Windows GGUF load 2x, server race + tokenizer fixes, AMD intrinsics; v0.1.30: short prompts faster (streaming from 1024 tokens), resident low-RAM variant, multi-GPU session carve, RDNA4; v0.1.29: sampled answers faster (split top-k), #154 correctness fixes; v0.1.28: the expert cache reserves the draft head, a cancelled request no longer fails the next; v0.1.27: RTX 20 (sm_75) in the ready-made engine, the HIP build without CUDA headers; v0.1.26: the draft layer's prompt pass in batches; v0.1.25: faster prompts (grouping off the copy engine, fused hyper-connection kernels), AMD HIP backend, --kv k8v4; v0.1.24: long prompts faster (QSA select on tensor cores); v0.1.23: image requests honor sampling, 8 GB cards start, batched verify window; v0.1.22: faster prompts (tensor-core attention), multi-GPU across images/steering/KV streaming; v0.1.21: multi-GPU layer split (--gpus); v0.1.20: system-prompt checkpoint, PCIe probe, hit rate; v0.1.19: penalties
+MIN_ENGINE = (0, 1, 46) # versions compare all numbers; v0.1.46: the 0.2 release engine (UPDATE.bat and setup fetch it from the signed feed, #3469 #3437); v0.1.45: Qwen3.6 parallel requests and pictures, fixes #2346 #2372 #2396; v0.1.44.1: CPUs without AVX2 start again (#2324: an illegal instruction before main in the NVFP4 CPU rows), the CUDA 12.0/12.1 toolkits build (#2322), a crashed engine start is named instead of retried as a GPU runtime start, Mac M2 (#2323); v0.1.44: UD-Q4/Q5/Q6_K_XL and Q8_0 Unsloth files in setup (K-quant experts, MMQ prompt library strata_kq), NVFP4 packs, Qwen3.6-35B-A3B, Intel Arc A-series kernels, the pipelined-decode and latency work; v0.1.43.1: the adaptive overlap no longer collapses decode on pinned-host desktops (#1964 #1975), Windows PLE mmap stall (#1984), pinned prompt uploads (#1996), the I/O watchdog allowance; v0.1.43: adaptive-tier copies overlap on CUDA, P100 (sm_60) kernel tables, RDNA4 (gfx1200/gfx1201) defaults (route tail skip, measured PCIe share, 16K prompt chunks, fused IQ3 prompt experts); v0.1.42.1: the GGUF-in-place file tier on Windows unmaps every shard, dual-socket starts judge each NUMA node and STRATA_DROP_CACHE_ON_EXIT, tools/strata_bench.py; v0.1.42: STRATA_ROUTE_TAIL_SKIP=7 default on CUDA (+13-16% decode), the AMD 2-GPU checkpoint fix (H13), Phase 2 fixes and opt-ins; v0.1.41.1 (folded into 0.1.42): the #1576 card order, #1616 stage size, #1595 CPU share, #1630/#1607 Windows HIP chunk cap and commit-limit warning, #1603 parallel-request bookkeeping; v0.1.41: CPU prefill share on by default (NVIDIA, chunks < 1024), --batch-groups auto, the file tier read in batches on Windows, Pascal decode fix, per-arch tables; v0.1.40.3: the #1357 MTP router guard, the #1376 Windows HIP cache floor, #461 runtime DLLs, Intel A750 first-request fix; v0.1.40.2: F4 verify windows, the Linux file tier (#1194), the stager wait (#1057), #1264/#1201/#1139 fixes, opt-in CPU share (#1282), Intel