- use HalfRope position encodings
- zero-initialize the caches for each Generate at position 0
The lack of the latter made the tests in gemma_test dependent on each other.
PiperOrigin-RevId: 694509054
Add Extents2D, Range2D vocab types
Matmul uses ConstMat for inputs and RowPtr for output
Move RowVectorBatch to basics.h
Separate threading.cc
Fix topology string: report cores not LPs, and #HT
Move QStride/IsMHA into LayerConfig
ImageTokens does not require make_unique.
matmul_test: no longer require template args
PiperOrigin-RevId: 692963605
Use ModelConfig values for ImageTokens.
Output timing info for image token generation.
Add a method to copy image data into Image class directly.
Minor changes: pipe ModelTraining to more places.
PiperOrigin-RevId: 690572283
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
Move image related config values from LayerConfig to ModelConfig.
Minor changes: Add a few comments, remove gcpp:: qualification where it wasn't needed in a few places, define local constants in VitAttention.DotSoftmaxWeightedSum()
PiperOrigin-RevId: 687210519
Changed CompressedLayer and CompressedWeights to be constructed with an instance of a LayerConfig and WeightsConfig respectively.
Added CompressedModel to remove ByteStorageT and get rid of most of the type casting, as well as allowing the default destructor to be used and work properly.
Adjusted WeightsWrapper and ForwardLayer etc to match.
The only remaining template arg is the weight type.
This enables all the instantiations to be deleted, apart from one per type.
It also enables (but not yet done) the config to be stored in the blob file instead of having to be specified separately.
Reduces the size of the gemma_lib and weights shared libraries by a factor of 4.3 and 3.2 respectively.
PiperOrigin-RevId: 686870060
add app.h comment
compress-inl: remove unused typedef
gemma-inl: add missing HWY_ATTR and cast
separate sum-inl.h and basics.h headers
replace more hwy::bfloat16_t with BF16
update include pragmas
update dot_test thresholds
update Highway version in Bazel for HWY_RCAST_ALIGNED fix
PiperOrigin-RevId: 684464326
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
Use in dot_test
app.h: add new flags and rename num_threads to max_threads
matmul: Parallelize MatMulSlow and enable spinning, more large/fewer medium test cases
PiperOrigin-RevId: 683216386
See https://arxiv.org/abs/2407.07726 for a description of the model.
Because PaliGemma operates as a prefix-LM on the image+prompt, add support for that.
PiperOrigin-RevId: 677841119
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
Use args only in GemmaEnv constructor, store everything else in RuntimeConfig.
Add runtime option to turn off thread spinning.
PiperOrigin-RevId: 670467320
Supports converting all weight/activation formats to native MulT (bf16/f32)
Also:
- ConstMat/MutableMat for const correctness
- Move RowVectorBatch to allocator.h so it can be used from Matmul
- Add matmul.h so MatMulEnv can be used from Activations
- Remove kMaxThreads, detect from PerClusterPools
- Build fix: -inl.h files must be textual_hdrs, and highway.h should precede -inl.h
```
zen4 new
64, 24576, 3072, add=0, MatTA=bf16, MatTB=sfp: 616.6 GFLOPS.
64, 3072, 24576, add=0, MatTA=bf16, MatTB=sfp: 460.7 GFLOPS.
64, 24576, 3072, add=0, MatTA=f32, MatTB=sfp: 598.6 GFLOPS.
64, 3072, 24576, add=0, MatTA=f32, MatTB=sfp: 435.6 GFLOPS.
zen4 old
64, 24576, 3072, add=0, MatTA=f32, MatTB=sfp: 257.5 GFLOPS.
64, 3072, 24576, add=0, MatTA=f32, MatTB=sfp: 231.9 GFLOPS.
```
PiperOrigin-RevId: 663729812
LayersOutputFunc is no longer invoked for "blocks" and "final_norm" outputs.
Instead, we directly expose the Activations structure.
PiperOrigin-RevId: 663409316
Split attention into functions, move into class.
Fuse Rope and MulBy, allow non-in-place version to avoid copy from q to KV.
Sink if() into MaybeLogitsSoftCap.
PiperOrigin-RevId: 661168418
Limit thread counts to detected. Add max_clusters arg.
Update detection logic to check for smt0 - previously we pinned to some siblings.
PiperOrigin-RevId: 659755311
- Allocate only the required KV caches and activation batch size
- Add flags for batch sizes
- Const-correct interface: Span of const int.
- Also clean up the KVCache arg to a span.
- Move kPrefillBatchSize into RuntimeConfig and remove related global constants.
PiperOrigin-RevId: 655893197
SfpCodec::Dec2F and ComressTraits<T>::Decompress2 for all supported types. It also allows to remove one of the specializations of GEMM_4x4_Tile, handling compressed MatB with one function. As before even when MatA is bf16 it is using 32-bit registers for computations.
Measurements for a 2b-it sfp-encoded model on a AMD Ryzen Threadripper PRO 3945WX 12-Cores:
baseline:
```
32.6254 prefill tokens / sec
8.91429 tokens / sec
115 milliseconds time to first token
```
this change:
```
54.3045 prefill tokens / sec
16.8191 tokens / sec
56 milliseconds time to first token
```
PiperOrigin-RevId: 651369694
This CL adds a new function to Gemma that allows for batching of multiple prompts. The function takes a vector of prompts and returns a vector of responses. The prompts are processed in parallel, and the responses are returned in the same order as the prompts.
