qwen2_model_descriptor

Qwen2 model descriptor for AnyModel compression.

Classes

Qwen2ModelDescriptor	Model descriptor for Qwen2 models.
Qwen2FFNIntermediateLayerDescriptor	Layer descriptor for Qwen2 FFN intermediate pruning.

class Qwen2FFNIntermediateLayerDescriptor

Bases: LlamaFFNIntermediateLayerDescriptor

Layer descriptor for Qwen2 FFN intermediate pruning.

Qwen2 uses the same FFN structure as Llama (gate_proj, up_proj, down_proj).

__init__(down_proj_name='mlp.down_proj', ffn_prefix_name='model.layers.{layer_idx}.mlp', linear_weight_names=<factory>)

Parameters:

• down_proj_name (str)

• ffn_prefix_name (str)

• linear_weight_names (List[str])

Return type:

None

class Qwen2ModelDescriptor

Bases: ModelDescriptor

Model descriptor for Qwen2 models.

static attn_no_op_post_init(decoder_layer)

Parameters:

decoder_layer (Module)

static block_config_to_layer_overrides(block_config)

Parameters:

block_config (BlockConfig)

classmethod create_dummy_block(original_layer, block_index)

Create a dummy block that preserves Qwen2-specific attributes like attention_type.

Qwen2’s forward pass accesses decoder_layer.attention_type for attention mask selection.

Parameters:

• original_layer (Module)

• block_index (int)

Return type:

Module

static decoder_layer_cls()

static final_norm_name()

static init_rotary_embedding(model, runtime)

Parameters:

model (Module)

static input_embedding_name()

static layer_block_name(index)

Parameters:

index (int)

static layer_name_predicates(num_layers)

Parameters:

num_layers (int)

Return type:

Dict[str, Pattern]

static mlp_no_op_post_init(decoder_layer)

Parameters:

decoder_layer (Module)

static output_embedding_name()