Arc; v0.1.40: --resident-experts on a layer split with the split+resident variant (#848), --kv k8v4 with KV streaming (#711); v0.1.39: the #577 file-tier regression fixed, the OpenAI Responses API (#451, Codex), a reply stuck on one token ended (#606), the head before the arena (#620), effort_position (#458), --vram-reserve hot resize opt-in (#533), PR batch; v0.1.38: prompts faster (one gather per expert group #372, the first chunk's PLE rows beside layer 0 #374, DeltaNet three heads per thread #413), --kv q4_0 prompts on tensor cores (#452), Q5_0 experts on the GPU (#473), IQ4_XS on AVX-2 (#415), unbuffered expert loading on Windows (#357 #362), --peer-device (#531), a 6 GB card starts (#496), PR batch; v0.1.37: a silent engine is restarted (#481), Windows AMD counts the desktop's VRAM (#380 #377 #497), a steadier PCIe probe (#485), fixes #496 #495 #498 #505 #493; v0.1.36: a cancelled prompt logged as read so far (#471), the draft-head hint (#474), UPDATE.bat (#475), --expert-profile-save (#477); v0.1.35: Windows AMD uses its bundled HIP runtime (#468 #461), the low-RAM resident mode on Windows 32 GB (#467), fixes #460 #459 #446 #447 #457 #448 #444; v0.1.34: AMD on Windows (a ready-made HIP engine), an MCP server for AI assistants (tools/strata_mcp.py), a shorter README; v0.1.33: a portable image encoder again (#411 #412), setup recommends instead of forcing (#406 #403 #364 #384), fixes #352 #365 #369 #371 #375 #393 #408 #414; v0.1.32: split prompts faster (#340), AMD router +12%, Unsloth Q4 in setup, faster Q4 prompts, #326/#327/#342/#344 fixes, PR batch; v0.1.31: Unsloth UD-Q4_K_XL (experimental), GGUF-in-place low-RAM mode, Windows GGUF load 2x, server race + tokenizer fixes, AMD intrinsics; v0.1.30: short prompts faster (streaming from 1024 tokens), resident low-RAM variant, multi-GPU session carve, RDNA4; v0.1.29: sampled answers faster (split top-k), #154 correctness fixes; v0.1.28: the expert cache reserves the draft head, a cancelled request no longer fails the next; v0.1.27: RTX 20 (sm_75) in the ready-made engine, the HIP build without CUDA headers; v0.1.26: the draft layer's prompt pass in batches; v0.1.25: faster prompts (grouping off the copy engine, fused hyper-connection kernels), AMD HIP backend, --kv k8v4; v0.1.24: long prompts faster (QSA select on tensor cores); v0.1.23: image requests honor sampling, 8 GB cards start, batched verify window; v0.1.22: faster prompts (tensor-core attention), multi-GPU across images/steering/KV streaming; v0.1.21: multi-GPU layer split (--gpus); v0.1.20: system-prompt checkpoint, PCIe probe, hit rate; v0.1.19: penalties
# KV bytes per context token and attention layer: 8-bit 1056, rotated 4-bit 576, hybrid K8V4 (8-bit K, 4-bit V) 816
KV_CELL_BYTES = {"q4_0": 576, "k8v4": 816}
PY_PACKAGES = ["numpy", "jinja2", "regex", "pyyaml", "tqdm", "requests", "pillow", "psutil"]
@@ -3792,7 +3792,7 @@ def amd_parse_gpus(text, amd) -> list:
