Compare commits

..

No commits in common. "489af21fd2615777461ec2556fa9360fc0e1a070" and "e2297f43e46ccce22e94c9cdc3f631ccec2729fa" have entirely different histories.

11 changed files with 15 additions and 119 deletions

View file

@ -7,16 +7,6 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
## [Unreleased]
### Added
- Per-profile chat streaming and chat-request timeouts: `stream`, `read_timeout`
and `connect_timeout` are now resolvable in `[default]`/`[model.<profile>]`
(CLI `--stream`/`--read-timeout`/`--connect-timeout` still override).
### Changed
- The chat request now honors the configured `(connect, read)` timeout instead
of a hard-coded 30 s; the default `read_timeout` is 600 s so slow reasoning
models are no longer cut off mid-generation.
## [0.1.0] - 2026-07-07
Initial release. Replaces the previous collection of shell scripts

View file

@ -13,7 +13,7 @@ Referenzmodell (zunächst): das Default-Modell aus `llama.cpp.config`
|---|---|
| `system_prompt_prosa.md` | `prosa_01_werkstatt.md`, `prosa_02_dystopie.md`, `prosa_03_essay.md`, `prosa_04_dialog.md` |
| `system_prompt_reden.md` | `reden_01_abifeier.md`, `reden_02_windkraft.md`, `reden_03_umstrukturierung.md`, `reden_04_gedenken.md` |
| `system_prompt_coding.md` | `coding_01_jsonl_validator.md`, `coding_02_refactor.md`, `coding_03_ts_retry.md`, `coding_04_performance.md`, `coding_05_rng_uniform.md` |
| `system_prompt_coding.md` | `coding_01_jsonl_validator.md`, `coding_02_refactor.md`, `coding_03_ts_retry.md`, `coding_04_performance.md` |
Jeder User-Prompt fordert gezielt Eigenschaften heraus, die der jeweilige
System-Prompt verspricht (z. B. „show, don't tell" und Schlussbild bei Prosa;
@ -37,20 +37,7 @@ Hinweise:
(`--max-tokens`) muss **Denk- plus Ausgabe-Tokens** abdecken — für Prosa/Reden
großzügig wählen (z. B. 600010000), sonst bricht die sichtbare Antwort ab
oder bleibt leer.
- **`--stream` benutzen.** Beim Streaming setzt jeder Token den Read-Timeout
zurück; ohne Streaming muss `read_timeout` die gesamte Generierung abdecken.
Chat-`read_timeout`/`stream` sind pro Profil in der Config setzbar (Default
`read_timeout = 600`), `--read-timeout`/`--stream` auf der CLI überschreiben.
## Coding-Prompts mit dem `ornith35b`-Profil
Die `coding_*`-Prompts lassen sich direkt gegen das Coding-Profil testen. Es
bringt deterministischeres Sampling mit und aktiviert `stream` + hohen
`read_timeout` bereits in der Config, es sind also keine Extra-Flags nötig:
```bash
llamacppctl --change --profile ornith35b # Coding-Modell laden
llamacppctl --chat --profile ornith35b \
--system-prompt-profile coding \
--prompt-file example_user_prompts/coding_05_rng_uniform.md
```
- **`--stream` benutzen.** Ohne Streaming wartet das Tool die komplette Antwort
in einem einzigen Read ab und läuft beim langen Reasoning-Output in den
Default-`--read-timeout` (30 s). Beim Streaming setzt jeder Token den Timeout
zurück. Für lange nicht-gestreamte Läufe zusätzlich `--read-timeout 600`.

View file

@ -1,10 +0,0 @@
Implementiere in Python eine Funktion `random_unit_float() -> float`, die eine möglichst „perfekte" gleichverteilte Zufallszahl im halboffenen Intervall [0.0, 1.0) liefert.
Anforderungen:
- Verwende eine kryptografisch starke Entropiequelle aus der Standardbibliothek (kein `random.random()`), z. B. `secrets`/`os.urandom`.
- Nutze die volle 53-Bit-Mantisse eines `float64` und erzeuge die Zahl bias-frei (kein Modulo-Bias, keine ungleichmäßige Rundung); begründe die Wahl von genau 53 Bit.
- Halte das Intervall sauber halboffen: 0.0 ist möglich, 1.0 darf niemals herauskommen.
- Type Hints, ein knapper Docstring und keine unnötigen Abhängigkeiten.
- pytest-Tests, die abdecken: Wertebereich [0.0, 1.0) über viele Ziehungen, dass 1.0 nie auftritt, und ein einfacher Verteilungs-Check (z. B. Bucket-/Mittelwert-Test) mit begründeter Toleranz.
Erkläre im Anschluss kurz die wichtigsten Designentscheidungen (warum 53 Bit, warum die gewählte Quelle, wie 1.0 ausgeschlossen wird).

