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.

Studio: avoid duplicate llama-server option families (#9621)

Apozee1540d009acf
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index ce7744ddc51..6bd4ad80445 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -93,6 +93,8 @@ from hub.utils.hf_tokens import (
)
from core.inference.llama_server_args import (
_CACHE_RAM_FLAGS,
+ _CACHE_TYPE_K_FLAGS,
+ _CACHE_TYPE_V_FLAGS,
_CTX_CHECKPOINTS_FLAGS,
_DEVICE_FLAGS,
_GPU_LAYER_FLAGS,
@@ -27950,16 +27952,20 @@ class LlamaCppBackend:
# takes a value -- older builds take -fa as a bare boolean and
# read a following "on" as a stray positional.
if _caps.get("supports_flash_attn", True):
- cmd.append("--flash-attn")
- if _caps.get("flash_attn_takes_value", True):
- cmd.append("on")
+ if not _extra_args_set_any_flag(extra_args, {"-fa", "--flash-attn"}):
+ cmd.append("--flash-attn")
+ if _caps.get("flash_attn_takes_value", True):
+ cmd.append("on")
else:
# Only ever reached on an authoritative --help that has no
# flash attention to lose, so there is no speed to give up.
_flash_attn_known_off = True
# Error out at n_ctx instead of silently rotating the KV cache; frontend catches it and points the user at "Context Length".
if _caps.get("supports_no_context_shift", True):
- cmd.append("--no-context-shift")
+ if not _extra_args_set_any_flag(
+ extra_args, {"--context-shift", "--no-context-shift"}
+ ):
+ cmd.append("--no-context-shift")
# A positive context is always passed (in auto-fit, --fit then
# optimizes the gpu-layer offload around it). When auto-fit has
# no explicit context, omit -c so --fit sizes it to fit VRAM:
@@ -28141,7 +28147,8 @@ class LlamaCppBackend:
# Expose Prometheus /metrics for the engine-stats logger, only
# when the binary advertises it (older/custom binaries may not).
if server_caps.get("supports_metrics"):
- cmd.append("--metrics")
+ if not _extra_args_set_any_flag(extra_args, {"--metrics", "--no-metrics"}):
+ cmd.append("--metrics")
self._slot_save_dir = None
self._slot_save_binary = None
self._prompt_cache_disabled = False
@@ -28220,7 +28227,8 @@ class LlamaCppBackend:
# Enable Jinja chat template rendering
if _caps.get("supports_jinja", True):
- cmd.extend(["--jinja"])
+ if not _extra_args_set_any_flag(extra_args, {"--jinja", "--no-jinja"}):
+ cmd.extend(["--jinja"])
cmd.extend(_trace_verbosity_args(_caps, extra_args, os.environ))
# KV cache data type
@@ -28235,14 +28243,10 @@ class LlamaCppBackend:
and cache_type_kv in _valid_cache_types
and not _cache_type_from_env
):
- cmd.extend(
- [
- "--cache-type-k",
- cache_type_kv,
- "--cache-type-v",
- cache_type_kv,
- ]
- )
+ if not _extra_args_set_any_flag(extra_args, _CACHE_TYPE_K_FLAGS):
+ cmd.extend(["--cache-type-k", cache_type_kv])
+ if not _extra_args_set_any_flag(extra_args, _CACHE_TYPE_V_FLAGS):
+ cmd.extend(["--cache-type-v", cache_type_kv])
self._cache_type_kv = cache_type_kv
logger.info(f"KV cache type: {cache_type_kv}")
else:
diff --git a/studio/backend/tests/test_llama_extra_args_end_to_end.py b/studio/backend/tests/test_llama_extra_args_end_to_end.py
index 36c7c195add..9e4234be9b6 100644