# loads, built by tools/hip/build_windows.bat); it needs only the AMD driver. Once it is installed, the cards are
# numbered as the HIP runtime numbers them (`strata-device --list-devices`): an integrated Radeon takes HIP's device 0
# and pushes the discrete card to 1, which the display-adapter order does not show (#325).
-WIN_HIP_ASSET = "strata-windows-x64-hip.zip"
+WIN_HIP_ASSET = "strata-windows-x64-hip.zip" # releases up to 0.1.45; 0.2 releases publish per-card packages in the signed feed (engine_from_feed)
WIN_HIP_MIN_ENGINE = max(MIN_ENGINE, (0, 1, 33)) # the first release with a Windows HIP engine
WIN_AMD_DRIVER = "https://www.amd.com/en/support/download/drivers.html"
# PCI device ids (VEN_1002) of the cards the AMD backend knows; the names below cover a card whose id is not listed
@@ -4071,6 +4071,57 @@ def hip_card(eng: Path, gpu: dict, listed: list[dict]) -> dict:
return {**gpu, "index": m["index"], "count": len(hip), "vram_gb": vram, "driver": "hip"}
+def engine_from_feed(eng: Path, updating: bool = False) -> Path | None:
+ """#3469 #3437: the 0.2 releases no longer carry the flat engine zips (strata-windows-x64*.zip); they publish one package
+ per card in the signed update feed. `strata update --no-control` fetches, verifies and test-starts the package for this PC
+ (into a folder of its own under the install), and its files then become the installed engine in `eng`, the same way a
+ zip's do. None when it cannot be had (no internet, nothing for this card): the caller goes on as before."""
+ acquisition.require("the ready-made engine")
+ from strata_control import engine_component as ec
+ store = home() / "engine-feed"
+ env = dict(os.environ)
+ env.pop("STRATA_NO_ACQUIRE", None)
+ env["STRATA_COMPONENTS_DIR"] = str(store)
+ say(" Fetching the ready-made engine from the signed update feed ...")
+ try:
+ run = subprocess.run([sys.executable, str(ROOT / "strata.py"), "update", "--no-control", "--output", "json"],
+ env=env, cwd=str(ROOT), capture_output=True, text=True, timeout=3600)
+ except (OSError, subprocess.SubprocessError) as e:
+ warn(f"the update feed could not be run ({e})")
+ return None
+ if run.returncode != 0:
+ why = (run.stderr or run.stdout or "").strip().splitlines()[-1:] or [f"exit {run.returncode}"]
+ warn(f"no engine from the update feed: {why[0][:200]}")
+ return None
+ try:
+ component = ec.resolve(store, full=True)
+ except ec.ComponentProblem as e:
+ warn(f"the engine from the update feed is not usable: {e}")
+ return None
+ eng.mkdir(exist_ok=True)
+ tmp = eng / "_unpack"
+ shutil.rmtree(tmp, ignore_errors=True)
+ shutil.copytree(component.root, tmp)
+ try:
+ meta = json.loads((tmp / "BUILD.json").read_text(encoding="utf-8"))
+ except (OSError, ValueError):
+ meta = {}
+ if meta.get("source") != "prebuilt": # the setup's own books: a ready-made engine, not one compiled here
+ meta["source"] = "prebuilt"
+ (tmp / "BUILD.json").write_text(json.dumps(meta, indent=1), encoding="utf-8")
+ (eng / "BUILD.json").replace(eng / "BUILD.json.prev") if (eng / "BUILD.json").exists() else None
+ install_unpacked(tmp, eng)
+ shutil.rmtree(tmp, ignore_errors=True)
+ if not (eng / EXE).exists():
+ fail("the engine package from the update feed has no " + EXE)
+ if not WIN:
+ for x in (EXE, VEXE):
+ if (eng / x).exists():
+ (eng / x).chmod(0o755)
+ ok(f"ready-made engine {meta.get('version', '')} from the update feed ({component.root.name})")
+ return eng
+
+
def get_prebuilt_hip(url_base, gpu, updating=False) -> Path | None:
"""The ready-made Windows HIP engine (WIN_HIP_ASSET) in engine/, kept between runs; None when it cannot be had
(not published for this version, no internet) or has no code for the card."""
@@ -4104,7 +4155,7 @@ def get_prebuilt_hip(url_base, gpu, updating=False) -> Path | None:
say(f" No ready-made AMD engine for v{source_version()} ({e}): the latest release instead")
continue
warn(f"no ready-made AMD engine at {base} ({e})")
- return None
+ return engine_from_feed(eng, updating)
say(" Downloading the ready-made Strata engine for AMD GPUs (with the ROCm libraries it uses) ...")