View file

@ -62,14 +62,6 @@ api_key =
max_tokens = 2048
# chat_temperature leer lassen -> die Server-Temperatur (temp) gilt.
chat_temperature =
# Chat-Antwort standardmäßig token-weise streamen (wie --stream). Für langsame
# Reasoning-Modelle angenehm, weil sofort Ausgabe erscheint.
stream = false
# (connect, read)-Timeout des Chat-Requests in Sekunden. read_timeout muss die
# volle Generierungszeit abdecken; Reasoning-Modelle brauchen viel -> hoch
# setzen. --read-timeout/--connect-timeout auf der CLI überschreiben das.
read_timeout = 600
connect_timeout = 10
# Example second model profile, pinned to the second GPU (e.g. RTX 3090 #2)
# with its own port and container name so it can run alongside [default].

View file

@ -233,8 +233,7 @@ def do_chat(cfg: ServerConfig, prompt_cfg: PromptConfig, args) -> int:
print("--chat requires a user prompt", file=sys.stderr)
return EXIT_GENERAL
# Streaming can be requested via --stream OR enabled per profile in config.
if getattr(args, "stream", False) or getattr(prompt_cfg, "stream", False):
if getattr(args, "stream", False):
return _do_chat_stream(cfg, prompt_cfg, args)
try:

View file

@ -103,11 +103,8 @@ def build_parser() -> argparse.ArgumentParser:
p.add_argument("--no-allow-symlinks", dest="allow_symlinks", action="store_false")
p.add_argument("--max-input-bytes", type=int, default=1_048_576)
p.add_argument("--url-allow-host", action="append", default=[])
# Default None so a config-provided value can win; a fallback is applied at
# the point of use (URL fetch: 3s/10s; chat request: profile read_timeout).
# When set, these also override the chat request's (connect, read) timeout.
p.add_argument("--connect-timeout", type=float, default=None)
p.add_argument("--read-timeout", type=float, default=None)
p.add_argument("--connect-timeout", type=float, default=3.0)
p.add_argument("--read-timeout", type=float, default=10.0)
# --- Chat request parameters (for --chat and the --start reply) ---
p.add_argument(
@ -169,9 +166,7 @@ def validate_args(args: argparse.Namespace, parser: argparse.ArgumentParser) ->
if args.max_tokens is not None and args.max_tokens <= 0:
parser.error("--max-tokens must be > 0")
if (args.connect_timeout is not None and args.connect_timeout <= 0) or (
args.read_timeout is not None and args.read_timeout <= 0
):
if args.connect_timeout <= 0 or args.read_timeout <= 0:
parser.error("timeouts must be > 0")
if args.allow_private_url and args.url_allow_host:

View file

@ -248,29 +248,11 @@ def resolve_effective_config(args) -> tuple:
if getattr(args, "chat_temp", None) is not None:
chat_temp = args.chat_temp
# Streaming: enabled per profile via `stream = true`, or with --stream.
stream = str(merged.get("stream", "")).strip().lower() in TRUE_STRINGS
if getattr(args, "stream", False):
stream = True
# Chat request timeouts: config value first, CLI flag overrides. These are
# the (connect, read) timeouts of the chat POST; read_timeout must cover the
# model's full generation time (raise it for slow reasoning models).
read_timeout = float(merged.get("read_timeout") or 600.0)
if getattr(args, "read_timeout", None) is not None:
read_timeout = args.read_timeout
connect_timeout = float(merged.get("connect_timeout") or 10.0)
if getattr(args, "connect_timeout", None) is not None:
connect_timeout = args.connect_timeout
prompt_cfg = PromptConfig(
system_prompt=system_prompt,
user_prompt=getattr(args, "_resolved_user_prompt", None),
max_tokens=max_tokens,
temperature=chat_temp,
stream=stream,
connect_timeout=connect_timeout,
read_timeout=read_timeout,
)
return server_cfg, prompt_cfg

