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.

feat(studio): price GLM-5.3-Flash oQ quants as unsloth-zoo loads them (#16619)

Long Yixing6938f3256ca0
diff --git a/studio/backend/core/inference/mlx_memory.py b/studio/backend/core/inference/mlx_memory.py
index 38561062582..15061a0e183 100644
--- a/studio/backend/core/inference/mlx_memory.py
+++ b/studio/backend/core/inference/mlx_memory.py
@@ -1364,12 +1364,17 @@ def _loader_view(model_dir: str, config: dict) -> tuple:
def _install_oq_renames(config: dict) -> None:
- """unsloth_zoo's sanitize renames for DeepSeek-V4.1 oQ quants, as its load installs them."""
+ """unsloth_zoo's sanitize renames for oQ quants, as its load installs them."""
try:
- from unsloth_zoo.mlx.loader import _ensure_deepseek_v41_omlx_layout
+ from unsloth_zoo.mlx import loader
except Exception:
return
- _ensure_deepseek_v41_omlx_layout(config.get("model_type"))
+ # Older unsloth_zoo installs its one oQ layout (DeepSeek-V4.1) by that name.
+ install = getattr(loader, "_ensure_omlx_layout", None) or getattr(
+ loader, "_ensure_deepseek_v41_omlx_layout", None
+ )
+ if install is not None:
+ install(config.get("model_type"))
@functools.lru_cache(maxsize = 4)
diff --git a/studio/backend/tests/test_memory_estimate.py b/studio/backend/tests/test_memory_estimate.py
index b6b75cdf631..cf53239383e 100644
--- a/studio/backend/tests/test_memory_estimate.py
+++ b/studio/backend/tests/test_memory_estimate.py
@@ -4157,6 +4157,105 @@ def test_a_deepseek_v41_oq_quant_is_priced_as_unsloth_zoo_loads_it(tmp_path, mon
assert not os.path.exists(view)
[email protected]("name", ["_ensure_omlx_layout", "_ensure_deepseek_v41_omlx_layout"])
+def test_unsloth_zoo_installs_its_oq_layout_by_either_name(monkeypatch, name):
+ loader = pytest.importorskip("unsloth_zoo.mlx.loader")
+ installed = []
+ for other in ("_ensure_omlx_layout", "_ensure_deepseek_v41_omlx_layout"):
+ monkeypatch.delattr(loader, other, raising = False)
+ monkeypatch.setattr(loader, name, installed.append, raising = False)
+ mm._install_oq_renames({"model_type": "glm5_next"})
+ assert installed == ["glm5_next"]
+
+
+def test_a_glm5_next_oq_quant_is_priced_as_unsloth_zoo_loads_it(tmp_path, monkeypatch):
+ if not _HAVE_MLX:
+ pytest.skip("needs MLX to build and load a checkpoint")
+ glm = pytest.importorskip("mlx_vlm.models.glm5_next.glm5_next")
+ loader = pytest.importorskip("unsloth_zoo.mlx.loader")
+ if "glm5_next" not in getattr(loader, "_OMLX_LAYOUTS", {}):
+ pytest.skip("this unsloth_zoo loads no GLM-5.3 oQ quant")
+ import re
+
+ import mlx.core as mx
+ import mlx.nn as nn
+ from mlx.utils import tree_flatten
+ from mlx_vlm.models.glm5_next.config import ModelConfig
+
+ linear = {
+ "num_heads": 2,
+ "head_dim": 64,
+ "short_conv_kernel_size": 4,
+ "kda_layers": [0],
+ "full_attn_layers": [1],
+ }
+ text = dict(
+ model_type = "glm5_next_text",
+ vocab_size = 512,
+ hidden_size = 256,
+ num_hidden_layers = 2,
+ intermediate_size = 256,
+ moe_intermediate_size = 128,
+ n_routed_experts = 16,
+ num_experts_per_tok = 2,
+ first_k_dense_replace = 1,
+ num_attention_heads = 2,
+ num_key_value_heads = 2,
+ q_lora_rank = 128,
+ kv_lora_rank = 128,
+ qk_head_dim = 64,
+ qk_nope_head_dim = 64,
+ qk_rope_head_dim = 0,
+ v_head_dim = 64,
+ index_n_heads = 2,
+ index_head_dim = 64,
+ linear_attn_config = linear,
+ layer_types = ["linear_attention", "deepseek_sparse_attention"],
+ mlp_layer_types = ["dense", "sparse"],
+ )
+ vision = dict(depth = 1, hidden_size = 64, intermediate_size = 128, num_heads = 1, out_hidden_size = 256)
+ config = {"model_type": "glm5_next", "text_config": text, "vision_config": vision}
+ model = glm.Model(ModelConfig.from_dict(json.loads(json.dumps(config))))
+ nn.quantize(
+ model,
+ 64,
+ 4,
+ class_predicate = lambda path, module: hasattr(module, "to_quantized")
+ and model.quant_predicate(path, module),
+ )
+ config["quantization"] = {"group_size": 64, "bits": 4, "mode": "affine"}
+ params = dict(tree_flatten(model.parameters()))
+
+ # The layout omlx writes: its forget gate module, the fused convolution as `conv1d`, `vision_model`.
+ def omlx(key):
+ key = key.replace("vision_tower.", "vision_model.").replace(".qkv_conv.conv.", ".conv1d.")
+ return re.sub(r"\.self_attn\.(f_b_proj|A_log|dt_bias)", r".self_attn.forget_gate.\1", key)
+
+ fbg = "language_model.model.layers.0.self_attn.fbg_a_proj."
+ stored = {omlx(k): v for k, v in params.items() if not k.startswith(fbg)}
+ for leaf in ("weight", "scales", "biases"):
+ for name, part in zip(
+ ("forget_gate.f_a_proj", "b_proj", "g_a_proj"),
+ mx.split(params[fbg + leaf], [64, 66], axis = 0),
+ ):
+ stored[f"language_model.model.layers.0.self_attn.{name}.{leaf}"] = part
+ stored["language_model.mtp.0.eh_proj.weight"] = mx.zeros((256, 512))
+ mx.save_safetensors(str(tmp_path / "model.safetensors"), stored)
+ (tmp_path / "config.json").write_text(json.dumps(config))
+
+ model_dir = str(tmp_path)
+ priced = mm._snapshot_config(model_dir)
+ sanitize = glm.Model.sanitize
+ if getattr(sanitize, "_unsloth_omlx_layout", False):
+ pytest.skip("an earlier load in this process installed the oQ layout")
+ monkeypatch.setattr(glm.Model, "sanitize", sanitize)
+ assert mm._load_is_refused(model_dir, priced, load_in_4bit = False) is None
+ glm.Model.sanitize = sanitize
+ experts = sum(v.nbytes for k, v in params.items() if ".switch_mlp." in k)
+ found = mm._offloadable(model_dir, priced)
+ assert (found["expert_bytes"], found["experts"]) == (experts, None)
+
+
try:
import importlib