RedHatAI/Llama-4-Scout-17B-16E-Instruct
Original108.6Bparameters
Fixture44.7Mparameters
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.
vision_model.model.layers
Original · 34
Fixture · 2
| Kind | Original | Fixture | |
|---|---|---|---|
| VisionEncoderLayer | 34 | 2 | ✓ kept |
language_model.model.layers
Original · 48
Fixture · 4
| Kind | Original | Fixture | |
|---|---|---|---|
| TextL2Norm | 36 | 3 | ✓ kept |
| no TextL2Norm | 12 | 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 |
|---|---|---|---|---|
| language_model.model.layers | 105.7B | 97.3% | 542.2K | 1.2% |
| language_model.model | 1.0B | 1.0% | 12.9M | 28.9% |
| language_model.lm_head | 1.0B | 1.0% | 12.9M | 28.9% |
| vision_model.model.layers | 809.5M | 0.7% | 100.0K | 0.2% |
| vision_model.vision_adapter | 39.8M | 0.0% | 17.8M | 39.9% |
| multi_modal_projector.linear_1 | 21.0M | 0.0% | 262.1K | 0.6% |
| vision_model.patch_embedding | 827.9K | 0.0% | 37.6K | 0.1% |
| 3 others | 819.5K | 0.0% | 37.2K | 0.1% |
Preserved properties
| Config | Property | Original | Fixture | |
|---|---|---|---|---|
| model | model_type | llama4 | llama4 | ✓ kept |
| model | tie_word_embeddings | False | False | ✓ kept |
| vision_config | RoPE type | default | default | ✓ kept |
| vision_config | model_type | llama4_vision_model | llama4_vision_model | ✓ kept |
| vision_config | hidden_act | gelu | gelu | ✓ kept |
| vision_config | image_size | 336 | 336 | ✓ kept |
| vision_config | patch_size | 14 | 14 | ✓ kept |
| vision_config | num_channels | 3 | 3 | ✓ kept |
| text_config | attention style | GQA (40 heads, 8 KV) | GQA (4 heads, 2 KV) | ✓ kept |
| text_config | MoE routing | top-1 of 16 experts | top-1 of 4 experts | ✓ kept |
| text_config | RoPE type | llama3 | llama3 | ✓ kept |
| text_config | model_type | llama4_text | llama4_text | ✓ kept |
| text_config | vocab_size | 202048 | 202048 | ✓ kept |
| text_config | tie_word_embeddings | False | False | ✓ kept |
| text_config | hidden_act | silu | silu | ✓ kept |
| text_config | max_position_embeddings | 10485760 | 10485760 | ✓ kept |
Module classes
Instances of each class in the built model (prefix Llama4 omitted).
| Class | Original | Fixture |
|---|---|---|
| Linear | 545 | 45 |
| TextRMSNorm | 97 | 9 |
| SiLUActivation | 96 | 8 |
| LayerNorm | 70 | 6 |
| Router | 48 | 4 |
| TextAttention | 48 | 4 |
| TextDecoderLayer | 48 | 4 |
| TextExperts | 48 | 4 |
| TextMLP | 48 | 4 |
| TextMoe | 48 | 4 |
| TextL2Norm | 36 | 3 |
| GELU | 35 | 3 |
| VisionAttention | 34 | 2 |
| VisionEncoderLayer | 34 | 2 |
| VisionMLP | 34 | 2 |
| ModuleList | 2 | 2 |
| Embedding | 1 | 1 |
| ForCausalLM | 1 | 1 |
| ForConditionalGeneration | 1 | 1 |
| MultiModalProjector | 1 | 1 |
| TextModel | 1 | 1 |
| TextRotaryEmbedding | 1 | 1 |
| UnfoldConvolution | 1 | 1 |
| VisionEncoder | 1 | 1 |
| VisionMLP2 | 1 | 1 |
| VisionModel | 1 | 1 |
| VisionPixelShuffleMLP | 1 | 1 |
| VisionRotaryEmbedding | 1 | 1 |
| Unfold | 1 | 1 |
Config changes
14 fields changed, 75 unchanged.
| Field | Original | Fixture |
|---|---|---|
| text_config.head_dim | 128 | 16 |
| text_config.hidden_size | 5120 | 64 |
| text_config.intermediate_size | 8192 | 128 |
| text_config.layer_types | 48 entries: 36× chunked_attention, 12× full_attention | ["chunked_attention", "chunked_attention", "chunked_attention", "full_attenti… |
| text_config.moe_layers | 48 entries | [0, 1, 2, 3] |
| text_config.no_rope_layers | 48 entries: 36× 1, 12× 0 | [1, 1, 1, 0] |
| text_config.num_attention_heads | 40 | 4 |
| text_config.num_hidden_layers | 48 | 4 |
| text_config.num_key_value_heads | 8 | 2 |
| text_config.num_local_experts | 16 | 4 |
| vision_config.hidden_size | 1408 | 64 |
| vision_config.intermediate_size | 5632 | 256 |
| vision_config.num_attention_heads | 16 | 4 |
| vision_config.num_hidden_layers | 34 | 2 |