Problem Description

I have Radeon R9700. I noticed in ollama logs:

rocblaslt error: Cannot read "TensileLibrary_lazy_gfx1201.dat": No such file or directory
rocblaslt error: Could not load "TensileLibrary_lazy_gfx1201.dat"

as I run ollama in podman container, I mounted the 7.2.3 libraries from the host in the quadlet:

Volume=/opt/rocm-7.2.3/lib/hipblaslt/library:/lib/ollama/rocm/hipblaslt/library:ro
Environment=HIPBLASLT_TENSILE_LIBPATH=/lib/ollama/rocm/hipblaslt/library

then the error in the podman/ollama logs was gone.

However, I got a MASSIVE hit to performance. A medium-sized prompt run against qwen3.6 took 4 minutes to complete. Mostly spent in cold pre-filling.

when i eventually commented out these 7.2.3 libraries and instead added:
Environment=ROCBLAS_USE_HIPBLASLT=0

the performace was back to normal, the same prompt took like 20 secs to execute.

Took me a whole day of scratching my head (and help from Opus) to figure it out... :)

Operating System

Debian 13 Trixie

CPU

AMD Ryzen 7 9700X

GPU

AMD Radeon AI PRO R9700

ROCm Version

ROCM 7.2.3

ROCm Component

No response

Steps to Reproduce

create an ollama podman container and provide the host ROCm 7.2.3 libraries to it:

Provide the missing hipBLASLt kernel directory from the host

Volume=/opt/rocm-7.2.3/lib/hipblaslt/library:/lib/ollama/rocm/hipblaslt/library:ro
Environment=HIPBLASLT_TENSILE_LIBPATH=/lib/ollama/rocm/hipblaslt/library

then run a medium sized prompt (50k long) against it:

time curl http://localhost:11434/api/chat -d @prompt_file.json
real 4m12.321s

when the quadlet config for ollama is changed to:
#Volume=/opt/rocm-7.2.3/lib/hipblaslt/library:/lib/ollama/rocm/hipblaslt/library:ro
#Environment=HIPBLASLT_TENSILE_LIBPATH=/lib/ollama/rocm/hipblaslt/library
Environment=ROCBLAS_USE_HIPBLASLT=0

the same prompt runs much faster:
time curl http://localhost:11434/api/chat -d @prompt_file.json
real 0m29.903s

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response