nixpkgs/pkgs/development/python-modules/compressed-tensors/default.nix

170 lines
5.9 KiB
Nix

{
lib,
stdenv,
buildPythonPackage,
fetchFromGitHub,
# build-system
setuptools,
setuptools-scm,
# dependencies
frozendict,
loguru,
pydantic,
torch,
transformers,
# tests
accelerate,
nbconvert,
nbformat,
pytestCheckHook,
writableTmpDirAsHomeHook,
}:
buildPythonPackage (finalAttrs: {
pname = "compressed-tensors";
version = "0.17.1";
pyproject = true;
# Release on PyPI is missing the `utils` directory, which `setup.py` wants to import
src = fetchFromGitHub {
owner = "neuralmagic";
repo = "compressed-tensors";
tag = finalAttrs.version;
hash = "sha256-14AHbokDlN5iXy/fvOq7Xp1OS8N1b+Xpxd33KOylWiU=";
};
postPatch = ''
substituteInPlace pyproject.toml \
--replace-fail "setuptools_scm==8.2.0" "setuptools_scm"
'';
build-system = [
setuptools
setuptools-scm
];
dependencies = [
frozendict
loguru
pydantic
torch
transformers
];
pythonImportsCheck = [ "compressed_tensors" ];
nativeCheckInputs = [
accelerate
nbconvert
nbformat
pytestCheckHook
writableTmpDirAsHomeHook
];
disabledTests = [
# these try to download models from HF Hub
"test_apply_tinyllama_dynamic_activations"
"test_compress_model"
"test_compress_model_meta"
"test_compressed_linear_from_linear_usage"
"test_decompress_model"
"test_get_observer_token_count"
"test_kv_cache_quantization"
"test_load_compressed_sharded"
"test_model_forward_pass"
"test_save_compressed_model"
"test_target_prioritization"
"test_expand_targets_with_llama_stories"
# AssertionError: Torch not compiled with CUDA enabled
"test_register_parameter"
"test_register_parameter_invalidates"
"test_set_item"
"test_set_item_buffers"
"test_update_offload_parameter_with_grad"
]
++ lib.optionals (stdenv.hostPlatform.isLinux && stdenv.hostPlatform.isAarch64) [
# aarch64-linux fails cpuinfo test, because /sys/devices/system/cpu/ does not exist in the sandbox:
# RuntimeError: Failed to initialize cpuinfo!
"test_mxfp4_scales_e2e"
"test_mxfp8_scales_e2e"
]
++ lib.optionals stdenv.hostPlatform.isDarwin [
# AssertionError: Numerical precision error
"test_quantization_enabled_disabled"
# torch._inductor.exc.InductorError: ImportError: dlopen(/nix/var/nix/builds/nix-25002-542173852/torchinductor__nixbld1/xo/cxovsevcfanmw7lgoddbnyhoxes3nzlu7ecugxedaq2zr4f6b2qh.main.so, 0x0002):
# symbol not found in flat namespace '___kmpc_barrier'
"test_compress_decompress_module"
# AssertionError: Torch not compiled with CUDA enabled
"test_match_quantizable_tensors"
"test_narrow_match_true_child_only"
"test_narrow_match_false_when_parent_also_matches"
"test_narrow_match_false_when_neither_matches"
"test_narrow_match_iterable_targets_any_true"
"test_narrow_match_with_explicit_module_argument"
"test_narrow_match_top_level_behavior_documented"
"test_complex_model_matching"
"test_parameter_and_module_consistency"
"test_all_functions_with_regex"
"test_match_quantizable_tensors"
"test_mtp_tensors_saved_correctly"
"test_index_updated"
"test_single_shard_dest_creates_index"
"test_no_mtp_tensors_raises"
"test_missing_dest_files_raises"
"test_custom_mtp_prefix"
"test_apply_config_detects_deepseekv3_attention_and_hooks"
"test_initialize_module_for_quantization_offloaded"
"test_correctness_model"
"test_random_matrix_device_handling"
"test_memory_sharing"
"test_attention_cache"
# TypeError: Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.
"test_compressed_model_inference_with_hook"
];
disabledTestPaths = [
# these try to download models from HF Hub
"tests/test_quantization/lifecycle/test_apply.py"
# RuntimeError: The weights trying to be saved contained shared tensors
"tests/test_transform/factory/test_serialization.py::test_serialization[True-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[True-random-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-random-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-True-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-True-random-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-False-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[False-False-random-hadamard]"
# AssertionError: Torch not compiled with CUDA enabled
"tests/test_transform/factory/test_serialization.py::test_serialization[True-True-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[True-True-random-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[True-False-hadamard]"
"tests/test_transform/factory/test_serialization.py::test_serialization[True-False-random-hadamard]"
# AttributeError: 'NoneType' object has no attribute 'type'
"tests/test_compressors/distributed/test_distributed_compression.py"
"tests/test_compressors/distributed/test_module_parallel.py"
"tests/test_compressors/model_compressors/test_model_compressor_distributed.py"
"tests/test_offload"
]
++ lib.optionals stdenv.hostPlatform.isDarwin [
# AssertionError: Torch not compiled with CUDA enabled
"tests/test_transform/utils/test_hadamard.py"
"tests/test_utils/test_match.py"
];
meta = {
description = "Safetensors extension to efficiently store sparse quantized tensors on disk";
homepage = "https://github.com/neuralmagic/compressed-tensors";
changelog = "https://github.com/neuralmagic/compressed-tensors/releases/tag/${finalAttrs.src.tag}";
license = lib.licenses.asl20;
maintainers = with lib.maintainers; [ sarahec ];
};
})