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++

Shared speed-switch table and Terminal: DS4_DEFAULTS=026 with the DeepSeek off switch, the 12 GB NVIDIA row, and a Qwen "0.2.6 behaviour" row (Advanced only)

Niko1221211fca87df43
diff --git a/strata_control/speed_switches.py b/strata_control/speed_switches.py
index f39d314ec..8f9871cdf 100644
--- a/strata_control/speed_switches.py
+++ b/strata_control/speed_switches.py
@@ -26,14 +26,19 @@ class Switch:
label: str # the words the app shows
effect: str # one plain sentence: what it does and what the other choice does
env: tuple[tuple[str, str], ...] # the environment entries written when the switch is NOT in its default state
- only: str # "qwen-gpu" | "qwen-radeon-9000" | "deepseek": where it applies
+ only: str # "qwen-gpu" | "qwen-cuda-12gb" | "qwen-radeon-9000" | "deepseek" | "qwen-any": where it applies
default: str = "on" # "on" | "off": what a launch without the entries does
+ legacy: tuple[tuple[str, str], ...] = () # entries an earlier 0.2.7 test build wrote for the same state: they read as the switch being set, and are removed with it
+ terminal_only: bool = False # shown in the Terminal's Advanced table only, never in the app
SWITCHES: tuple[Switch, ...] = (
Switch("faster_long_prompts", "Faster long prompts",
"Reads very long prompts faster by approximating part of the work. Off by default: answers about long documents can miss or mix up details.",
(("STRATA_KVA", "1"),), "qwen-gpu", default="off"),
+ Switch("faster_12gb_nvidia", "Faster answers on 12 GB NVIDIA cards",
+ "Prefers the parts of the model already on the graphics card, which makes answers much faster on 12 GB cards. Turn off if answers look worse than before.",
+ (("STRATA_ROUTE_PRIOR", "0"),), "qwen-cuda-12gb"),
Switch("faster_radeon_9000", "Faster answers on Radeon RX 9000 / AI PRO R9700",
"Shares the work between the graphics card and the processor in a way that makes answers faster on these cards. "
"Turn off to go back to the previous setting, which is slower.",
@@ -41,8 +46,15 @@ SWITCHES: tuple[Switch, ...] = (
Switch("deepseek_speedups", "DeepSeek speed-ups",
"Reads long prompts faster and loads the parts of the model it needs faster. "
"Turn off if DeepSeek stops working or its answers look wrong; it then runs as in version 0.2.6.",
- (("DS4_FAST_ATTN", "0"), ("DS4_ATTN_SLAB", "0"), ("DS4_COPY_PIN", "0"), ("DS4_COPY_HUGETLB", "0"),
- ("DS4_STORE_SPARE", "0"), ("DS4_STORE_PUT", "0")), "deepseek"),
+ # DS4_DEFAULTS=026: the engine's master switch back to the 0.2.6 defaults; the six are what older engines (and the test builds) read
+ (("DS4_DEFAULTS", "026"), ("DS4_FAST_ATTN", "0"), ("DS4_ATTN_SLAB", "0"), ("DS4_COPY_PIN", "0"), ("DS4_COPY_HUGETLB", "0"),
+ ("DS4_STORE_SPARE", "0"), ("DS4_STORE_PUT", "0")), "deepseek",
+ legacy=(("DS4_FAST_ATTN", "0"), ("DS4_ATTN_SLAB", "0"), ("DS4_COPY_PIN", "0"), ("DS4_COPY_HUGETLB", "0"),
+ ("DS4_STORE_SPARE", "0"), ("DS4_STORE_PUT", "0"))),
+ # Troubleshooting only, in the Terminal's Advanced table and nowhere else: every Qwen 0.2.7 default back to its 0.2.6 behaviour.