--- a/studio/backend/tests/test_llama_extra_args_end_to_end.py
+++ b/studio/backend/tests/test_llama_extra_args_end_to_end.py
@@ -99,6 +99,87 @@ def test_a_multi_token_value_stays_one_argv_entry(tmp_path):
assert cmd[cmd.index("--chat-template") + 1] == template
+ "extra_args,family",
+ [
+ pytest.param(["--cache-type-k", "q4_0"], {"-ctk", "--cache-type-k"}, id = "cache-k"),
+ pytest.param(["-ctk", "q4_0"], {"-ctk", "--cache-type-k"}, id = "cache-k-short"),
+ pytest.param(["--cache-type-v", "q4_0"], {"-ctv", "--cache-type-v"}, id = "cache-v"),
+ pytest.param(["-ctv", "q4_0"], {"-ctv", "--cache-type-v"}, id = "cache-v-short"),
+ pytest.param(["--flash-attn", "off"], {"-fa", "--flash-attn"}, id = "flash-off"),
+ pytest.param(["-fa", "off"], {"-fa", "--flash-attn"}, id = "flash-short"),
+ pytest.param(
+ ["--context-shift"],
+ {"--context-shift", "--no-context-shift"},
+ id = "context-shift",
+ ),
+ pytest.param(
+ ["--no-context-shift"],
+ {"--context-shift", "--no-context-shift"},
+ id = "no-context-shift",
+ ),
+ pytest.param(["--jinja"], {"--jinja", "--no-jinja"}, id = "jinja"),
+ pytest.param(["--no-jinja"], {"--jinja", "--no-jinja"}, id = "no-jinja"),
+ ],
+)
+def test_an_explicit_semantic_override_suppresses_the_automatic_atom(tmp_path, extra_args, family):
+ cmd = _cmd(tmp_path, cache_type_kv = "q8_0", extra_args = extra_args)
+
+ assert cmd[-len(extra_args) :] == extra_args
+ assert family.isdisjoint(cmd[: -len(extra_args)])
+
+
[email protected]("extra_arg", ["--metrics", "--no-metrics"])
+def test_a_supported_metrics_override_suppresses_the_automatic_flag(tmp_path, extra_arg):
+ backend, gguf = _backend(tmp_path, vulkan = False, memory = [(0, 20_000, 24_000)])
+ caps = backend.probe_server_capabilities()
+ caps["supports_metrics"] = True
+ backend.probe_server_capabilities = lambda _binary = None: caps
+
+ cmd = _launch(backend, gguf, extra_args = [extra_arg])["cmd"]
+
+ assert cmd[-1] == extra_arg
+ assert not {"--metrics", "--no-metrics"}.intersection(cmd[:-1])
+
+
+ "extra_args,other_axis",
+ [
+ pytest.param(["-ctk", "q4_0"], "--cache-type-v", id = "k-keeps-v"),
+ pytest.param(["-ctv", "q4_0"], "--cache-type-k", id = "v-keeps-k"),
+ ],
+)
+def test_cache_axis_overrides_are_independent(tmp_path, extra_args, other_axis):
+ cmd = _cmd(tmp_path, cache_type_kv = "q8_0", extra_args = extra_args)
+
+ assert other_axis in cmd[: -len(extra_args)]
+ assert cmd[-len(extra_args) :] == extra_args
+
+
+def test_a_supported_flash_override_is_not_treated_as_a_missing_flag(tmp_path):
+ extra_args = ["-fa", "on"]
+ cmd = _cmd(tmp_path, cache_type_kv = "q8_0", extra_args = extra_args)
+
+ cache_v = cmd.index("--cache-type-v")
+ assert cmd[cache_v + 1] == "q8_0"
+ assert cmd[-len(extra_args) :] == extra_args
+
+
+def test_unknown_and_repeated_extra_args_remain_verbatim_and_last(tmp_path):
+ extra_args = [
+ "--future-flag",
+ "alpha",
+ "--metrics",
+ "--future-flag",
+ "omega",
+ "--no-metrics",
+ ]
+ cmd = _cmd(tmp_path, extra_args = extra_args)
+
+ assert cmd[-len(extra_args) :] == extra_args
+ assert not {"--metrics", "--no-metrics"}.intersection(cmd[: -len(extra_args)])
+
+
@pytest.mark.parametrize(
"denied",
[