comfyui_therock_gpu_test.py --generate-cat derives checkpoint dir from 'folder:' instead of 'models path:' → checkpoint never found
Nobody has claimed this yet.
- Dominant language
- Rust
- Stars
- 40
- Forks
- 9
- Avg merge
- 4d 20h
- Merged PRs (30d)
- 59
Description
Current behavior
scripts/comfyui_therock_gpu_test.py --generate-cat puts (or looks for) the test checkpoint in the wrong directory, so ComfyUI never lists it and the image-generation step fails:
- After
rocm comfyui install(default flow) it downloads the 2 GB checkpoint to~/.rocm/envs/<runtime>/apps/comfyui/models/checkpoints/and then fails with
RuntimeError: ComfyUI does not list checkpoint 'sd-v1-5-tiny.safetensors'. ... Available checkpoints: [] - With
--skip-installit looks in~/.rocm/envs/<runtime>/models/checkpoints/and fails with
RuntimeError: checkpoint is missing and downloads are disabled: ~/.rocm/envs/<runtime>/models/checkpoints/sd-v1-5-tiny.safetensors
ComfyUI's actual models folder, printed by both rocm comfyui install and rocm comfyui status, is ~/.rocm/envs/<runtime>/apps/comfyui/source/models.
Expected behavior
The harness should place the checkpoint under the models path: the product reports (<models path>/checkpoints/), so the default --generate-cat flow works without --comfyui-root.
Steps to reproduce
# 1. default flow (downloads 2 GB, then fails)
python scripts/comfyui_therock_gpu_test.py --rocm target/release/rocm --runtime-id <key> --generate-cat --timeout 1200
# 2. cheaper reproduction of the path derivation
python scripts/comfyui_therock_gpu_test.py --rocm target/release/rocm --runtime-id <key> --skip-install --generate-cat --no-download-checkpoint
Reproduced on a6fa76dbca5d767c1190ee47b765d16bad0e4a53 (current main).
Root cause (script)
parse_comfyui_root (line 479) takes the folder: line of the CLI output and ensure_checkpoint (line 544) assumes <folder>/models/checkpoints. But folder: is not the ComfyUI source folder:
rocm comfyui installprintsfolder: <env>/apps/comfyui(apps/rocm/src/comfyui.rs:287,app_root), and the source lives in<app_root>/source;rocm comfyui statusprintsfolder: <env>(apps/rocm/src/comfyui.rs:135,manifest.runtime_root).
The models path: line, which both commands print right after folder: and which is the correct location, is ignored by the harness. (The inconsistency of folder: itself is filed as #391.)
Possible solution
Parse models path: (^\s*models path:\s*(.+?)\s*$) and use <models path>/checkpoints for ensure_checkpoint; keep --comfyui-root as the override. Workaround today: pass --comfyui-root <env>/apps/comfyui/source explicitly.
Your environment
- rocm-cli channel: nightly (TheRock
10.1.0a20260822,gfx120X-all) - rocm-cli version (
rocm --version):rocm 0.1.0, built froma6fa76dbca5d767c1190ee47b765d16bad0e4a53 - Platform: Linux
- OS / Distro: CachyOS (Arch), kernel 7.2
- GPU/APU: AMD Radeon AI PRO R9700 (gfx1201)
- ComfyUI 0.34.0 installed by
rocm comfyui install
Relevant log output
ComfyUI
action: install
ROCm install: 10.1.0a20260822 (build 2026-08-22)
folder: ~/.rocm/envs/therock-10.1.0a20260822/apps/comfyui
models path: ~/.rocm/envs/therock-10.1.0a20260822/apps/comfyui/source/models
installed: yes
AMD GPU: ready (2 devices)
[comfyui-gpu-test] Downloading checkpoint for the cat workflow (2.0 GB expected).
[comfyui-gpu-test] Using checkpoint ~/.rocm/envs/therock-10.1.0a20260822/apps/comfyui/models/checkpoints/sd-v1-5-tiny.safetensors.
...
RuntimeError: ComfyUI does not list checkpoint `sd-v1-5-tiny.safetensors`. If it was just downloaded, restart ComfyUI and run this test again.
Available checkpoints: []
ComfyUI
installed: yes
folder: ~/.rocm/envs/therock-10.1.0a20260822
models path: ~/.rocm/envs/therock-10.1.0a20260822/apps/comfyui/source/models
...
RuntimeError: checkpoint is missing and downloads are disabled: ~/.rocm/envs/therock-10.1.0a20260822/models/checkpoints/sd-v1-5-tiny.safetensors
Additional context
With --comfyui-root <env>/apps/comfyui/source the harness gets past this step. On this machine it then fails inside ComfyUI with hipErrorInvalidValue in the CLIP text encoder; that is a separate problem in the TheRock nightly torch wheel (torch-2.15.0a0+rocm10.1.0a20260822 ships libaotriton_v2.so.0.13.0 without aotriton.images/, so flash/mem-efficient SDPA output is unusable on gfx1201; SDPBackend.MATH and ComfyUI --use-split-cross-attention work). Mentioned only so the GPU lane's red status is not attributed to this script bug; it is not a rocm-cli defect. Related harness issue: #389.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start in scripts/comfyui_therock_gpu_test.py at parse_comfyui_root (line 479) and ensure_checkpoint (line 544), then review the install and status output showing folder: and models path:. Reproduce with --skip-install --generate-cat --no-download-checkpoint. Done means the checkpoint uses the reported models path/checkpoints directory while --comfyui-root remains an override and the harness reaches the ComfyUI generation step.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- cli, testing-qa
- Issue type
- Bug
- Difficulty
- 2/5
- Estimated time
- 1-3 hours
- Activity status
- Active
- Clarity
- Clearly specified
- Newbie friendliness
- 88/100