GitLab

* [xpu/triton] Add trtion dequantization kernel

This PR adds xpu backend and trtion kernel for dequantization nf4 dtype.
Trtion is an optional import.
Tests:
	tests/test_functional.py::TestQuantize4BitFunctional supported nf4/fp4 cases
	tests/test_functional.py::Test8BitBlockwiseQuantizeFunctional
implemented quantize_blockwise with binary search that works faster for XPU
        tests/test_linear4bit.py
Signed-off-by: Dmitrii Makarenko <dmitrii.makarenko@intel.com>

* align with ipex code

* enable test for ipex

* test_kbit_backprop: skip no longer needed

* remove unused

---------
Signed-off-by: Dmitrii Makarenko <dmitrii.makarenko@intel.com>

* Deprecation cleanup: remove histogram_scatter_add_2d

* Deprecation cleanup: vectorwise_mm_dequant

* Deprecation cleanup: vectorwise_quant

* Remove unused test

* Optimizer test cleanup

* Deprecations: remove estimate_quantiles, create_quantile_map

* Move deprecated test

* Tests: add linux x64 cpu+ipex to nightly CI workflow

* typo

* Tests: guard linear8bit compile test for ipex cpu issue

* Tests: xfail opcheck for 4bit quantization with floating storage dtypes

* Tests: xfail opcheck for 4bit quantization with floating storage dtypes

* Tests: skip test_gemv_eye_4bit on CPU with bf16 when not supported by torch

* Tests: skip test_gemv_eye_4bit on CPU with bf16 when not supported by torch

* enable ipex
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix cpu 8bit quantization
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix int8 and nf4 cpu inference
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* add cpu fp4 and rem
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix dequantize nf4 xpu
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix ipex op
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix dequantize nf4 name
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix dequantize nf4 ipex
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix matmul8bitfp
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* enable cpu tests
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix format
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix quantize blockwise output shape
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix quant_storage bf16 and gemv cpu
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix cpu tests
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix xpu tests
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix lib
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* skip xpu dequantize blockwise op check
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix matmul8bit
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* skip not used function teests
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix matmul8bit fp
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* check ipex before MatMul8bitFp
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* update ipex install guide
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* update install guide
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix error log
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix error lof
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* update comment
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* move torch op to default
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* revert ipex check
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix code tabledevice
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix code table device
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix xpu ops
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

---------
Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* Add torch.compile tests

* Tests: WA aarch64 CPU regressions for torch 2.6.0; add Windows torch==2.7.0+cu118 test config

* Tests: skip torch.compile for cuda on windows

* General cleanup & test improvements

* Tests: WA numpy 2 compat issue for torch<2.3

* Tests: update aarch64 cpu min torch version

* Tests: update aarch64 cpu min torch version

* Tests: update aarch64 cpu min torch version

* Test g5g runner

* Switch L4 to L40S runner; swap GitHub Linux T4 runner for AWS g4dn

* Run tests on last 2 pytorch stable releases

* Run tests on last 2 pytorch stable releases

* Improvements for testing suite

* Add workflow for macOS arm64 CPU tests

* Run unit tests on GH Actions

* fix

* fix

* trigger workflow

* Update

* Update

* Update

* Run tests nightly

* Disable paged optimizer test on Windows

* Skip unit tests on Windows for CUDA 12.x (driver on runner is too old)

* Additional 4bit CPU ops

* Additional 4bit CPU ops

* Implement additional device-agnostic ops and test updates

* More test fixes

* int8 tests passing

* Fix feature flag for multi_backend

* Include device support tags for transformers multi-backend compatability; add xpu() and cpu() to Params4bit

* Make test suite more device-agnostic

* Additional device agnostic tests

* Additional device agnosticism for tests

* Add BNB_TEST_DEVICE env var to manually select device for unit tests

* Include device support tags for transformers multi-backend compatability; add xpu() and cpu() to Params4bit

* Make test suite more device-agnostic

* Additional device agnostic tests

* Additional device agnosticism for tests

* Add BNB_TEST_DEVICE env var to manually select device for unit tests

* Small bugfix for int8 test

* Exclude backward() from code coverage reports

* Params4bit: don't try to quantize when moving to meta device

* Testing cleanup

* More test cleanup

* Additional deprecations/removals.

* Skip benchmark, deprecated, slow tests by default

* Drop Python 3.8 support.

* Formatting

* Sketch out first custom op registration

* Add note

* Initial int8 op registration

* Cleanup some deprecated functions.

