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: stop listing Bonsai DSpark drafter files as models (#11265)

kevin902ef70dffc2
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index a76e5a24660..a46f2576196 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -3227,7 +3227,8 @@ _IMATRIX_TOKEN_RE = re.compile(r"^imatrix(?:[._\-]|$)|[._\-]imatrix$", re.IGNORE
def _is_companion_gguf_path(path: str) -> bool:
"""True for a non-main GGUF: vision mmproj, a calibration imatrix, or a separate
drafter (repo-root ``mtp-*.gguf``, the ``MTP/`` subdir copies for Gemma 4, the ``dspark/``
- drafters for DeepSeek V4 Flash, the ``eagle3-*.gguf`` draft heads). Mirrors
+ drafters for DeepSeek V4 Flash, the ``<model>-dspark-<quant>.gguf`` drafters for
+ prism-ml Bonsai, the ``eagle3-*.gguf`` draft heads). Mirrors
hub.utils.gguf so variant resolution never picks a companion as the main model --
a Gemma ``Q8_0`` request must not resolve to ``MTP/...-Q8_0-MTP.gguf``, which sorts
ahead of the real weight.
@@ -3243,6 +3244,8 @@ def _is_companion_gguf_path(path: str) -> bool:
# holding calibration statistics, so only the name rules it out.
if _IMATRIX_TOKEN_RE.search(Path(p).stem):
return True
+ if _DSPARK_BEFORE_QUANT_RE.search(Path(p).name):
+ return True
return _drafter_path_kind(path) is not None
@@ -3311,6 +3314,11 @@ _GGUF_KNOWN_QUANT_RE = re.compile(
r"|BF16|F16|F32)",
re.IGNORECASE,
)
+# Mirrors hub.utils.gguf._DSPARK_BEFORE_QUANT_RE, for _is_companion_gguf_path.
+_DSPARK_BEFORE_QUANT_RE = re.compile(
+ rf"-dspark-(?:{_GGUF_KNOWN_QUANT_RE.pattern})(?:-[0-9]{{5}}-of-[0-9]{{5}})?\.gguf$",
+ re.IGNORECASE,
+)
def _is_big_endian_gguf_path(path: str, variant_key: str = "") -> bool:
diff --git a/studio/backend/hub/utils/gguf.py b/studio/backend/hub/utils/gguf.py
index 7b775925a14..1cca0f91a20 100644
--- a/studio/backend/hub/utils/gguf.py
+++ b/studio/backend/hub/utils/gguf.py
@@ -108,16 +108,24 @@ _DRAFTER_KINDS = ("mtp", "dspark", "dflash", "eagle3")
# Directories only: mtp/ and dspark/ are always a publisher's companion folder, while dflash/ is a family name a user picks for real weights.
_DRAFTER_DIR_KINDS = ("mtp", "dspark")
+# <model>-dspark-<quant>.gguf drafters (prism-ml Bonsai, third-party DSpark GGUFs); a name going on past the quant (DeepSeek-V4-Flash-Dspark-Abliterated-MXFP4-BF16.gguf) is the model.
+_DSPARK_BEFORE_QUANT_RE = re.compile(
+ rf"-dspark-(?:{_GGUF_QUANT_RE.pattern})(?:-[0-9]{{5}}-of-[0-9]{{5}})?\.gguf$",
+ re.IGNORECASE,
+)
+
def is_mtp_drafter_path(path: str) -> bool:
- """True for a separate-file drafter, a companion to the main model rather than a selectable quant: the repo-root ``mtp-*.gguf`` (the Q8_0 copy unsloth ships for llama.cpp ``-hf`` auto-discovery), the ``MTP/`` subdir copies (Gemma 4), the ``dspark/`` drafters (DeepSeek V4 Flash) and the ``eagle3-*.gguf`` draft heads (ggml-org gpt-oss). Repos that bake the head into the main GGUF (Qwen) have no such file, so this is False for them. Must be excluded from main-model selection everywhere mmproj is. Matched by basename prefix, or by an exact parent dir for ``_DRAFTER_DIR_KINDS``; never a substring, since the kind names double as family names, so ``Qwen3.6-27B-MTP-Q4_K_M.gguf`` and ``Qwen3.6-35B-A3B-DFlash-Q4_K_M.gguf`` ARE the model. CANONICAL COPY. Two mirrors must change in lockstep: utils/models/model_config.py ``_is_mtp_drafter`` (utils cannot import hub) and core/inference/llama_cpp.py ``_is_companion_gguf_path`` (core avoids hub imports; bundles the mmproj check)."""