+ Switch("qwen_defaults_026", "Qwen: 0.2.6 behaviour",
+ "Starts every Qwen speed setting new in 0.2.7 the way 0.2.6 did, for troubleshooting. Use it if something that worked before no longer does.",
+ (("STRATA_DEFAULTS", "026"),), "qwen-any", default="off", terminal_only=True),
)
BY_ID = {s.id: s for s in SWITCHES}
@@ -51,6 +63,20 @@ _GFX12 = re.compile(r"^gfx120[01]$", re.IGNORECASE)
_RADEON_9000 = re.compile(r"\b(?:RX\s*9\d{3}|AI\s*PRO\s*R97\d{2}|R97\d{2})\b", re.IGNORECASE)
+def _small_nvidia(gpus: Iterable[Mapping[str, Any]] | None) -> bool:
+ """A discrete NVIDIA card of 13 GiB or less (the engine's rule for residency-biased routing: not an integrated chip, not a card that shares system memory)."""
+ for g in gpus or []:
+ if g.get("usable") is False or g.get("uma") or g.get("integrated") or (g.get("unified_gb") or 0) > 0:
+ continue
+ vendor = str(g.get("vendor") or "").lower()
+ if vendor not in ("nvidia", "") or (vendor == "" and not re.search(r"nvidia|geforce|\brtx\b|\bgtx\b|quadro|tesla", str(g.get("name") or ""), re.IGNORECASE)):
+ continue
+ vram = float(g.get("vram_gb") or 0)
+ if 0 < vram <= 13.0:
+ return True
+ return False
+
+
def is_deepseek(cfg: Mapping[str, Any] | None) -> bool:
cfg = cfg or {}
return cfg.get("arch") == "deepseek41" or bool(re.search(r"strata-dsv41", str(cfg.get("exe") or ""), re.IGNORECASE))
@@ -74,11 +100,16 @@ def applies(switch: Switch | str, cfg: Mapping[str, Any] | None, gpus: Iterable[
return is_deepseek(cfg)
if is_deepseek(cfg):
return False
+ if sw.only == "qwen-any":
+ return True
backend = (cfg or {}).get("backend") or backend
if sw.only == "qwen-gpu":
# KVA was measured on the Qwen3.8-Flash-Next shape (48 layers, 512 experts); the engine declines any other. Setup writes "arch" only for the
# families with an engine path of their own (Qwen3.6-35B-A3B: qwen35moe), so a Flash-Next config (the original, Swift, Coder, Unsloth, Gyro, NVFP4) has none.
return not (cfg or {}).get("arch") and backend in ("cuda", "hip")
+ if sw.only == "qwen-cuda-12gb":
+ # the engine's residency-biased routing needs the Flash-Next shape (512 experts, top-10) and a CUDA build; a Qwen3.6 config has an "arch"
+ return not (cfg or {}).get("arch") and backend == "cuda" and _small_nvidia(gpus)
if sw.only == "qwen-radeon-9000":
return backend == "hip" and _radeon_9000(gpus)
return False
@@ -89,11 +120,17 @@ def applicable(cfg: Mapping[str, Any] | None, gpus: Iterable[Mapping[str, Any]]
return [s for s in SWITCHES if applies(s, cfg, gpus, backend)]
+def app_switches() -> list[Switch]:
+ """The switches the app shows (apps/host/server/native-settings.mjs SPEED_SWITCHES): all but the Terminal-only ones."""
+ return [s for s in SWITCHES if not s.terminal_only]
+
+
def is_set(env: Mapping[str, Any] | None, switch: Switch | str) -> bool:
"""Every entry the switch writes is there with its value (the switch is out of its default). A file with only some of them, or other values
(``STRATA_KVA=0``), is in the default state and is left alone."""
