Commit Graph

8 Commits

Author SHA1 Message Date
Jan Wassenberg a60b564b88 Infra improvements (2)
ops.h: move CreateInvTimescale to allow calling without depending on gemma
Pass around MatMulEnv instead of pools to avoid re-creating the env
profiler.h can now be used outside SIMD code
allocator: add StepBytes and QuantumSteps
rename worker thread with package/cluster in the name
threading: add Visit* to IndexRange
PiperOrigin-RevId: 718766704
2025-01-23 01:55:19 -08:00
Jan Wassenberg 02ce1e344f Use NestedPools, add NUMA infra
Improved threading.h, fix thread counts for single package/cluster systems
Temporarily forces to a single socket. Prefill 29.28 tps, decode 6.92.

Also fix benchmarks.cc build, update tensor allocator to Allocator

PiperOrigin-RevId: 687307167
2024-10-18 08:11:18 -07:00
Ray Smith 85958f5fd3 Added MatPtr/MatPtrT/MatStorageT/MatStorage as a dynamically-sized replacement for CompressedArray.
Definition of array size is moved to the constructor.
Allocation is separate and parallelized.
All users of weights_raw.h migrated to CompressedWeights and weights_raw.h deleted.
Replaced all previous ForEachTensor functions with a single unified function.

PiperOrigin-RevId: 684451604
2024-10-10 08:22:30 -07:00
Jan Wassenberg 5a71d819cb Also enable f64 dot/sum for <f32 inputs
Add bf16 support to Dot/SumKernelDouble in the same way as *Compensated.

PiperOrigin-RevId: 682308683
2024-10-04 07:12:10 -07:00
Jan Wassenberg 5e812f07f5 Use f64 Dot and sum in softmax - faster than Cascaded
Also let the kernel specify the Raw and State types,
rename WeightT/VecT -> WT/VT.

PiperOrigin-RevId: 680464427
2024-09-30 01:22:09 -07:00
Jan Wassenberg 35fdf848c7 Cascaded summation for Softmax
This can affect generation results after a few hundred tokens.

Also remove profiler from DecompressAndCall, use Add instead of +=,
use PackedSpan for args and remove alignment requirement.
Changing accumulation order in AssimilateCascadedSums updates dot_test thresholds.

PiperOrigin-RevId: 676891797
2024-09-20 10:31:23 -07:00
Jan Wassenberg 8c0a8834c1 Major compression update, arbitrary-len unpack + new Dot
Compression:
* Implement {any packed} x {bf16, f32} 'Load2' and DecompressAndZeroPad
* New compression test for all packed formats, add to GEMMA_TEST_FILES, remove from sfp/nuq_test
* Decompress->DecompressAndZeroPad, use PackedSpan for args with bounds checking
* NUQ: support arbitrary-length enc/dec
* New compression/shared, remove sfp.h and nuq.h
* Move Store2 into Traits and provide Compress2 wrapper
* Remove unused Decompress()-with-pool overload
* Simplify CompressedArrayLen, rename to CompressedArrayElements
* Remove unused DistortionStats b_l1_

Misc:
* Add compensated and Kahan dot, support any length
* Use same Dot function everywhere
* Move exact arithmetic functions into fp_arith
* use FloatPtr and MatPtr typedefs in tests; less stack usage
* Rename args to packed/raw
* Remove Traits::Name, instead TypeName<T>()
* Move kMaxSFP and kClusters/kGroupSize into Sfp/NuqStream
PiperOrigin-RevId: 672868468
2024-09-10 02:22:19 -07:00
Jan Wassenberg 2308514e5a Experiment with compensated dot product.
ULP difference vs exact is 0..1, vs 200-5000 for previous.
Runtime overhead is 2.5-4x for f32 input.

PiperOrigin-RevId: 668084019
2024-08-27 12:05:35 -07:00