For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

EplbStatsMetadata

EplbStatsMetadata

class max.pipelines.lib.eplb_stats.EplbStatsMetadata(num_layers, num_moe_layers, moe_layer_indices, num_logical_experts, num_experts_per_token)

source

Bases: object

Static metadata describing the shape and semantics of a snapshot.

Parameters:

• num_layers (int)
• num_moe_layers (int)
• moe_layer_indices (tuple[int, ...])
• num_logical_experts (int)
• num_experts_per_token (int)

from_dict()

classmethod from_dict(data)

source

Constructs an instance from a dict produced by to_dict.

Tolerates legacy snapshots that only had num_moe_layers (which then meant the row count): those map to num_layers with every row treated as MoE.

Parameters:

data (dict[str, Any])

Return type:

EplbStatsMetadata

moe_layer_indices

moe_layer_indices: tuple[int, ...]

source

Histogram row indices that correspond to MoE layers (row i == layer i).

num_experts_per_token

num_experts_per_token: int

source

Top-k experts selected per token.

num_layers

num_layers: int

source

Total transformer layers profiled. Equals the histogram row count.

num_logical_experts

num_logical_experts: int

source

Number of logical experts per layer.

num_moe_layers

num_moe_layers: int

source

Count of true MoE (routing) layers. Dense layers are all-zero rows.

to_dict()

to_dict()

source

Returns a JSON-safe dict representation.

Return type:

dict[str, Any]