# Extra dependencies for test_llm.py (LLM quantized models)
# These are NOT needed by test_hf_transformers.py
#
# Versions below are hard-pinned (==) intentionally. transformers/gptqmodel are bumped
# deliberately, not automatically: a silent upgrade changes the GPTQ backend selection
# (e.g. gptqmodel auto-selecting TorchFusedQuantLinear -> QUANT_TYPE "torch_fused") and the
# generated graph, which previously broke conversion. Bump these together and re-validate
# the opt_gptq entry in test_llm.py before raising them.

transformers==5.5.3
huggingface-hub==1.10.1

# kernels (and kernels-data) are pulled transitively by transformers and gptqmodel, neither of
# which caps the version. kernels>=0.15 made LayerRepository require a version/revision, which
# transformers 5.5.3's hub_kernels.py constructs without -> ImportError at "import transformers".
# Keep at the validated 0.14.1.
kernels==0.14.1
kernels-data==0.14.1

# quantized model deps
autoawq==0.2.9; platform_system == "Linux" and platform_machine == "x86_64"
triton==3.6.0; platform_system == "Linux" and platform_machine == "x86_64"
gptqmodel==6.0.3; platform_system == "Linux" and platform_machine == "x86_64" and python_version < "3.12"

peft==0.18.1; platform_system == "Linux" and platform_machine == "x86_64" and python_version < "3.12"

# optimum (required by transformers' GptqHfQuantizer for GPTQ model loading)
optimum==2.1.0; python_version < "3.12"

# compressed-tensors (required for compressed-tensors quantized models, e.g. cyankiwi/Qwen3.5-4B-AWQ-4bit)
compressed-tensors==0.17.1