sw = BY_ID[switch] if isinstance(switch, str) else switch
- return all((env or {}).get(name) == value for name, value in sw.env)
+ have = lambda entries: bool(entries) and all((env or {}).get(name) == value for name, value in entries) # noqa: E731
+ return have(sw.env) or have(sw.legacy)
def is_on(env: Mapping[str, Any] | None, switch: Switch | str) -> bool:
@@ -112,6 +149,9 @@ def set_switch(env: Mapping[str, Any] | None, switch: Switch | str, on: bool) ->
out[name] = value
else:
out.pop(name, None)
+ if not write:
+ for name, _ in sw.legacy:
+ out.pop(name, None)
return out
diff --git a/tests/terminal/test_speed_switches.py b/tests/terminal/test_speed_switches.py
index 8f3386327..b7b596a88 100644
--- a/tests/terminal/test_speed_switches.py
+++ b/tests/terminal/test_speed_switches.py
@@ -21,11 +21,17 @@ ROOT = Path(__file__).resolve().parents[2]
ENV = { # what each switch writes outside its default
"faster_long_prompts": {"STRATA_KVA": "1"},
+ "faster_12gb_nvidia": {"STRATA_ROUTE_PRIOR": "0"},
"faster_radeon_9000": {"STRATA_PCIE_FRAC_DEFAULT": "old"},
- "deepseek_speedups": {"DS4_FAST_ATTN": "0", "DS4_ATTN_SLAB": "0", "DS4_COPY_PIN": "0", "DS4_COPY_HUGETLB": "0",
+ "deepseek_speedups": {"DS4_DEFAULTS": "026", "DS4_FAST_ATTN": "0", "DS4_ATTN_SLAB": "0", "DS4_COPY_PIN": "0", "DS4_COPY_HUGETLB": "0",
"DS4_STORE_SPARE": "0", "DS4_STORE_PUT": "0"},
+ "qwen_defaults_026": {"STRATA_DEFAULTS": "026"}, # the Terminal's troubleshooting switch: not in the app
}
-DEFAULT = {"faster_long_prompts": "off", "faster_radeon_9000": "on", "deepseek_speedups": "on"}
+DEFAULT = {"faster_long_prompts": "off", "faster_12gb_nvidia": "on", "faster_radeon_9000": "on", "deepseek_speedups": "on", "qwen_defaults_026": "off"}
+SIX = ("DS4_FAST_ATTN", "DS4_ATTN_SLAB", "DS4_COPY_PIN", "DS4_COPY_HUGETLB", "DS4_STORE_SPARE", "DS4_STORE_PUT")
+RTX3060 = {"name": "NVIDIA GeForce RTX 3060", "vendor": "nvidia", "arch": "86", "vram_gb": 12.0}
+RTX4070S = {"name": "NVIDIA GeForce RTX 4070 SUPER", "vendor": "nvidia", "arch": "89", "vram_gb": 12.0}
+RTX4060TI16 = {"name": "NVIDIA GeForce RTX 4060 Ti 16GB", "vendor": "nvidia", "arch": "89", "vram_gb": 16.0}
R9700 = {"name": "AMD Radeon AI PRO R9700", "vendor": "amd", "arch": "gfx1201"}
RX7900 = {"name": "AMD Radeon RX 7900 XTX", "vendor": "amd", "arch": "gfx1100"}
STRIX = {"name": "AMD Radeon 8060S", "vendor": "amd", "arch": "gfx1151", "uma": True}
@@ -37,11 +43,12 @@ DEEPSEEK = {"arch": "deepseek41", "exe": "C:/s/engine-dsv41/0.1.52-win/strata-ds
def ids(cfg, gpus, backend=None):
- return [s.id for s in W.applicable(cfg, gpus, backend)]
+ """The switches the app would show (the Terminal-only one is tested apart)."""
+ return [s.id for s in W.applicable(cfg, gpus, backend) if not s.terminal_only]
class Contract(unittest.TestCase):
- def test_three_switches_with_the_contract_names_and_their_defaults(self):
+ def test_five_switches_with_the_contract_names_and_their_defaults(self):
self.assertEqual([s.id for s in W.SWITCHES], list(ENV))
for s in W.SWITCHES:
self.assertEqual(dict(s.env), ENV[s.id], s.id)
@@ -49,7 +56,8 @@ class Contract(unittest.TestCase):
def test_an_effect_text_says_what_the_feature_does_in_plain_words(self):
for s in W.SWITCHES:
- self.assertRegex(s.effect, r"Off by default" if s.default == "off" else r"[Tt]urn off", s.id)
+ if not s.terminal_only:
+ self.assertRegex(s.effect, r"Off by default" if s.default == "off" else r"[Tt]urn off", s.id)
self.assertNotRegex(s.effect, r"STRATA_|DS4_|--|env", s.id)
self.assertEqual(W.SWITCHES[0].effect, "Reads very long prompts faster by approximating part of the work. Off by default: answers about long documents can miss or mix up details.")
