lllyasviel / lllyasviel/FramePack

RTX 5070 (Blackwell) Compatibility: Severe Slowdown/Freeze due to DynamicSwap I/O Bottleneck

Open
#783 24 comments 3 reactions 0 assignees View on GitHub
Dominant language
Python
Stars
17.3k
Forks
1.7k
PR merge metrics
No merged PRs in 30d

Description

Hello,

I am experiencing a critical performance issue when attempting to generate videos using FramePack on the new NVIDIA RTX 5070 (Blackwell) GPU. The issue appears to be related to model swapping/loading overhead rather than GPU calculation speed.

## 1. Environment Details

* **GPU:** NVIDIA GeForce RTX 5070 (12 GB VRAM)
* **CPU:** AMD Ryzen 7 5700X
* **RAM:** 64 GB
* **OS:** Windows 11
* **FramePack Setup:** Custom install with Python 3.12, CUDA 12.8, PyTorch nightly build (cu128), and latest `sageattention`.

## 2. Issue Description: High Usage / Low Power Freeze

Video generation does complete, but it is extremely slow due to the constant, lengthy waiting periods during model loading.

The key observation is a significant contradiction in GPU telemetry during the wait periods:

* **GPU Usage:** Consistently **99%**
* **GPU Power Draw:** Extremely low, only **~49 Watts**
* **GPU Temperature:** Very low, **~50°C**
* **Clock Speed:** High (**~2880 MHz**)

This indicates the GPU is **stuck waiting** for data, not performing heavy calculation, suggesting an I/O or memory access bottleneck caused by the DynamicSwap feature.

## 3. Specific Bottleneck Log

The process frequently freezes/stalls at the model loading and unloading points, specifically with the `Autoencoder` and `Transformer` models.

**Example Log of Stall:**

Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete. ... (Waiting for seconds/minutes) Loaded AutoencoderKLHunyuanVideo to cuda:0 as complete. ... (Processing starts: ~6.55 s/it) ... (Processing finishes, then stall occurs) Moving DynamicSwap_HunyuanVideoTransformer3DModelPacked to cuda:0 with preserved memory: 6 GB

## 4. Current Performance

After a full clean reinstall, the actual calculation speed is good (~6.55 s/it), but the overall generation time is poor due to the prolonged loading/offloading stalls.

We suspect this is a compatibility issue with the DynamicSwap mechanism and the RTX 5000 series (Blackwell) architecture, possibly related to how the custom `sageattention` or memory allocation handles the new hardware.

Thank you for your attention to this critical issue.

Contributor guide

No contributing guide indexed for this repository

Research direction

Start with the DynamicSwap load and unload path for the Autoencoder and HunyuanVideo Transformer, then reproduce the stalls on the listed RTX 5070 environment while checking the provided telemetry and log sequence. Done means identifying and resolving the prolonged model-swapping waits without regressing the reported processing speed.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.