[BugFix] Release unused unquantized weights for fused MC2 (#15216)
### What this PR does / why we need it? When fused MC2 uses CANN MegaMoe, unquantized expert weights are cloned into per-expert `w13_weight_list` and `w2_weight_list` tensors. The original fused `w13_weight` and `w2_weight` parameters remain registered even though the forward and dynamic EPLB paths consume the cloned lists, which keeps a redundant full copy of the local expert weights on the NPU. This change: - creates split expert-weight lists for CANN MegaMoe or dynamic EPLB through one shared path; - removes the original fused parameters after the independent clones are ready; - releases cached NPU allocator blocks after removing the unused parameters. The execution tensors are unchanged: fused MC2 and dynamic EPLB continue to receive the same cloned per-expert weights. For a BF16 example with 43 MoE layers, 4 local experts per rank, hidden size 4096, and expert intermediate size 2048, the calculated persistent saving per rank is: `43 × 4 × 3 × 4096 × 2048 × 2 bytes = 8.0625 GiB` This is a calculated weight-memory saving. No end-to-end throughput improvement is claimed. ### Does this PR introduce _any_ user-facing change? No API or inference-output behavior changes. It reduces NPU memory usage for unquantized fused-MC2 deployments that materialize per-expert weight lists. ### How was this patch tested? - `git diff --check github/main...HEAD` - `python3 -m py_compile vllm_ascend/ops/fused_moe/routed_experts.py` - Reviewed the fused-MC2 and dynamic-EPLB consumers to confirm they use `w13_weight_list` and `w2_weight_list` after split-list creation. The focused pytest file was not run on the development control host because `pytest` is not installed there (`No module named pytest`). - vLLM main: https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3 Signed-off-by: Zheng Shoujian <zheng.shoujian@outlook.com>
J
Jade Zheng committed
e41435157d3a80937bb01f60efb46b28d663f2d5
Parent: 895ca07
Committed by GitHub <noreply@github.com>
on 8/29/2026, 1:52:10 AM