@@ -67,6 +75,59 @@ class Contract(unittest.TestCase):
self.assertEqual(ids({**DEEPSEEK, "backend": "cuda"}, [RTX]), ["deepseek_speedups"])
self.assertEqual(ids({"exe": "C:/x/strata-dsv41.exe"}, [RTX]), ["deepseek_speedups"], "found by the engine's name too")
+ def test_faster_answers_on_12gb_nvidia_cards_only_where_the_engine_routing_default_applies(self):
+ """The engine turns residency-biased routing on by default on a DISCRETE NVIDIA card of 13 GiB or less in a CUDA build (the Flash-Next shape)."""
+ self.assertEqual(ids(QWEN, [RTX3060]), ["faster_long_prompts", "faster_12gb_nvidia"])
+ self.assertEqual(ids(QWEN, [RTX4070S]), ["faster_long_prompts", "faster_12gb_nvidia"])
+ self.assertEqual(ids(QWEN, [{**RTX3060, "vram_gb": 13.0}]), ["faster_long_prompts", "faster_12gb_nvidia"], "13 GiB counts")
+ self.assertEqual(ids(QWEN, [{**RTX3060, "vram_gb": 13.5}]), ["faster_long_prompts"])
+ self.assertEqual(ids(QWEN, [RTX4060TI16]), ["faster_long_prompts"], "16 GB is not a 12 GB card")
+ self.assertEqual(ids(QWEN, [RTX]), ["faster_long_prompts"])
+ self.assertEqual(ids(QWEN, [RTX3060, RTX]), ["faster_long_prompts", "faster_12gb_nvidia"], "one small card among others")
+ self.assertEqual(ids(QWEN, [{**RTX3060, "uma": True}]), ["faster_long_prompts"], "not a chip that shares system memory")
+ self.assertEqual(ids(QWEN, [{**RTX3060, "integrated": True}]), ["faster_long_prompts"])
+ self.assertEqual(ids(QWEN, [{**RTX3060, "usable": False}]), ["faster_long_prompts"])
+ self.assertEqual(ids({**QWEN, "backend": "hip"}, [{"name": "AMD Radeon RX 7700 XT", "vendor": "amd", "arch": "gfx1101", "vram_gb": 12.0}]), ["faster_long_prompts"], "not a Radeon")
+ self.assertEqual(ids({**QWEN, "backend": "sycl"}, [RTX3060]), [], "not a CUDA build")
+ self.assertEqual(ids(QWEN36, [RTX3060]), [], "Qwen3.6-35B-A3B: the routing needs 512 experts")
+ self.assertEqual(ids(DEEPSEEK, [RTX3060]), ["deepseek_speedups"])
+ self.assertEqual(ids(QWEN, [{"name": "GeForce RTX 3060", "vram_gb": 12.0}]), ["faster_long_prompts", "faster_12gb_nvidia"], "no vendor reported: the name decides")
+ self.assertEqual(ids(QWEN, [{"name": "Mystery card", "vram_gb": 12.0}]), ["faster_long_prompts"])
+
+ def test_the_12gb_switch_off_writes_route_prior_zero_and_on_removes_it(self):
+ env = {"PATH": "/bin"}
+ off = W.set_switch(env, "faster_12gb_nvidia", False)
