Commit Graph

9 Commits

Author SHA1 Message Date
Jan Wassenberg c8d92948f4 Move fields, io* and blob* from compression/ into io/
PiperOrigin-RevId: 755445712
2025-05-06 11:17:19 -07:00
Jan Wassenberg 275135d7e8 Rename-only: remove Allocator2 etc suffixes now that refactoring is complete
PiperOrigin-RevId: 755397220
2025-05-06 09:12:43 -07:00
Jan Wassenberg 8d0882b966 Huge refactor of weight handling and model loading.
Weight handling:
- new ModelStore2 supports both pre-2025 multi-file and single-file formats
- simpler ForEachTensor with TensorArgs
- tensors are constructed with their full suffixed name

I/O:
- support mmap and stride
- Simplified SbsWriter, single insert(); add SbsReader

Misc:
- kMockTokenizer: allow creating with unavailable tokenizer
- configs.h: Simpler enum validity checks via kSentinel
- matmul.h: remove unused enable_bind (now in allocator.h)
- tensor_info: single TensorInfoRegistry class, rename from tensor_index.h

Frontends:
- Replace Allocate/CreateGemma with ctor(LoaderArgs, MatMulEnv&)
- Deduce model/weight type, remove --model and parsing
- Replace most common.h includes with configs.h
- Remove --compressed_weights, use --weights instead
- Remove ModelInfo, replaced by ModelConfig.

Backprop:
- Reduce max loss, remove backward_scalar_test (timeout)
- Update thresholds because new RandInit changes rng eval order and thus numerics
PiperOrigin-RevId: 755317484
2025-05-06 04:44:21 -07:00
Jan Wassenberg 8532da47f7 Major refactor of allocator/args:
use new ThreadingContext2 instead of monostate/init in each frontend
Add ThreadingArgs(replaces AppArgs)

backprop: use Packed() accessor and MakePacked factory and row-based access to allow for stride
compress_weights: remove, moving to py-only exporter instead

Move MatPtr to mat.h and revise interface:
- Generic MatOwner
- rename accessors to Packed*
- support stride/row accessors, fix RowPtr stride

Add TypeBits(Type)
Move GenerateMat to test_util-inl for sharing between matmul test/bench
Move internal init to gemma.cc to avoid duplication
Rename GemmaEnv model_ to gemma_ for disambiguating vs upcoming ModelStorage
Remove --compressed_weights, use --weights instead.
tensor_index: add ExtentsFromInfo and TensorIndexLLM/Img
Allocator: use normal unique_ptr for AllocBytes so users can call directly
threading: use -> because AlignedPtr no longer assumes arrays
PiperOrigin-RevId: 745918637
2025-04-10 01:29:54 -07:00
Jan Wassenberg 1b72c22345 Refactor Gemma ctor and improve pool NUMA support
Gemma receives a MatMulEnv arg, with comment on lifetime
Split threading into topology so the latter can be used in allocator
Add AllocClasses() for non-POD (ThreadPool)
Support binding pool to NUMA node
Update threading_test with latency measurements
Also update Highway version.

PiperOrigin-RevId: 736904748
2025-03-14 10:19:00 -07:00
Jan Wassenberg b0fe9a43e6 Further speed up blob_compare: single alloc, use dual sockets
PiperOrigin-RevId: 724947361
2025-02-09 10:53:49 -08:00
Jan Wassenberg f31e12e63b Improved blob diff: parallel, tolerance for float
PiperOrigin-RevId: 724060325
2025-02-06 13:46:28 -08:00
Daniel Keysers 7af2e70321 Add python wrappers for configs and inference.
Enable building compression/python/compression_test using bazel.
Add default image path for image_test and paligemma_test.

PiperOrigin-RevId: 720583438
2025-01-28 08:22:03 -08:00
Ray Smith 96513a8dc3 Added a blob_compare tool that compares two sbs files that may have the blobs in a different order
PiperOrigin-RevId: 696458888
2024-11-14 03:26:32 -08:00