Chroma 1 HD crashes with cuDNN segfault on RTX 3090 (gRPCServerCLI Docker)

[drockthedoc]

Environment

• GPU: NVIDIA GeForce RTX 3090 (24GB VRAM)
• Driver Version: 580.82.09
• CUDA Version: 12.4.1 (from container)
• Container: drawthingsai/draw-things-grpc-server-cli:latest
• Host OS: Linux
• Command: gRPCServerCLI /grpc-models --no-flash-attention

Issue Description

The gRPCServerCLI crashes with a segmentation fault when attempting to generate images using the Chroma 1 HD model. FLUX models work perfectly on the same setup.

Model Files

chroma_1_hd_r0.1_q5p.ckpt (6.1GB)
t5_xxl_encoder_q6p.ckpt (352KB)
t5_xxl_encoder_q6p.ckpt-tensordata (3.7GB)
flux_1_vae_f16.ckpt (161MB)
clip_vit_l14_f16.ckpt (236MB)

Steps to Reproduce

1. Start gRPCServerCLI with --no-flash-attention flag
2. Connect from Draw Things iOS app
3. Attempt to generate an image using Chroma 1 HD model (28 steps)
4. Server crashes with SIGSEGV

Expected Behavior

Chroma 1 HD should generate images successfully like FLUX models do.

Actual Behavior

Server crashes with null pointer dereference in libcudnn_graph.so.9.1.0

Crash Log

2025-10-05T19:14:55+0000 info com.draw-things.image-generation-service : [GRPCServer] Received request for files exist: [“chroma_1_hd_r0.1_q5p.ckpt”, “t5_xxl_encoder_q6p.ckpt”, “flux_1_vae_f16.ckpt”, “clip_vit_l14_f16.ckpt”]
2025-10-05T19:14:55+0000 info com.draw-things.image-generation-service : [GRPCServer] Received image processing request, begin.
2025-10-05T19:14:55+0000 info com.draw-things.image-generation-service : [GRPCServer] Received image processing request with configuration steps: 28

*** Signal 11: Backtracing from 0x7fc71724c1aa… done ***

*** Program crashed: Bad pointer dereference at 0x0000000000000000 ***

Thread 6 “NIO-ELT-0-#0” crashed:

0      0x00007fc71724c1aa <unknown> in [libcudnn_graph.so](http://libcudnn_graph.so).9.1.0
1 [ra] 0x00007fc71724e01c <unknown> in [libcudnn_graph.so](http://libcudnn_graph.so).9.1.0

Working Comparison

FLUX dev model works perfectly on the same setup:

2025-10-05T19:12:39+0000 info com.draw-things.image-generation-service : [GRPCServer] Received request for files exist: [“flux_1_dev_q5p.ckpt”, “t5_xxl_encoder_q6p.ckpt”, “flux_1_vae_f16.ckpt”, “clip_vit_l14_f16.ckpt”]
2025-10-05T19:12:39+0000 info com.draw-things.image-generation-service : [GRPCServer] Received image processing request, begin.
2025-10-05T19:12:39+0000 info com.draw-things.image-generation-service : [GRPCServer] Received image processing request with configuration steps: 20
2025-10-05T19:14:24+0000 info com.draw-things.image-generation-service : [GRPCServer] Image processed
2025-10-05T19:14:24+0000 info com.draw-things.image-generation-service : [GRPCServer] Image processed successfully

Additional Notes

• FlashAttention was disabled due to RTX 3090 being sm86 (Ampere)
• GPU UUID is correctly pinned in Docker configuration
• RTX 3090 has 24GB VRAM available with no other processes using the GPU
• This appears to be specific to Chroma's CUDA/cuDNN implementation

Docker Compose Configuration

version: "3.9"
services:
drawthings:
image: drawthingsai/draw-things-grpc-server-cli:latest
container_name: drawthings
restart: unless-stopped
ports:
- "7859:7859"
environment:
- NVIDIA_VISIBLE_DEVICES=GPU-*******
- NVIDIA_DRIVER_CAPABILITIES=compute,utility
volumes:
- /srv/drawthings/models:/grpc-models:Z
command: ["gRPCServerCLI", "/grpc-models", "--no-flash-attention"]
deploy:
resources:
reservations:
devices:
- driver: nvidia
capabilities: ["gpu"]
count: 1