* Int8 ops updates; tests

* Implement 4bit quant/dequant ops

* Fix nested quant

* cleanup

* Test improvements

* Clean up and improve tests

* Add higher level custom op for int8 matmul + dequant + bias

* Add gemv 4bit custom op

* Cleanup

* Implement out kwarg overloads for custom ops

* Update PyTorch minimum to 2.1

* Deprecation updates

* Deprecation updates

* Cleanup; rename int8_linear_dequant -> int8_scaled_mm

* Bump min pytorch to 2.2

* cleanup

* Test reorganization

* Remove deprecated supports_igemmlt

* More cleanup

* Cleanup obsolete C++/CUDA code

* Cleanup

* Create 'default' backend for fallback op implementations; initial CPU nf4 work

* Stub out for multi-platform

* Fix serialization tests for torch>=2.6.0

* Add example for torch.compile e2e inference

* Test update

---------
Co-authored-by: Titus von Koeller <9048635+Titus-von-Koeller@users.noreply.github.com>

* Start of int8 refactor: remove col32/col_ampere/col_turing transforms in new igemmlt implementation

* Fix unintended change

* New naive mm_dequant kernel for row-major; cleanup

* fix

* int8 refactor: initial sparse decomp, cleanup

* Int8 refactoring: remove separate NO_CUBLASLT build; more cleanup

* int8: inference optimizations, some cleanup

* int8: more tests passing, cleanup

* int8 - more cleanup, most tests passing

* int8: specify CUDA stream for int8 ops

* perf: reduce overhead from getting cudaStream ptr

* Mark some functions for deprecation.

* int8 sparse decomp: small perf improvement

* update setup.py

* Update bitsandbytes/autograd/_functions.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update bitsandbytes/research/autograd/_functions.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* int8 - perf improvement for sparse decomposition inference; deprecate get_tensor_stream() in favor of new private fn

* int8 cleanup

* Ignore ruff rule ISC001 (incompatible with formatter)

* add comment

* int8 more cleanup

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* int8: rename / deprecate old fn signatures

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* type annotation

* format update

* Update bitsandbytes/research/autograd/_functions.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* cleanup

* Add comment to explain division optimization

* more cleanup

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Update bitsandbytes/functional.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* cleanup

* Type annotations, cleanup

* remove unused kernels; improved type annotations

* small perf optimization for single-GPU systems

* small perf optimization for single-GPU systems

* update docstrings

* Improve docs and tests

* Update docstring

* Update test

* add benchmarking script

* test cleanup: add deprecated marker, move benchmarks out

* Add int8 dequant function; misc improvements

* int8 matmul fallback for inner dims not divisible by 4

* improve register usage of kInt8VectorQuant - especially for A100/H100

* disable fail-fast for package build

* maxwell compat

* ptxas verbose

* docs update

* doc update

* backward fix

* Bugfix sparse decomp

* Int8 fix for PEFT OLoRA init

* Fix test for deprecated spmm_coo

* test improvement

* doc update

* typo

* doc cleanup

* docs

* add inference benchmark script

* Add benchmarks, doc update

---------
Co-authored-by: Aarni Koskela <akx@iki.fi>

* Change 8bit optimizer blocksize 2048->256; additional bf16 support
* Update tolerances for 8bit optimizer tests

* Add AdEMAMix optimizer

* Add PagedAdEMAMix32bit, AdEMAMix32bit

* Add PagedAdEMAMix32bit, AdEMAMix32bit

* AdEMAMix: add support for alpha/beta3 scheduling

* Update paged AdEMAMix

* Embedding4bit and Embedding8bit implementation

* lint

* Update bitsandbytes/nn/modules.py
Co-authored-by: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>

* Update bitsandbytes/nn/modules.py
Co-authored-by: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>

* Update bitsandbytes/nn/modules.py
Co-authored-by: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>

* saving -> Saving

---------
Co-authored-by: Matthew Douglas <38992547+matthewdouglas@users.noreply.github.com>

* fixed test_4bit_warnings on cpu-only platforms

* fixed linear8bit-based tests for cpu only platforms

This requires to implement the __deepcopy__ method in Int8Params.
Moreover, there was an issue in the Linear8BitLT constructor that would
assign instance attributes to the class, which is now fixed.

Please review carefully that this does not impact existing code.

Tests that I ran:

- pytest tests/test_linear8bitlt.py
- in PEFT: python -m pytest -m "single_gpu_tests and bitsandbytes" tests/test_gpu_examples.py
- in PEFT: python -m pytest -m "single_gpu_tests and bitsandbytes" tests/test_common_gpu.py
- in transformers: RUN_SLOW=1 python -m pytest tests/quantization/bnb -x

As discussed internally, use state = self.__dict__.copy(), which is also
what the Python docs recommend.

* fix deepcopy and copy

* add tests

* remove line

* ruff fix

* ruff

* Update tests/test_linear4bit.py
Co-authored-by: Aarni Koskela <akx@iki.fi>

* add missing state

* ruff format

* ignore formatting commit for git blame

* Params4bit should be initialized as frozen by default

* add test for serialization round-tripping

* add comparison capability for QuantSate

* add back accidentally remove line

---------
Co-authored-by: Aarni Koskela <akx@iki.fi>
Co-authored-by: Titus von Koeller <9048635+Titus-von-Koeller@users.noreply.github.com>

Co-authored-by: Titus von Koeller <titus@vonkoeller.com>

fix erroneous correction