+ self.assertEqual(off, {"PATH": "/bin", "STRATA_ROUTE_PRIOR": "0"})
+ self.assertFalse(W.is_on(off, "faster_12gb_nvidia"))
+ self.assertTrue(W.is_on({}, "faster_12gb_nvidia"))
+ self.assertEqual(W.set_switch(off, "faster_12gb_nvidia", True), env)
+ self.assertTrue(W.is_on({"STRATA_ROUTE_PRIOR": "2.0"}, "faster_12gb_nvidia"), "any other value is the user's own and reads as the default")
+
+ def test_deepseek_off_writes_the_master_switch_and_the_six_and_on_removes_all_seven(self):
+ env = {"DS4_ASYNC_T": "4"}
+ off = W.set_switch(env, "deepseek_speedups", False)
+ self.assertEqual(list(off)[:2], ["DS4_ASYNC_T", "DS4_DEFAULTS"], "the master switch first")
+ self.assertEqual(off, {"DS4_ASYNC_T": "4", "DS4_DEFAULTS": "026", **{k: "0" for k in SIX}})
+ self.assertFalse(W.is_on(off, "deepseek_speedups"))
+ self.assertEqual(W.set_switch(off, "deepseek_speedups", True), env)
+ # an earlier test build wrote only the six: it still reads as off, and turning it on removes them
+ six = {"DS4_ASYNC_T": "4", **{k: "0" for k in SIX}}
+ self.assertFalse(W.is_on(six, "deepseek_speedups"))
+ self.assertEqual(W.set_switch(six, "deepseek_speedups", True), env)
+ self.assertTrue(W.is_on({"DS4_DEFAULTS": "026"}, "deepseek_speedups"), "the master switch alone is not the whole switch: a hand-edited file is left alone")
+
+ def test_the_terminal_only_switch_is_for_qwen_and_never_in_the_app(self):
+ sw = W.BY_ID["qwen_defaults_026"]
+ self.assertTrue(sw.terminal_only)
+ self.assertEqual(dict(sw.env), {"STRATA_DEFAULTS": "026"})
+ self.assertNotIn("qwen_defaults_026", [s.id for s in W.app_switches()])
+ for cfg, gpus in ((QWEN, [RTX]), (QWEN36, [RTX]), ({**QWEN, "backend": "sycl"}, []), ({**QWEN, "backend": "cpu"}, [])):
+ self.assertIn("qwen_defaults_026", [s.id for s in W.applicable(cfg, gpus)])
+ self.assertNotIn("qwen_defaults_026", [s.id for s in W.applicable(DEEPSEEK, [RTX])], "DeepSeek has its own master switch (DS4_DEFAULTS)")
+ self.assertEqual(W.set_switch({"PATH": "/bin"}, sw, True), {"PATH": "/bin", "STRATA_DEFAULTS": "026"})
+ self.assertFalse(W.is_on({}, sw))
+ self.assertEqual(W.set_switch({"STRATA_DEFAULTS": "026"}, sw, False), {})
+
def test_faster_long_prompts_is_for_qwen38_flash_next_only(self):
"""KVA was measured on the 48-layer / 512-expert Flash-Next shape and the engine declines other shapes: no row for Qwen3.6-35B-A3B (arch qwen35moe) or DeepSeek."""
