thinkingmachines/Inkling
Original947.1Bparameters
Fixture33.8Mparameters
Layer stacks2compared
Findings0to review
✓ Same architecture, smaller
Every module class, layer kind and checked property of the original is in the fixture.
Layer stacks
One square per layer, colored by kind: the layer's module classes and type attributes. Hover a square for its index and kind.
model.language_model.layers
Original · 66
Fixture · 6
| Kind | Original | Fixture | |
|---|---|---|---|
| hybrid_sliding, MLP | 2 | 2 | ✓ kept |
| hybrid_sliding, Experts, MoE … | 53 | 3 | ✓ kept |
| hybrid, Experts, MoE … | 11 | 1 | ✓ kept |
model.vision_tower.encoder_layers
Original · 4
Fixture · 2
| Kind | Original | Fixture | |
|---|---|---|---|
| RMSNorm | 3 | 1 | ✓ kept |
| no RMSNorm | 1 | 1 | ✓ kept |
Where the parameters are
Share of each model's parameters per component. Vocabulary and patch sizes are kept, so embeddings usually take a larger share of the fixture.
Original
Fixture
| Component | Original | Share | Fixture | Share |
|---|---|---|---|---|
| model.language_model.layers | 944.6B | 99.7% | 7.2M | 21.4% |
| model.language_model | 1.2B | 0.1% | 12.9M | 38.1% |
| lm_head | 1.2B | 0.1% | 12.9M | 38.1% |
| model.vision_tower.encoder_layers | 83.7M | 0.0% | 751.7K | 2.2% |
| model.audio_tower | 7.9M | 0.0% | 82.0K | 0.2% |
| model.vision_tower | 6.1K | 0.0% | 64 | 0.0% |
Preserved properties
| Config | Property | Original | Fixture | |
|---|---|---|---|---|
| model | model_type | inkling_mm_model | inkling_mm_model | ✓ kept |
| text_config | attention style | GQA (64 heads, 8 KV) | GQA (4 heads, 2 KV) | ✓ kept |
| text_config | MoE routing | top-6 of 256 experts | top-2 of 4 experts | ✓ kept |
| text_config | model_type | inkling_text | inkling_text | ✓ kept |
| text_config | vocab_size | 201024 | 201024 | ✓ kept |
| text_config | hidden_act | silu | silu | ✓ kept |
| text_config | max_position_embeddings | 1048576 | 1048576 | ✓ kept |
| audio_config | model_type | inkling_audio | inkling_audio | ✓ kept |
| vision_config | model_type | inkling_vision | inkling_vision | ✓ kept |
| vision_config | patch_size | 40 | 40 | ✓ kept |
| vision_config | num_channels | 3 | 3 | ✓ kept |
Module classes
Instances of each class in the built model (prefix Inkling omitted).
| Class | Original | Fixture |
|---|---|---|
| Linear | 341 | 39 |
| RMSNorm | 271 | 29 |
| Conv1d | 264 | 24 |
| ShortConvolution | 264 | 24 |
| SiLUActivation | 130 | 10 |
| Attention | 66 | 6 |
| DecoderLayer | 66 | 6 |
| RelativeLogits | 66 | 6 |
| Experts | 64 | 4 |
| MoE | 64 | 4 |
| SharedExperts | 64 | 4 |
| TopkRouter | 64 | 4 |
| VisionEncoderLayer | 4 | 2 |
| MLP | 2 | 2 |
| ModuleList | 2 | 2 |
| Embedding | 1 | 1 |
| AudioModel | 1 | 1 |
| AudioModelEmbeddings | 1 | 1 |
| ForConditionalGeneration | 1 | 1 |
| Model | 1 | 1 |
| NormedEmbedding | 1 | 1 |
| TextModel | 1 | 1 |
| VisionModel | 1 | 1 |
Config changes
16 fields changed, 102 unchanged.
| Field | Original | Fixture |
|---|---|---|
| audio_config.text_hidden_size | 6144 | 64 |
| text_config.head_dim | 128 | 16 |
| text_config.hidden_size | 6144 | 64 |
| text_config.intermediate_size | 24576 | 128 |
| text_config.layer_types | 66 entries: 55× hybrid_sliding, 11× hybrid | ["hybrid_sliding", "hybrid_sliding", "hybrid_sliding", "hybrid_sliding", "hyb… |
| text_config.mlp_layer_types | 66 entries: 64× sparse, 2× dense | ["dense", "dense", "sparse", "sparse", "sparse", "sparse"] |
| text_config.moe_intermediate_size | 3072 | 32 |
| text_config.n_routed_experts | 256 | 4 |
| text_config.num_attention_heads | 64 | 4 |
| text_config.num_experts_per_tok | 6 | 2 |
| text_config.num_hidden_layers | 66 | 6 |
| text_config.num_key_value_heads | 8 | 2 |
| vision_config.hidden_size | 1024 | 64 |
| vision_config.n_layers | 4 | 2 |
| vision_config.num_attention_heads | 16 | 4 |
| vision_config.text_hidden_size | 6144 | 64 |