download(base + WIN_HIP_ASSET, z, "Strata AMD engine")
try:
@@ -4879,6 +4930,10 @@ def get_prebuilt(url_base, gpu, vision, updating=False, toolkit=13) -> Path | No
if i + 1 < len(bases): # #214: this checkout's release is not published (yet)
say(f" No ready-made engine for v{source_version()} ({e}): the latest release instead")
continue
+ if int(toolkit) != 12: # 0.2 releases: the signed feed has it (#3469)
+ got = engine_from_feed(eng, updating)
+ if got is not None:
+ return got
warn(f"no ready-made engine at {base} ({e})" + ("" if updating else ": compiling instead"))
return None
say(" Downloading the ready-made Strata engine" + (" (CUDA 12, experimental)" if int(toolkit) == 12 else "") + " ...")
diff --git a/tools/test_setup_feed_engine.py b/tools/test_setup_feed_engine.py
new file mode 100644
index 000000000..8b961633c
--- /dev/null
+++ b/tools/test_setup_feed_engine.py
@@ -0,0 +1,83 @@
+"""#3469 #3437: the 0.2 releases carry no flat engine zip, so a source checkout (UPDATE.bat, setup) fetches the engine package
+for its card from the signed update feed and installs its files in engine/.
+
+ python -m unittest tools.test_setup_feed_engine
+"""
+from __future__ import annotations
+
+import json
+import sys
+import tempfile
+import types
+import unittest
+from pathlib import Path
+from unittest import mock
+
+ROOT = Path(__file__).resolve().parents[1]
+sys.path.insert(0, str(ROOT))
+sys.path.insert(0, str(ROOT / "tools"))
+import setup # noqa: E402
+from strata_control import engine_component as ec # noqa: E402
+
+
+class FeedEngine(unittest.TestCase):
+ def setUp(self):
+ self.tmp = tempfile.TemporaryDirectory()
+ self.addCleanup(self.tmp.cleanup)
+ self.root = Path(self.tmp.name)
+ self.eng = self.root / "engine"
+ self.eng.mkdir()
+ (self.eng / "BUILD.json").write_text(json.dumps({"version": "0.1.45", "source": "prebuilt", "archs": [89]}), encoding="utf-8")
+ (self.eng / setup.EXE).write_bytes(b"old")
+ self.package = self.root / "components" / "engine" / "0.1.46-test"
+ self.package.mkdir(parents=True)
+ (self.package / "BUILD.json").write_text(json.dumps({"version": "0.1.46", "archs": [89]}), encoding="utf-8")
+ (self.package / setup.EXE).write_bytes(b"new")
+ for name, value in (("home", lambda: self.root), ("say", lambda *a, **k: None), ("ok", lambda *a, **k: None),
+ ("warn", lambda *a, **k: None)):
+ patch = mock.patch.object(setup, name, value)
+ patch.start()
+ self.addCleanup(patch.stop)
+ patch = mock.patch.object(setup.acquisition, "require", lambda what: None)
+ patch.start()
+ self.addCleanup(patch.stop)
+
+ def run_update(self, code=0):
+ return mock.patch.object(setup.subprocess, "run", return_value=types.SimpleNamespace(returncode=code, stdout="", stderr="no feed"))
+
+ def test_the_package_becomes_the_installed_engine(self):
+ with self.run_update() as run, mock.patch.object(ec, "resolve", return_value=types.SimpleNamespace(root=self.package)):
+ got = setup.engine_from_feed(self.eng, updating=True)
+ self.assertEqual(got, self.eng)
+ cmd = run.call_args.args[0]
+ self.assertIn("update", cmd)
+ self.assertIn("--no-control", cmd)
+ self.assertNotIn("STRATA_NO_ACQUIRE", run.call_args.kwargs["env"]) # this one step may fetch
+ self.assertEqual((self.eng / setup.EXE).read_bytes(), b"new")
