vllm-project/vllm

[Bug]: ModelOpt Llama-4 Checkpoints Take 5+ minutes to load

Open

#31,624 opened on 2026年1月2日

GitHub で見る
 (12 comments) (0 reactions) (0 assignees)Python (80,034 stars) (16,816 forks)batch import
bugfeature requestgood first issuehelp wanted

説明

🚀 The feature, motivation and pitch

In working on some MoE refactors, I discovered that L4 for ModelOpt takes 5+minutes to load weights even from CPU page cache.

The root cause is basically this hack logic to load the state dict that ModelOpt uses

What happens is that the CPU tensor (loaded weight) that we are going to load into the GPU tensor (param) becomes non-contiguous due to this logic. As a result, when we eventually call _copy() from CPU->GPU we are calling this on a non-contiguous cpu tensor which takes 3-4s per weight.

To hack around this for local R&D, I simply immediately move the loaded_weight to the GPU. This makes the gather happen on the GPU which accelerates things a lot. This isn't reasonable as an actual solution though

We should investigate where the logic in the weight loader can avoid creating non-contiguous CPU tensors

Alternatives

No response

Additional context

No response

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

コントリビューターガイド

技術スタック
pythonpytorch
領域
backendperformance
Issue 種別
bug
難度新しい貢献者にとっての実装難度です。1 は非常に小さな変更、5 は専門的な作業です。
3
推定時間経験ある貢献者が調査、実装、テスト、pull request 準備にかけるおおよその時間範囲です。
1-2 days
活動状況Issue が今どれだけ取り組みやすいかを示します。新しい、活発、古い、ブロック中、またはメンテナー入力待ちなどです。
active
明確さ期待される変更、受け入れ条件、次の手順がどれだけ明確かを示します。
mostly clear
前提条件
PyTorch tensor operationsvLLM model loadingcontiguous memory
初心者向け度初回貢献者にどれだけ取り組みやすいかを 1-100 で推定したスコアです。
30
調査方針
Investigate the weight loading logic in vllm/model executor/models/llama4.py lines 439-523 (ModelOpt fused case). Focus on why the CPU tensor becomes non contiguous after the custom state dict loading. Propose a way to ensure tensors remain contiguous before the CPU >GPU copy, e.g., by calling `.contiguous()` on the loaded weight tensor. Alternatively, consider modifying the gather logic to avoid creating non contiguous slices. The goal is to avoid the 3-4 second delay per weight when copying non contiguous tensors.