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-V4.1: slots on by default (2 GB cap, one week), docs
Niko1221be26cb79957e
diff --git a/docs/DEEPSEEK41.md b/docs/DEEPSEEK41.mdindex 32bd2a22c..7aefeec4c 100644--- a/docs/DEEPSEEK41.md+++ b/docs/DEEPSEEK41.md@@ -70,6 +70,11 @@ The server speaks to the engine as it does for the other models, so `/v1/chat/coserver turns them into OpenAI `tool_calls` and Anthropic `tool_use`, including `tool_choice`, streamed arguments and calls written inside the thinking.A value marked `string="true"` stays a string.- Defaults when a request names none: temperature 1.0, top_p 1.0.+- Agents and sub-agents that take turns do not read their prompts again. When another request takes the model, the engine sets the+ conversation aside (about 17 MB per 6.5K tokens) and takes it back when the next prompt continues it; the system prompt and tools shared by+ an agent and its sub-agents are kept as a checkpoint too. Slots survive a restart, live in `ds4-slots` next to the run log, are capped at+ 2 GB and are removed after a week. `STRATA_DS4_SLOT_KEEP=0` turns it off, `STRATA_DS4_SLOT_GB` changes the cap, `STRATA_DS4_AUTO_PIN=0`+ turns off the shared-start checkpoint. The answers are the same as a fresh read of the prompt.Pictures and the DSpark drafter are in the engine (strata-dsv41 reads them from the unsloth mmproj and dspark files beside the model; the server side is not wired yet). Not in this version: several requests at once (`--parallel` is ignored with a note), tuning(`--tune` is skipped with a note).diff --git a/serve/server.py b/serve/server.pyindex 55a371967..fdfd59495 100644--- a/serve/server.py+++ b/serve/server.py@@ -3477,16 +3477,18 @@ def hip_auto_devices(cfg: dict, env: dict) -> list[int] | None:return visible-DS41_SLOT_KEEP_DEFAULT = "0" # the engine keeps a displaced conversation in RAM and in the slot folder (see below)+DS41_SLOT_KEEP_DEFAULT = "1" # the engine keeps a displaced conversation in RAM and in the slot folder (see below)+DS41_SLOT_GB_DEFAULT = "2" # the folder's cap in GB (a 6.5K-token conversation is about 17 MB)def deepseek_slot_env(env: dict, cfg: dict) -> None:"""DeepSeek-V4.1 only. The engine sets a conversation aside when another request takes the model and takes it backat its checkpoints (agents and sub-agents that take turns do not read their prompts again). STRATA_DS4_SLOT_KEEPswitches it (the default above; a config's env or the shell wins); the folder is `ds4-slots` next to the run log, in- Strata's data folder (STRATA_DS4_SLOT_DIR names another), and the engine caps it (STRATA_DS4_SLOT_GB, 16) and- removes what is older than a week or not this model's."""+ Strata's data folder (STRATA_DS4_SLOT_DIR names another), and the engine caps it (STRATA_DS4_SLOT_GB, 2 here, 16 in the engine) and+ removes what is older than a week or not this model's. On by default; STRATA_DS4_SLOT_KEEP=0 turns it off."""env.setdefault("STRATA_DS4_SLOT_KEEP", DS41_SLOT_KEEP_DEFAULT)+ env.setdefault("STRATA_DS4_SLOT_GB", DS41_SLOT_GB_DEFAULT)log = cfg.get("log")if "STRATA_DS4_SLOT_DIR" not in env and log:env["STRATA_DS4_SLOT_DIR"] = str(Path(log).parent / "ds4-slots")diff --git a/serve/test_dsv41.py b/serve/test_dsv41.pyindex b928e8162..15315ee59 100644--- a/serve/test_dsv41.py+++ b/serve/test_dsv41.py@@ -592,7 +592,8 @@ class SlotEnv(unittest.TestCase):def test_deepseek_gets_a_folder_in_the_data_folder_and_the_default_switch(self):env = self.env({"arch": "deepseek41", "log": str(Path("data") / "strata-deepseek.log"), "args": []})self.assertEqual(Path(env["STRATA_DS4_SLOT_DIR"]), Path("data") / "ds4-slots")- self.assertEqual(env["STRATA_DS4_SLOT_KEEP"], server.DS41_SLOT_KEEP_DEFAULT)+ self.assertEqual(env["STRATA_DS4_SLOT_KEEP"], "1")+ self.assertEqual(env["STRATA_DS4_SLOT_GB"], "2")def test_the_configs_own_settings_win(self):env = self.env({"arch": "deepseek41", "log": "x/y.log",