Built For RX 7900 XTX on Win 11/ROCm With GPT5.5 High - Errors In Python Scripts Fixed

[HaiderGill]

Hi,

Pre Reqs: -
Installed Windows ROCm to C:\ROCm
Installed Python 3.14
Installed Git for Windows
My local GIT repo is f:\temp\git_hub_repos\llama_cpp_turboquant\

GPT5.5 High Python test scripts fixes were: -
Added explicit UTF-8 encoding when writing NIAH output files
In scripts/niah_test.py, the JSON and Markdown result files now use:
open(path, "w", encoding="utf-8")
This avoids Windows default codepage problems when output contains non-ASCII text.

Added explicit UTF-8 encoding when reading those files in tests
In tests/test_niah.py and tests/test_turbo_hardware_diag.py, file reads now use:
read_text(encoding="utf-8")
open(path, encoding="utf-8")
This makes the tests match the script behavior and avoids Windows encoding failures.

Fixed a path comparison in the NIAH test
The test was comparing a Path object after converting it to a string:
assert str(call_args[0][3]) == "/tmp/custom_out"
It was changed to compare Path to Path:
assert call_args[0][3] == Path("/tmp/custom_out")
That is cleaner and more portable.

Fixed os.getloadavg mocks for Windows
Windows Python may not have os.getloadavg, so the test patches now use:
patch("os.getloadavg", ..., create=True)
That lets the test create the mocked attribute even when it does not exist on Windows.

Made the Linux cache-path test handle Windows path separators
The test expected paths containing:
cache/index
But Windows paths use backslashes. The mocked path string is now normalized first:
s = str(self).replace("\", "/")
That allowed the expected "Could not get Linux cache hierarchy" path to be exercised correctly.

Overall result
The changes were not logic changes to the main hardware diagnostic behavior; they were mostly Windows portability fixes for the tests, especially encoding, path handling, and mocking Linux/macOS-only APIs on Windows.

In DOS prompt (run as admin): -
winget install Kitware.CMake
winget install Microsoft.VisualStudio.2022.BuildTools
cd f:\temp\git_hub_repos\llama_cpp_turboquant
git clone https://github.com/TheTom/turboquant_plus.git
cd turboquant_plus
..venv\Scripts\activate.bat
pip install -e ".[dev]"
pip install ninja
pip install numpy
pip install SciPy
python -m pytest tests/ -v
git clone --branch feature/turboquant-kv-cache https://github.com/TheTom/llama-cpp-turboquant.git llama-cpp-tq
cd f:\temp\git_hub_repos\llama_cpp_turboquant\turboquant_plus\turboquant\llama-cpp-tq

set PATH=C:\ROCm\7.1\bin;C:\Program Files\CMake\bin;%PATH%
set HIP_PATH=C:\ROCm\7.1

In
cmake -S . -B build -G Ninja
-DGPU_TARGETS=gfx1100
-DGGML_HIP=ON
-DCMAKE_C_COMPILER=clang
-DCMAKE_CXX_COMPILER=clang++
-DCMAKE_BUILD_TYPE=Release
-DCMAKE_C_FLAGS_RELEASE="-O3 -DNDEBUG"
-DCMAKE_CXX_FLAGS_RELEASE="-O3 -DNDEBUG"
cmake --build build --config Release

I'm using the ByteShape Qwen3.6-35B-A3B-MTP-GGUF, it not standard quantization: -
ShapeLearn: Their core algorithm that adaptively taps into the AI training process to learn optimal datatypes for specific neural-network parameters and inputs. It selectively assigns lower or higher precision per-tensor (e.g., INT2 to FP8), rather than using a flat, static rate. This allows models to retain high accuracy while vastly reducing their arithmetic complexity and memory footprint.

ShapeSqueeze: A lossless compression layer that applies per-value encoding to minimize off-chip data transfers, saving up to an extra 40% in model size

Batch file settings
set LLAMA_CACHE="E:\programs\llama_cache\Byteshape-Qwen3.6-35B-A3B-IQ4"
set TURBO_LAYER_ADAPTIVE=7
@echo off
.\bin\llama-server.exe ^
-m "E:\programs\llama_models\Byteshape-Qwen3.6-35B-A3B-IQ4_XS-3.53bpw\Qwen3.6-35B-A3B-IQ4_XS-3.53bpw.gguf" ^
--verbose ^
--mmproj-auto ^
--spec-type draft-mtp ^
--spec-draft-n-max 4 ^
--cache-type-k q8_0 ^
--cache-type-v turbo2 ^
--spec-draft-type-k q8_0 ^
--spec-draft-type-v turbo2 ^
--host 127.0.0.1 ^
--port 8080 ^
--alias Byteshape-qwen3.6-35b-a3b ^
--ctx-size 262144 ^
--parallel 1 ^
--jinja ^
--reasoning auto ^
--reasoning-format deepseek ^
--cache-reuse 256 ^
--n-gpu-layers auto ^
--temp 0.6 ^
--top-p 0.95 ^
--min-p 0.0 ^
--top-k 20 ^
--repeat-penalty 1.0 ^
--presence-penalty 0.0 ^
--chat-template-kwargs "{"preserve_thinking":true}"

pause

set TURBO_LAYER_ADAPTIVE=7 is important.

If anybody wants the compiled version can probably put it on Dropbox...