[GH-ISSUE #2115] ROCm Support for Dual AMD Radeon Pro V620 (GLM-5.2) #1283

New issue

Open

opened 2026-10-29 14:50:50 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-29 14:50:50 +00:00

Owner

Originally created by @DrSchnitzel57 on GitHub (Jul 31, 2026).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/2115

Reminder

I have read the above rules and searched the existing issues.

System Info

System Specifications & Environment
OS: Ubuntu 26.04
GPUs: 2x AMD Radeon Pro V620 (RDNA 2 / gfx1030)
Environment: Miniconda3 (kt_amd environment)
Dependencies: ROCm-compatible PyTorch, ktransformers v0.6.4 (installed from source via pip install .)
Native ROCm: 7.1 (System default)
Target ROCm: 6.4 (Attempted for compatibility)
Model: GLM-5.2 (GGUF format: UD-IQ2_M) or deepseek v4 flash

Reproduction

Put your message here.

Others

I cannot get it to run. After a ton of tinkering I got it to build, but it will not run.
Initial Attempt (ktransformers): Tried running via local_chat.py, but encountered structural changes in ktransformers v0.6.4 that required transitioning to the sglang server-client model.

ROCm Version Conflict: Encountered environment conflicts between Ubuntu 26.04's native ROCm 7.1 packages and the ROCm 6.4 requirement for the ktransformers/sglang stack.

Source Compilation Attempt: Attempted to compile the sgl-kernel dependency from source to explicitly support gfx1030.

Compilation Failures: Installed the ROCm 6.4 developer toolchain (rocm-developer-tools, hip-dev, rocm-llvm) alongside a Conda environment, but compilation repeatedly failed due to:

Missing hipcc compiler detection.

Persistent path misconfigurations during the build process, preventing the C++ GPU kernels from compiling for AMD hardware.

Is there a planned path for ROCm 7.1 support?

Are there pre-compiled wheels planned for gfx1030, or is there a recommended workaround for the hipcc pathing issues when compiling sgl-kernel from source on newer Ubuntu distributions?

Originally created by @DrSchnitzel57 on GitHub (Jul 31, 2026). Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/2115 ### Reminder - [x] I have read the above rules and searched the existing issues. ### System Info System Specifications & Environment OS: Ubuntu 26.04 GPUs: 2x AMD Radeon Pro V620 (RDNA 2 / gfx1030) Environment: Miniconda3 (kt_amd environment) Dependencies: ROCm-compatible PyTorch, ktransformers v0.6.4 (installed from source via pip install .) Native ROCm: 7.1 (System default) Target ROCm: 6.4 (Attempted for compatibility) Model: GLM-5.2 (GGUF format: UD-IQ2_M) or deepseek v4 flash ### Reproduction ```text Put your message here. ``` ### Others I cannot get it to run. After a ton of tinkering I got it to build, but it will not run. Initial Attempt (ktransformers): Tried running via local_chat.py, but encountered structural changes in ktransformers v0.6.4 that required transitioning to the sglang server-client model. ROCm Version Conflict: Encountered environment conflicts between Ubuntu 26.04's native ROCm 7.1 packages and the ROCm 6.4 requirement for the ktransformers/sglang stack. Source Compilation Attempt: Attempted to compile the sgl-kernel dependency from source to explicitly support gfx1030. Compilation Failures: Installed the ROCm 6.4 developer toolchain (rocm-developer-tools, hip-dev, rocm-llvm) alongside a Conda environment, but compilation repeatedly failed due to: Missing hipcc compiler detection. Persistent path misconfigurations during the build process, preventing the C++ GPU kernels from compiling for AMD hardware. Is there a planned path for ROCm 7.1 support? Are there pre-compiled wheels planned for gfx1030, or is there a recommended workaround for the hipcc pathing issues when compiling sgl-kernel from source on newer Ubuntu distributions?