self.assertEqual(ids(QWEN36, [RTX]), [], "Qwen3.6-35B-A3B on CUDA: no KVA row")
@@ -130,10 +191,23 @@ class Terminal(unittest.TestCase):
def names(cfg, gpus, extra=None):
rows = P.build_rows(None, cfg["args"], None, cfg["env"], None, extra or {}, P.switch_params(W.applicable(cfg, gpus)))
return [r.name for r in rows if r.name in W.env_names()]
- self.assertEqual(names(QWEN, [RTX]), ["STRATA_KVA"])
- self.assertEqual(names({**QWEN, "backend": "hip"}, [R9700]), ["STRATA_KVA", "STRATA_PCIE_FRAC_DEFAULT"])
- self.assertEqual(names(DEEPSEEK, [R9700]), list(ENV["deepseek_speedups"]), "the six, in the contract order")
- self.assertEqual(names({**QWEN, "backend": "sycl"}, [R9700]), [])
+ # (STRATA_DEFAULTS, the Qwen "0.2.6 behaviour" row, closes every Qwen table; the DeepSeek table has DS4_DEFAULTS instead)
+ self.assertEqual(names(QWEN, [RTX]), ["STRATA_KVA", "STRATA_DEFAULTS"])
+ self.assertEqual(names(QWEN, [RTX3060]), ["STRATA_KVA", "STRATA_ROUTE_PRIOR", "STRATA_DEFAULTS"])
+ self.assertEqual(names({**QWEN, "backend": "hip"}, [R9700]), ["STRATA_KVA", "STRATA_PCIE_FRAC_DEFAULT", "STRATA_DEFAULTS"])
+ self.assertEqual(names(DEEPSEEK, [R9700]), list(ENV["deepseek_speedups"]), "the master switch and the six, in the contract order")
+ self.assertEqual(names({**QWEN, "backend": "sycl"}, [R9700]), ["STRATA_DEFAULTS"])
+ self.assertEqual(names(QWEN36, [RTX]), ["STRATA_DEFAULTS"])
+
+ def test_the_qwen_0_2_6_row_in_the_advanced_table(self):
+ rows = P.build_rows(None, QWEN["args"], None, QWEN["env"], None, {}, P.switch_params(W.applicable(QWEN, [RTX])))
+ row = next(r for r in rows if r.name == "STRATA_DEFAULTS")
+ self.assertEqual((row.default, row.yours, row.known), ("off", "", True))
+ self.assertIn("Qwen: 0.2.6 behaviour", row.description)
+ self.assertIn("off by default; set 026 to turn on", row.description)
+ cfg = derive_config(QWEN, S.LaunchSettings(extra=S.settings_from_pairs(["STRATA_DEFAULTS=026"])), bool_flags=frozenset())
+ self.assertEqual(cfg["env"], {**QWEN["env"], "STRATA_DEFAULTS": "026"})
+ self.assertEqual(derive_config(QWEN, S.LaunchSettings(), bool_flags=frozenset())["env"], QWEN["env"], "a default launch carries nothing")
def test_a_row_shows_its_default_until_it_is_changed_and_explains_itself(self):
rows = P.build_rows(None, QWEN["args"], None, QWEN["env"], None, {}, P.switch_params(W.applicable(QWEN, [RTX])))
@@ -151,7 +225,7 @@ class Terminal(unittest.TestCase):
def test_the_six_deepseek_rows_say_they_go_together(self):
text = W.env_words("DS4_COPY_PIN")
- self.assertIn("all 6 speed-up rows go together", text)
+ self.assertIn("all 7 speed-up rows go together", text)
self.assertEqual({W.env_names()[n].id for n in ENV["deepseek_speedups"]}, {"deepseek_speedups"})
@@ -159,15 +233,16 @@ class SameAsTheApp(unittest.TestCase):
@unittest.skipUnless(shutil.which("node"), "node is not installed")
def test_the_apps_table_has_the_same_names_and_words(self):
script = ("import {SPEED_SWITCHES} from './apps/host/server/native-settings.mjs';"
- "console.log(JSON.stringify(SPEED_SWITCHES.map(s=>({id:s.id,label:s.label,hint:s.hint,default:s.default,env:s.env}))))")
+ "console.log(JSON.stringify(SPEED_SWITCHES.map(s=>({id:s.id,label:s.label,hint:s.hint,default:s.default,env:s.env,legacy:s.legacy??{}}))))")
out = subprocess.run(["node", "--input-type=module", "-e", script], cwd=ROOT, capture_output=True, text=True, timeout=60, check=True).stdout
app = json.loads(out)
- self.assertEqual([a["id"] for a in app], [s.id for s in W.SWITCHES])
- for a, s in zip(app, W.SWITCHES):
+ self.assertEqual([a["id"] for a in app], [s.id for s in W.app_switches()], "the Terminal-only switch is not in the app's table")
+ for a, s in zip(app, W.app_switches()):
self.assertEqual(a["label"], s.label)
self.assertEqual(a["hint"], s.effect, s.id)
self.assertEqual(a["env"], dict(s.env), s.id)
self.assertEqual(a["default"], s.default, s.id)
+ self.assertEqual(a["legacy"], dict(s.legacy), s.id)
if __name__ == "__main__":