Trim public component surface and export build_from_gguf

Tighten the public API ahead of the first release.

components.__all__ leaked 37 model-specific classes (vision towers,
per-model attention/projector variants, codec/SSM blocks) that are
implementation details of individual models, not reusable building
blocks. Drop them from the public contract while keeping them importable
internally via redundant-alias re-exports (`X as X`), so existing
`from mobius.components import ...` call sites in model files are
unaffected. components.__all__ shrinks from 112 to 75 names.

Also export build_from_gguf at the top level: it was documented as
public but never added to mobius.__all__ or re-exported. The gguf
third-party import stays lazy, so `import mobius` remains safe without
the optional gguf package installed.

Config export policy (base-only vs export-all) is intentionally left to
the config-redesign track.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <11205048+justinchuby@users.noreply.github.com>

Author: justinchuby

src/mobius/__init__.py

@@ -34,6 +34,7 @@
     "build",
     "build_context",
     "build_diffusers_pipeline",
+    "build_from_gguf",
     "build_from_module",
     "components",
     "ep_capabilities",
@@ -87,4 +88,5 @@
     registry,
 )
 from mobius._weight_loading import apply_weights
+from mobius.integrations.gguf import build_from_gguf
 from mobius.tasks import CausalLMTask, ModelTask

src/mobius/components/__init__.py

@@ -11,8 +11,6 @@
     "CausalConvNd",
     "CausalDepthwiseConv1d",
     "CausalTransConv1d",
-    "CodecDecoderTransformerModel",
-    "CodecEncoderTransformerModel",
     "ConformerEncoder",
     "Conv1d",
     "Conv2d",
@@ -22,7 +20,6 @@
     "DecoderBlock",
     "DecoderLayer",
     "DecoderResidualUnit",
-    "DeepSeekMLA",
     "DiffusionFFN",
     "DiffusionSelfAttention",
     "Embedding",
@@ -34,59 +31,28 @@
     "GatedDeltaNet",
     "GatedMLP",
     "GatedRMSNorm",
-    "Gemma3MultiModalProjector",
-    "Gemma4AudioEncoder",
     "ClippableLinear",
     "GroupNorm",
     "GQAContext",
     "INT64_MAX",
     "InputMixer",
-    "JambaSelectiveScan",
     "LayerNorm",
     "LayerNormNoAffine",
     "LayerNormNoBias",
     "OffsetLayerNorm",
     "LayerScale",
-    "LightningAttention",
     "Linear",
     "LinearMultiModalProjector",
     "LoRALinear",
     "MLP",
-    "Mistral3MultiModalProjector",
     "MLPMultiModalProjector",
-    "Mamba2Block",
-    "MambaBlock",
     "MoELayer",
     "OffsetRMSNorm",
     "PatchEmbed",
     "PatchEmbedding",
-    "PixtralVisionTower",
     "PostGatedRMSNorm",
     "PostNormDecoderLayer",
-    "QFormer",
-    "QFormerAttention",
-    "QFormerLayer",
     "QuantizedLinear",
-    "Qwen25VLPatchEmbed",
-    "Qwen25VLPatchMerger",
-    "Qwen25VLVisionAttention",
-    "Qwen25VLVisionBlock",
-    "Qwen25VLVisionModel",
-    "Qwen25VLVisionRotaryEmbedding",
-    "Qwen2VLVisionBlock",
-    "Qwen2VLVisionModel",
-    "Qwen35Attention",
-    "Qwen3ASRAudioAttention",
-    "Qwen3ASRAudioEncoderLayer",
-    "Qwen3VLPatchEmbed",
-    "Qwen3VLPatchMerger",
-    "Qwen3VLVisionAttention",
-    "Qwen3VLVisionBlock",
-    "Qwen3VLVisionModel",
-    "Qwen3VLVisionRotaryEmbedding",
-    "SANMAttention",
-    "SANMEncoderLayer",
-    "SANMFFN",
     "RMSNorm",
     "SelectiveScan",
     "SiLU",
@@ -103,9 +69,6 @@
     "VisionEncoder",
     "VisionEncoderLayer",
     "VisionModel",
-    "WhisperAttention",
-    "WhisperDecoderLayer",
-    "WhisperEncoderLayer",
     "apply_rms_norm",
     "create_attention_bias",
     "create_decoder_layer",
@@ -120,9 +83,11 @@
 from mobius.components._attention import (
     Attention,
     GQAContext,
-    Qwen35Attention,
     StaticCacheState,
 )
+from mobius.components._attention import (
+    Qwen35Attention as Qwen35Attention,
+)
 from mobius.components._audio import ConformerEncoder
 from mobius.components._codec_conv import (
     CausalConv1d,
@@ -135,8 +100,10 @@
     SnakeBeta,
 )
 from mobius.components._codec_transformer import (
-    CodecDecoderTransformerModel,
-    CodecEncoderTransformerModel,
+    CodecDecoderTransformerModel as CodecDecoderTransformerModel,
+)
+from mobius.components._codec_transformer import (
+    CodecEncoderTransformerModel as CodecEncoderTransformerModel,
 )
 from mobius.components._codec_vq import SplitResidualVectorQuantizer
 from mobius.components._common import (
@@ -164,7 +131,7 @@
     PostNormDecoderLayer,
     create_decoder_layer,
 )
-from mobius.components._deepseek_mla import DeepSeekMLA
+from mobius.components._deepseek_mla import DeepSeekMLA as DeepSeekMLA
 from mobius.components._diffusion import (
     AdaLayerNormOutput,
     AdaLayerNormZero,
@@ -183,10 +150,12 @@
     EncoderDecoderAttention,
 )
 from mobius.components._gated_deltanet import GatedDeltaNet
-from mobius.components._gemma4_audio import ClippableLinear, Gemma4AudioEncoder
