[Bugfix][MoE] Filter packed expert weights during EP loading

[aoshen02]

Summary

• Treat per-expert .weight_packed tensors as heavy weights eligible for EP-local filtering.
• Keep scale and metadata tensors from every expert unchanged.
• Cover local packed weights, remote packed weights, and remote scale tensors in the existing unit suite.

Root cause

The EP filter previously recognized only names ending in .weight. Quantized expert checkpoints can store the primary expert payload in a packed representation whose serialized key ends in .weight_packed. These keys are produced when quantized expert weights are packed for storage or loading; the suffix describes the payload layout, not a different ownership rule.

The filter runs on checkpoint tensor names before the loader materializes or unpacks the tensor. If .weight_packed is not recognized as a heavy expert weight, non-local expert payloads bypass EP filtering and are redundantly read on every EP rank. Scale and metadata tensors intentionally remain unfiltered because they are small and may be required for quantized-kernel setup across experts.

Changes

• Recognize both .weight and .weight_packed as filterable heavy expert payloads.
• Continue filtering by the numeric expert ID and the rank's local_expert_ids.
• Leave scale, metadata, and other auxiliary tensors unchanged.
• Add focused tests for local and remote packed payloads and remote scale tensors.

Scope and compatibility

This change applies only to per-expert tensor names with a numeric expert ID. Fused expert layouts and tensors without an explicit numeric expert ID remain unchanged and continue through their existing loader paths.

Related upstream work

vLLM #48891 fixes a separate multithreaded safetensors path that previously did not receive local_expert_ids at all. This PR is complementary: it fixes recognition of .weight_packed payloads once EP filtering is enabled.

Validation

• Pure-function packed EP filter checks: passed.
• uvx ruff check vllm/model_executor/model_loader/ep_weight_filter.py tests/model_executor/model_loader/test_ep_weight_filter.py: passed.
• uvx ruff format --check vllm/model_executor/model_loader/ep_weight_filter.py tests/model_executor/model_loader/test_ep_weight_filter.py: passed.
• git diff --check: passed.
• Focused pytest was attempted, but collection on the host was blocked before reaching these tests because the local environment lacks the compiled vLLM CUDA flash-attention extension.

The available loading experiments measured the filtering path only; later post-load failures were unrelated to this change. No end-to-end accuracy claim is made.

AI assistance

AI assistance was used to investigate the root cause, prepare this minimal patch, and run the checks above. The human submitter will review every changed line and validate the runtime behavior before merge.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.