PiperOrigin-RevId: 648367559
For regular (non-SSM) Gemma models, kGriffinLayers is by definition always zero
and kGemmaLayers is just the number of layers.
PiperOrigin-RevId: 644384531
Helper functions to tokenize/wrap
Move LayersOutputFunc into RuntimeConfig
AcceptFunc passes the probability
Implement StringFromType using the parser, and verify results match
PiperOrigin-RevId: 643255119
accept_token: allow default, check if empty when using
allow mixing sample_func and stream_func, call the latter after the former
Also fix missing includes/deps.
PiperOrigin-RevId: 642240012
This changes the command line flags, but the default value retains the previous behavior.
Also add a CreateGemma helper to enable extra args without interface changes.
PiperOrigin-RevId: 641266411
With this addition the ComputeCrossEntropy function can be moved
to its own library, because now we can compute it using only the
public API functions from gemma.h
Split common and weights into separate lib
Remove common-inl (does not have to be SIMD code), activations.cc
Centralize switch(Model) to avoid duplication
Move CompressWeightsT to compress_weights.cc
Move LoadWeights to weights.cc
PiperOrigin-RevId: 640869202
This is still in progress / experimental, currently it is only
implemented for normal gemma MQA attention layers, and no
parallelism is added yet for backward pass.
Since we need to remember all activations from all layers, the
forward pass was also reimplemented with a new activation data
structure.
Remove extra Dot() overload
MatVecAdd always adds, use MatVecT<kAdd> if conditional.
Remove ununsed MatVecAddLoop and MatVecLoop
No longer tsan-verify even_odd
PiperOrigin-RevId: 631377279
Move the loop over the tokens inside the attention block and
then create kHeads * num_tokens threads.
This helps the multi-threaded speed only in case of the 2b gemma
model, but to be consistent we move the loop over the tokens inside
the griffin recurrent layer and the FFW layer as well. This is
also a preparation for using the MatMul operation later.
Benchmark results (summarization with 1600 tokens for prefill
and essay writing with 500 tokens for generation):
```
Prefill speed
Num threads BEFORE AFTER
32 61.76 t/s 65.08 t/s
64 89.46 t/s 98.62 t/s
```
We compute all three projections with one MatVec and then copy
the kv part to the cache.
Benchmark results for 7b-it model that uses MHA blocks (summarization with
1600 tokens for prefill and essay writing with 500 tokens for generation):
```
Prefill speed Generation speed
Num threads BEFORE AFTER BEFORE AFTER
32 13.75 t/s 14.80 t/s 9.22 t/s 9.77 t/s
64 19.89 t/s 24.83 t/s 12.46 t/s 13.66 t/s
```
We use MatVec instead of MatVecLoop for the per-head dense layers,
because we can parallelize more on the rows of the matrix than
on the number of heads. This will be even more efficient after
we rearrange the weights and can have a single MatVec operation.
Benchmark results (summarization with 1600 tokens for prefill
and essay writing with 500 tokens for generation):
```
Prefill speed Generation speed
Num threads BEFORE AFTER BEFORE AFTER
32 58.24 t/s 61.79 t/s 32.11 t/s 32.62 t/s
64 83.62 t/s 92.00 t/s 41.10 t/s 41.80 t/s
```
Instead of MatVecLoop, we use MatVec and we combine k and v
into one 2 * kQKVDim long vector so that K and V projections
can be combined into one MatVec operation.
Benchmark results (summarization with 1600 tokens for prefill
and essay writing with 500 tokens for generation):
```
Prefill speed Generation speed
Num threads BEFORE AFTER BEFORE AFTER
4 9.81 t/s 9.96 t/s 8.39 t/s 8.46 t/s
18 31.50 t/s 36.67 t/s 23.10 t/s 25.83 t/s
32 45.36 t/s 58.91 t/s 27.60 t/s 31.25 t/s
64 57.72 t/s 80.64 t/s 35.40 t/s 39.76 t/s
```
We only used inner_pool in the prefill FFW function, and there we
can achieve sufficient parallelism on the rows of the matrix-vector
multiplications.
Benchmark results on a 1600-token summarization task:
```
Prefill speed
Num threads BEFORE AFTER
4 9.24 t/s 9.76 t/s
18 31.41 t/s 31.16 t/s
32 31.41 t/s 45.13 t/s
64 31.03 t/s 57.85 t/s
```
Move Path into io.h and use for opening files.
Removes dependency of gemma_lib on args.
Separate Windows codepath instead of emulating POSIX functions.
Plus lint fixes.
PiperOrigin-RevId: 626279004
Daniel Keysers
Jan Wassenberg
Ray Smith
Krzysztof Ostrowski
RangerUFO
Zoltan Szabadka
Apoorv Reddy
Paul Chang
Copybara-Service
The gemma.cpp Authors
Phil Culliton
Andrey Vlasov
Kan Wu
Zelalem Aweke
Sam Kaufman
Charles Chan
Andrey Mikhaylov