Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

unsloth start dsh: warn when the loaded context leaves replies no room (#14352)

Daniel Han64814142defe
diff --git a/unsloth_cli/commands/start.py b/unsloth_cli/commands/start.py
index 9a27aa37e1d..c6a0c4c8084 100644
--- a/unsloth_cli/commands/start.py
+++ b/unsloth_cli/commands/start.py
@@ -82,6 +82,8 @@ _DSH_PACKAGE = "@deepseek-ai/dsh"
# dsh picks its sandbox+approval preset from DSH_PERMISSION_MODE via ??, so omitting it would inherit a danger-full-access exported in the parent shell, and "" is not unset to ??. Pin the mode in both directions instead of only setting it for --yolo.
_DSH_SAFE_PERMISSION_MODE = "workspace-write"
_DSH_YOLO_PERMISSION_MODE = "danger-full-access"
+# dsh's pi-ai provider sends max_tokens = contextWindow - its prompt estimate - 4,096 (pi-ai CONTEXT_SAFETY_TOKENS), floored at 1, and dsh's own instructions and tools are about 6,500 tokens: at 8,192 every reply is cut after one token (#14315).
+_DSH_MIN_CONTEXT = 16384
_PI_PROVIDER = "unsloth"
_SUBAGENT_NAME = "unsloth"
_SUBAGENT_DESCRIPTION = (
@@ -7576,6 +7578,16 @@ def dsh(
launch = launch,
server_options = server_options,
)
+ window = entry.get("context_length") or entry.get("max_context_length")
+ if window and int(window) < _DSH_MIN_CONTEXT:
+ typer.echo(
+ f"Warning: {entry['id']} is loaded with {int(window):,} tokens of context, too few "
+ "for DeepSeek Harness: its instructions and tools take about 6,500 and it keeps "
+ "4,096 in reserve, so replies are cut short (at 8,192, after a single token). "
+ f"Load it with --max-seq-length {_DSH_MIN_CONTEXT} or more (a smaller quant if "
+ "that does not fit in memory) and re-run unsloth start dsh.",
+ err = True,
+ )
with _session_config("dsh", launch, persist = persist) as home:
patch = home / _DSH_PATCH_FILE
write_dsh_patch(base, entry, patch, server_options.request_body())
diff --git a/unsloth_cli/tests/test_start.py b/unsloth_cli/tests/test_start.py
index 3b94077c913..517676a363e 100644
--- a/unsloth_cli/tests/test_start.py
+++ b/unsloth_cli/tests/test_start.py
@@ -7532,6 +7532,16 @@ def test_connect_dsh_no_launch(fake_studio, tmp_path):
assert not (home / "settings.yaml").exists()
[email protected]("window, warned", [(8_192, True), (12_288, True), (16_384, False)])
+def test_dsh_warns_when_the_context_leaves_replies_no_room(fake_studio, window, warned):
+ # #14315: at 8,192 dsh sends max_tokens 1, so every reply is "The" and nothing else.
+ fake_studio.state["models"] = [{**MODEL, "context_length": window}]
+ result = CliRunner().invoke(start.start_app, ["dsh", "--no-launch"])
+ assert result.exit_code == 0, result.output
+ assert (f"loaded with {window:,} tokens of context, too few" in result.output) is warned
+ assert ("--max-seq-length 16384" in result.output) is warned
+
+
def _vibe_settings(env: dict) -> tuple:
return json.loads(env["VIBE_PROVIDERS"]), json.loads(env["VIBE_MODELS"])