+ meta = json.loads((self.eng / "BUILD.json").read_text(encoding="utf-8"))
+ self.assertEqual((meta["version"], meta["source"]), ("0.1.46", "prebuilt"))
+ self.assertEqual((self.eng / ".previous" / setup.EXE).read_bytes(), b"old") # one generation to roll back to
+
+ def test_no_feed_keeps_the_installed_engine(self):
+ with self.run_update(code=6):
+ self.assertIsNone(setup.engine_from_feed(self.eng, updating=True))
+ self.assertEqual((self.eng / setup.EXE).read_bytes(), b"old")
+
+ def test_a_release_without_the_zip_falls_back_to_the_feed(self):
+ gpu = {"arch": 89, "archs": [89], "vram_gb": 8.0, "index": 0}
+ (self.eng / "BUILD.json").write_text(json.dumps({"version": "0.1.45", "source": "prebuilt", "archs": [89]}), encoding="utf-8")
+ with mock.patch.object(setup, "engine_dir", lambda toolkit=13: self.eng), mock.patch.object(setup, "_urlopen", side_effect=OSError("HTTP Error 404")), mock.patch.object(setup, "engine_from_feed", return_value=self.eng) as feed:
+ got = setup.get_prebuilt(setup.PREBUILT_URL, gpu, "none", updating=True)
+ self.assertEqual(got, self.eng)
+ feed.assert_called_once()
+ feed.reset_mock()
+ setup.get_prebuilt(setup.PREBUILT_URL, gpu, "none", updating=True, toolkit=12)
+ feed.assert_not_called() # the experimental CUDA 12 engine has no feed package here
+
+ def test_the_minimum_is_the_0_2_engine(self):
+ self.assertLess((0, 1, 45), setup.MIN_ENGINE)
+ self.assertGreaterEqual((0, 1, 46), setup.WIN_HIP_MIN_ENGINE)
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/tools/test_setup_hotfix_tag.py b/tools/test_setup_hotfix_tag.py
index be4a8d379..fc22e39a4 100644
--- a/tools/test_setup_hotfix_tag.py
+++ b/tools/test_setup_hotfix_tag.py
@@ -23,7 +23,8 @@ class HotfixTag(unittest.TestCase):
ver = tuple(int(x) for x in "0.1.40.3".split(".")) # a hotfix version has four numbers; MIN_ENGINE may have three
self.assertEqual(len(ver), 4)
self.assertLess(ver, setup.MIN_ENGINE) # an installed hotfix of the older release is replaced
- self.assertGreaterEqual((0, 1, 45), setup.MIN_ENGINE) # the minimum itself (or a later hotfix) is accepted
+ self.assertGreaterEqual((0, 1, 46), setup.MIN_ENGINE) # the minimum itself (or a later hotfix) is accepted
+ self.assertLess((0, 1, 45), setup.MIN_ENGINE) # 0.1.45 is replaced (#3469)
self.assertLess((0, 1, 44, 1), setup.MIN_ENGINE) # the 0.1.44.1 hotfix is replaced
self.assertLess((0, 1, 44), setup.MIN_ENGINE) # the release before is replaced
self.assertLess((0, 1, 43, 1), setup.MIN_ENGINE) # the 0.1.43.1 hotfix is replaced
diff --git a/tools/test_setup_menu.py b/tools/test_setup_menu.py
index 284f77039..d59e9d19c 100644
--- a/tools/test_setup_menu.py
+++ b/tools/test_setup_menu.py
@@ -153,7 +153,7 @@ class UnslothDefaultEnv(unittest.TestCase):
class EnginePlaceholders(unittest.TestCase):
def test_every_engine_minimum_is_this_release(self):
self.assertEqual(setup.NVFP4_ENGINE, (0, 1, 44))
- self.assertEqual(setup.MIN_ENGINE, (0, 1, 45))
+ self.assertEqual(setup.MIN_ENGINE, (0, 1, 46))
for m in ("UD-Q5_K_XL", "UD-Q6_K_XL", "NVFP4", "NVFP4-NVIDIA"):
self.assertEqual(setup.MODELS[m]["engine"], (0, 1, 44), m)
# No minimum may name an engine newer than this tree (a placeholder for a release that does not exist yet would make setup