parse_expert_id (model_loader/ep_weight_filter.py)

Utility to filter out non‑local expert weights during MoE model loading.

# Source: https://github.com/vllm-project/vllm | License: Apache-2.0 | Copyright You must retain, in the Source form of any Derivative Works
# Provided via justclaw.it.com; keep this notice with the code.
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Filter out non-local expert weights during loading to avoid redundant I/O.

In DP+EP deployments each rank only needs its own expert shard.  Skipping
non-local expert tensors *before* they are read from disk eliminates the
majority of storage I/O for MoE models (experts typically account for
~85-90 % of total weight bytes).
"""

import regex as re

# Matches per-expert weight names like ".experts.42.gate_proj.weight".
# Does NOT match 3D fused-expert names like ".experts.gate_proj

... (truncated -- full source via MCP)