View file

@ -88,21 +88,18 @@ def chat_completion(cfg: ServerConfig, prompt_cfg: PromptConfig) -> requests.Res
def chat_completion_text(
cfg: ServerConfig, prompt_cfg: PromptConfig, timeout: Optional[tuple] = None
cfg: ServerConfig, prompt_cfg: PromptConfig, timeout: float = 30.0
) -> ChatReply:
"""High-level chat helper for --chat and the --start/--check post-checks.
Returns a ChatReply (content + finish_reason), raising HttpError on
transport, HTTP status, JSON decoding, or malformed-response failures.
The (connect, read) timeout defaults to the prompt profile's configured
values so slow reasoning models are not cut off mid-generation."""
transport, HTTP status, JSON decoding, or malformed-response failures."""
try:
resp = requests.post(
chat_url(cfg),
json=_chat_payload(cfg, prompt_cfg),
headers=_auth_headers(cfg),
timeout=timeout or (prompt_cfg.connect_timeout, prompt_cfg.read_timeout),
timeout=timeout,
)
except requests.RequestException as exc:
raise HttpError(f"chat completion request failed: {exc}") from exc
@ -140,7 +137,7 @@ def stream_chat(
json=payload,
headers=_auth_headers(cfg),
stream=True,
timeout=timeout or (prompt_cfg.connect_timeout, prompt_cfg.read_timeout),
timeout=timeout or (10.0, 600.0),
)
except requests.RequestException as exc:
raise HttpError(f"chat completion request failed: {exc}") from exc

View file

@ -24,8 +24,8 @@ from .prompt_io import InputPolicy, PromptSourceError, load_prompt_source, resol
def _build_input_policy(args) -> InputPolicy:
return InputPolicy(
max_input_bytes=args.max_input_bytes,
connect_timeout=args.connect_timeout if args.connect_timeout is not None else 3.0,
read_timeout=args.read_timeout if args.read_timeout is not None else 10.0,
connect_timeout=args.connect_timeout,
read_timeout=args.read_timeout,
allow_insecure_http=args.allow_insecure_http,
allow_private_url=args.allow_private_url,
allow_ip_host=args.allow_ip_host,

View file

@ -73,12 +73,6 @@ class PromptConfig:
# None => omit from the request so the server's configured --temp applies.
temperature: Optional[float] = None
stream: bool = False
# HTTP timeouts for the chat request itself (seconds). read_timeout must
# cover the model's full generation time: reasoning models can "think" for
# a long time before the first/last token, so the default is generous. Both
# are configurable per profile (read_timeout/connect_timeout in the config).
connect_timeout: float = 10.0
read_timeout: float = 600.0
@dataclass

View file

@ -258,33 +258,3 @@ def test_chat_params_cli_overrides_config(tmp_path):
_, prompt_cfg = resolve_effective_config(args)
assert prompt_cfg.max_tokens == 512
assert prompt_cfg.temperature == 0.1
def test_stream_and_timeouts_default(tmp_path):
cfg_path = _write_config(tmp_path)
_, prompt_cfg = resolve_effective_config(_parse(["--start", "--config", str(cfg_path)]))
assert prompt_cfg.stream is False
assert prompt_cfg.read_timeout == 600.0 # generous default for reasoning models
assert prompt_cfg.connect_timeout == 10.0
def test_stream_and_read_timeout_from_config(tmp_path):
cfg = CONFIG_BASIC.replace(
"poll_interval = 2",
"poll_interval = 2\nstream = true\nread_timeout = 900\nconnect_timeout = 5",
)
cfg_path = _write_config(tmp_path, cfg)
_, prompt_cfg = resolve_effective_config(_parse(["--start", "--config", str(cfg_path)]))
assert prompt_cfg.stream is True
assert prompt_cfg.read_timeout == 900.0
assert prompt_cfg.connect_timeout == 5.0
def test_stream_and_read_timeout_cli_overrides_config(tmp_path):
cfg = CONFIG_BASIC.replace("poll_interval = 2", "poll_interval = 2\nread_timeout = 900")
cfg_path = _write_config(tmp_path, cfg)
# --stream forces streaming on; --read-timeout overrides the config value.
args = _parse(["--chat", "-p", "hi", "--config", str(cfg_path), "--stream", "--read-timeout", "42"])
_, prompt_cfg = resolve_effective_config(args)
assert prompt_cfg.stream is True
assert prompt_cfg.read_timeout == 42.0