+ """True for a separate-file drafter, a companion to the main model rather than a selectable quant: the repo-root ``mtp-*.gguf`` (the Q8_0 copy unsloth ships for llama.cpp ``-hf`` auto-discovery), the ``MTP/`` subdir copies (Gemma 4), the ``dspark/`` drafters (DeepSeek V4 Flash), the ``<model>-dspark-<quant>.gguf`` drafters (prism-ml Bonsai) and the ``eagle3-*.gguf`` draft heads (ggml-org gpt-oss). Repos that bake the head into the main GGUF (Qwen) have no such file, so this is False for them. Must be excluded from main-model selection everywhere mmproj is. Matched by basename prefix, by an exact parent dir for ``_DRAFTER_DIR_KINDS``, or by ``_DSPARK_BEFORE_QUANT_RE``; never a bare substring, since the kind names double as family names, so ``Qwen3.6-27B-MTP-Q4_K_M.gguf`` and ``Qwen3.6-35B-A3B-DFlash-Q4_K_M.gguf`` ARE the model. CANONICAL COPY. Two mirrors must change in lockstep: utils/models/model_config.py ``_is_mtp_drafter`` (utils cannot import hub) and core/inference/llama_cpp.py ``_is_companion_gguf_path`` (core avoids hub imports; bundles the mmproj check)."""
p = path.replace("\\", "/").lower()
if not p.endswith(".gguf"):
return False
parts = [segment for segment in p.split("/") if segment]
name, parents = parts[-1], parts[:-1]
- return any(name.startswith(f"{kind}-") for kind in _DRAFTER_KINDS) or any(
- kind in parents for kind in _DRAFTER_DIR_KINDS
+ return (
+ any(name.startswith(f"{kind}-") for kind in _DRAFTER_KINDS)
+ or any(kind in parents for kind in _DRAFTER_DIR_KINDS)
+ or _DSPARK_BEFORE_QUANT_RE.search(name) is not None
)
@@ -128,8 +136,11 @@ def is_reclaimable_drafter_path(path: str) -> bool:
return False
parts = [segment for segment in p.split("/") if segment]
name, parents = parts[-1], parts[:-1]
- return any(name.startswith(f"{kind}-") for kind in _DRAFTER_DIR_KINDS) or any(
- kind in parents for kind in _DRAFTER_DIR_KINDS
+ return (
+ any(name.startswith(f"{kind}-") for kind in _DRAFTER_DIR_KINDS)
+ or any(kind in parents for kind in _DRAFTER_DIR_KINDS)
+ # <model>-dspark-<quant>.gguf: hidden from the menu as a companion, so it goes with the last variant too.
+ or _DSPARK_BEFORE_QUANT_RE.search(name) is not None
)
diff --git a/studio/backend/tests/test_mtp_drafter_companion.py b/studio/backend/tests/test_mtp_drafter_companion.py
index f52b2ea8b7d..b1ec5a23eee 100644
--- a/studio/backend/tests/test_mtp_drafter_companion.py
+++ b/studio/backend/tests/test_mtp_drafter_companion.py
@@ -91,6 +91,12 @@ DRAFTER_CASES = [
("quants/eagle3-gpt-oss-20b-Q8_0.gguf", True),
("Llama-3.1-8B-Eagle3-Q4_K_M.gguf", False),
("eagle3/Llama-3.1-8B-Eagle3-Q4_K_M.gguf", False),
+ # <model>-dspark-<quant>.gguf drafters; a name going on past the kind is the model.
+ ("Ternary-Bonsai-27B-dspark-Q4_1.gguf", True),
+ ("Bonsai-27B-dspark-bf16.gguf", True),
+ ("Qwen3.6-27B-DSpark-Q4_K_M.gguf", True),
+ ("model-dspark-Q8_0-00001-of-00002.gguf", True),
+ ("DeepSeek-V4-Flash-Dspark-Abliterated-MXFP4-BF16-00001-of-00004.gguf", False),
]
@@ -180,6 +186,24 @@ def test_eagle3_draft_head_is_not_a_variant_or_the_default():
assert plans["mxfp4"].target_filenames == ("gpt-oss-20b-MXFP4.gguf",)
+BONSAI_FILES = [
+ "Bonsai-27B-F16.gguf",
+ "Bonsai-27B-Q1_0.gguf",
+ "Bonsai-27B-dspark-Q4_1.gguf",
+ "Bonsai-27B-dspark-bf16.gguf",
+ "Bonsai-27B-mmproj-BF16.gguf",
+]
+
+
+def test_bonsai_dspark_drafter_is_not_a_variant():
+ # #11259: the drafter was the smallest row, and llama-server cannot load it as a model.