-from mobius.components._lightning_attention import LightningAttention
+from mobius.components._gemma4_audio import ClippableLinear
+from mobius.components._gemma4_audio import Gemma4AudioEncoder as Gemma4AudioEncoder
+from mobius.components._lightning_attention import LightningAttention as LightningAttention
 from mobius.components._lora import LoRALinear
-from mobius.components._mamba_block import Mamba2Block, MambaBlock
+from mobius.components._mamba_block import Mamba2Block as Mamba2Block
+from mobius.components._mamba_block import MambaBlock as MambaBlock
 from mobius.components._mlp import FCMLP, MLP, FusedGateUpMLP, GatedMLP
 from mobius.components._moe import (
     MoELayer,
@@ -196,45 +165,79 @@
     TopKGate,
 )
 from mobius.components._multimodal import (
-    Gemma3MultiModalProjector,
+    Gemma3MultiModalProjector as Gemma3MultiModalProjector,
+)
+from mobius.components._multimodal import (
     InputMixer,
     LinearMultiModalProjector,
     MLPMultiModalProjector,
 )
 from mobius.components._pixtral_vision import (
-    Mistral3MultiModalProjector,
-    PixtralVisionTower,
+    Mistral3MultiModalProjector as Mistral3MultiModalProjector,
+)
+from mobius.components._pixtral_vision import (
+    PixtralVisionTower as PixtralVisionTower,
+)
+from mobius.components._qformer import (
+    QFormer as QFormer,
 )
 from mobius.components._qformer import (
-    QFormer,
-    QFormerAttention,
-    QFormerLayer,
+    QFormerAttention as QFormerAttention,
+)
+from mobius.components._qformer import (
+    QFormerLayer as QFormerLayer,
 )
 from mobius.components._quantized_linear import (
     QuantizedLinear,
     make_quantized_linear_factory,
 )
 from mobius.components._qwen3_asr_audio import (
-    Qwen3ASRAudioAttention,
-    Qwen3ASRAudioEncoderLayer,
+    Qwen3ASRAudioAttention as Qwen3ASRAudioAttention,
+)
+from mobius.components._qwen3_asr_audio import (
+    Qwen3ASRAudioEncoderLayer as Qwen3ASRAudioEncoderLayer,
+)
+from mobius.components._qwen3_vl_vision import (
+    Qwen3VLPatchEmbed as Qwen3VLPatchEmbed,
+)
+from mobius.components._qwen3_vl_vision import (
+    Qwen3VLPatchMerger as Qwen3VLPatchMerger,
+)
+from mobius.components._qwen3_vl_vision import (
+    Qwen3VLVisionAttention as Qwen3VLVisionAttention,
+)
+from mobius.components._qwen3_vl_vision import (
+    Qwen3VLVisionBlock as Qwen3VLVisionBlock,
 )
 from mobius.components._qwen3_vl_vision import (
-    Qwen3VLPatchEmbed,
-    Qwen3VLPatchMerger,
-    Qwen3VLVisionAttention,
-    Qwen3VLVisionBlock,
-    Qwen3VLVisionModel,
-    Qwen3VLVisionRotaryEmbedding,
+    Qwen3VLVisionModel as Qwen3VLVisionModel,
+)
+from mobius.components._qwen3_vl_vision import (
+    Qwen3VLVisionRotaryEmbedding as Qwen3VLVisionRotaryEmbedding,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen2VLVisionBlock as Qwen2VLVisionBlock,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen2VLVisionModel as Qwen2VLVisionModel,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen25VLPatchEmbed as Qwen25VLPatchEmbed,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen25VLPatchMerger as Qwen25VLPatchMerger,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen25VLVisionAttention as Qwen25VLVisionAttention,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen25VLVisionBlock as Qwen25VLVisionBlock,
+)
+from mobius.components._qwen25_vl_vision import (
+    Qwen25VLVisionModel as Qwen25VLVisionModel,
 )
 from mobius.components._qwen25_vl_vision import (
-    Qwen2VLVisionBlock,
-    Qwen2VLVisionModel,
-    Qwen25VLPatchEmbed,
-    Qwen25VLPatchMerger,
-    Qwen25VLVisionAttention,
-    Qwen25VLVisionBlock,
-    Qwen25VLVisionModel,
-    Qwen25VLVisionRotaryEmbedding,
+    Qwen25VLVisionRotaryEmbedding as Qwen25VLVisionRotaryEmbedding,
 )
 from mobius.components._rms_norm import (
     GatedRMSNorm,
@@ -245,12 +248,18 @@
 )
 from mobius.components._rotary_embedding import initialize_rope
 from mobius.components._sanm_attention import (
-    SANMFFN,
-    SANMAttention,
-    SANMEncoderLayer,
+    SANMFFN as SANMFFN,
+)
+from mobius.components._sanm_attention import (
+    SANMAttention as SANMAttention,
+)
+from mobius.components._sanm_attention import (
+    SANMEncoderLayer as SANMEncoderLayer,
+)
+from mobius.components._ssm import (
+    JambaSelectiveScan as JambaSelectiveScan,
 )
 from mobius.components._ssm import (
-    JambaSelectiveScan,
     SelectiveScan,
 )
 from mobius.components._vision import (
@@ -262,7 +271,13 @@
 )
 from mobius.components._whisper import (
     Conv1d,
-    WhisperAttention,
-    WhisperDecoderLayer,
-    WhisperEncoderLayer,
+)
+from mobius.components._whisper import (
+    WhisperAttention as WhisperAttention,
+)
+from mobius.components._whisper import (
+    WhisperDecoderLayer as WhisperDecoderLayer,
+)
+from mobius.components._whisper import (
+    WhisperEncoderLayer as WhisperEncoderLayer,
 )