+ plans = build_gguf_variant_plans(
+ [_sib(name, 1_000, f"sha-{i}") for i, name in enumerate(BONSAI_FILES)]
+ )
+ assert set(plans) == {"f16", "q1_0"}
+ assert plans["q1_0"].main_filenames == frozenset({"Bonsai-27B-Q1_0.gguf"})
+
+
def test_baked_in_repo_plans_unchanged():
plans = build_gguf_variant_plans([_sib("Qwen3.6-27B-MTP-Q4_K_M.gguf", 4_000, "q4")])
assert plans["q4_k_m"].target_filenames == ("Qwen3.6-27B-MTP-Q4_K_M.gguf",)
@@ -2790,6 +2814,8 @@ def test_dflash_stays_unreclaimable_even_though_auto_now_launches_it(tmp_path):
assert is_reclaimable_drafter_path("dflash-kquant.gguf") is False
assert is_reclaimable_drafter_path("dspark-model-Q8_0.gguf") is True
+ assert is_reclaimable_drafter_path("Ternary-Bonsai-27B-dspark-Q4_1.gguf") is True
+ assert is_reclaimable_drafter_path("DeepSeek-V4-Flash-Dspark-Abliterated-MXFP4.gguf") is False
repo, snap = _cache_repo(
tmp_path,
@@ -2803,6 +2829,22 @@ def test_dflash_stays_unreclaimable_even_though_auto_now_launches_it(tmp_path):
assert not (snap / "dspark-model-Q8_0.gguf").is_symlink()
+def test_deleting_the_last_bonsai_variant_reclaims_its_dspark_drafter(tmp_path):
+ """The menu hides <model>-dspark-<quant>.gguf as a companion, so once the last real
+ variant is deleted nothing could ever launch it; its blob must go too."""
+ from hub.services.models.deletion import _delete_gguf_variant_from_repos
+
+ repo, snap = _cache_repo(
+ tmp_path,
+ "prism-ml/Bonsai-27B-gguf",
+ ["Bonsai-27B-Q1_0.gguf", "Bonsai-27B-dspark-Q4_1.gguf"],
+ )
+ _delete_gguf_variant_from_repos("prism-ml/Bonsai-27B-gguf", "Q1_0", [repo], None, root = tmp_path)
+
+ assert not (snap / "Bonsai-27B-Q1_0.gguf").is_symlink()
+ assert not (snap / "Bonsai-27B-dspark-Q4_1.gguf").is_symlink()
+
+
def test_detect_dflash_file_skips_a_sidecar_named_for_another_weight(tmp_path):
"""A multi-model folder must not attach a foreign drafter.
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index e32805f3cfe..62158f11957 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -1901,7 +1901,8 @@ _DRAFTER_DIR_KINDS = ("mtp", "dspark")
def _is_mtp_drafter(path: str) -> bool:
"""True for a separate-file drafter, a companion to the main model rather
than a selectable quant: the repo-root ``mtp-*.gguf``, the ``MTP/`` subdir
- copies (Gemma 4), the ``dspark/`` drafters (DeepSeek V4 Flash) or the
+ copies (Gemma 4), the ``dspark/`` drafters (DeepSeek V4 Flash), the
+ ``<model>-dspark-<quant>.gguf`` drafters (prism-ml Bonsai) or the
``eagle3-*.gguf`` draft heads (ggml-org gpt-oss).
Mirrors hub.utils.gguf.is_mtp_drafter_path (utils cannot import hub). Must be
@@ -1910,18 +1911,20 @@ def _is_mtp_drafter(path: str) -> bool:
resolve to ``MTP/...-Q8_0-MTP.gguf``, which sorts ahead of the real weight,
or to ``dspark/dspark-...-Q8_0.gguf``.
- Prefix, or an exact directory for ``_DRAFTER_DIR_KINDS``; never a substring,
- since the kind names double as family names
- (``Qwen3.6-35B-A3B-DFlash-Q4_K_M.gguf`` IS the model, as is anything in a
- user's ``dflash/`` folder).
+ Prefix, an exact directory for ``_DRAFTER_DIR_KINDS``, or
+ ``_DSPARK_BEFORE_QUANT_RE``; never a bare substring, since the kind names
+ double as family names (``Qwen3.6-35B-A3B-DFlash-Q4_K_M.gguf`` IS the model,
+ as is anything in a user's ``dflash/`` folder).
"""
p = path.replace("\\", "/").lower()
if not p.endswith(".gguf"):
return False
parts = [segment for segment in p.split("/") if segment]
name, parents = parts[-1], parts[:-1]
- return any(name.startswith(f"{kind}-") for kind in _DRAFTER_KINDS) or any(
- kind in parents for kind in _DRAFTER_DIR_KINDS
+ return (
+ any(name.startswith(f"{kind}-") for kind in _DRAFTER_KINDS)
+ or any(kind in parents for kind in _DRAFTER_DIR_KINDS)
+ or _DSPARK_BEFORE_QUANT_RE.search(name) is not None
)
@@ -2806,6 +2809,11 @@ _GGUF_KNOWN_QUANT_RE = re.compile(
r"|BF16|F16|F32)",
re.IGNORECASE,
)
+# Mirrors hub.utils.gguf._DSPARK_BEFORE_QUANT_RE, for _is_mtp_drafter.
+_DSPARK_BEFORE_QUANT_RE = re.compile(
+ rf"-dspark-(?:{_GGUF_KNOWN_QUANT_RE.pattern})(?:-[0-9]{{5}}-of-[0-9]{{5}})?\.gguf$",
+ re.IGNORECASE,
+)
_FLOAT_PRECISION_QUANTS = frozenset({"BF16", "F16", "F32"})