* add support for l3b, new tokenizer
* add todo
* Add todo and use k_s model
* Use the official tokenizers.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* Utilize batches in Stable Diffusion that were already there, but unutilized.
Also refactor out the `save_image` function.
* Clippy + cosmetic fixes.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* Start adding the recurrent-gemma model.
* More griffin.
* Add the example + get the weights to load from the HF version.
* More inference code.
* Rope + kv-cache on the attention side.
* Add to the inference code.
* Add more to the recurrent gemma inference.
* Get some first inference to run.
* Add the softcap on logits.
* Fixes.
* Use partial rotary embeddings.
* Get inference to work.
* Add a comment.
* And add a readme.
* moondream implementation
* add moondream example
* change config default activation
* Add assets and integrate phi mixformer with example
* Make use of kv cache and fix seq_len bug; Clean up example code
* Add README link to example
* Remove pos_embed scaling; Remove assets; Add to README; Expand VisionConfig
* Delete image
* Use apply instead of forward
* Use latest release special token; Fix token/s accuracy; Use GeluPytorchTanh in VisionConfig v2
* Add flag to use f16
* Avoid breaking the quantized version on cuda.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* moondream implementation
* add moondream example
* change config default activation
* Add assets and integrate phi mixformer with example
* Make use of kv cache and fix seq_len bug; Clean up example code
* Add README link to example
* Remove pos_embed scaling; Remove assets; Add to README; Expand VisionConfig
* Delete image
* Use apply instead of forward
* Use latest release special token; Fix token/s accuracy; Use GeluPytorchTanh in VisionConfig v2
* moondream implementation
* add moondream example
* change config default activation
* Add assets and integrate phi mixformer with example
* Make use of kv cache and fix seq_len bug; Clean up example code
* Add README link to example
* Remove pos_embed scaling; Remove assets; Add to README; Expand VisionConfig
* Delete image
* Use apply instead of forward
* Pass bos token at the beginning of tensor.
* Quantize moondream.
* Forward with image bos token.
* Clippy.
* Use q4_0 quantization.
* Add pointers for sequence and tokens; Remove seq_len conditional
* Add more cuda kernels for quantized matmul.
* Add the vec-dot bits.
* Expose the quantized matmul-vec kernels.
* Also include the quantize-q8-1 kernel.
* Glue code for the q8-1 quantization.
* mm-vec product via q8-1 quantization.
* Add a test.
* Add a mm test.
* Get the test to return some sensible results.
* Also test dmmv.
* Fix the launch params.
* Allow for tweaking the force_dmmv parameter while it's experimental.
* moondream implementation
* add moondream example
* change config default activation
* Add assets and integrate phi mixformer with example
* Make use of kv cache and fix seq_len bug; Clean up example code
* Add README link to example
* Remove pos_embed scaling; Remove assets; Add to README; Expand VisionConfig
* Delete image
* Use apply instead of forward
* CLIP model implementation with example
* CLIP Implementation fixes, batch images
* CLIP model remove images from git
* CLIP model remove unnecessary use of batch_indices
* Avoid copying the data on squeeze and unsqueeze.
* Fix the quantized llama example.
* Unrelated fix for the quantized stable-lm example on cuda.
* Fix for mamba on cuda (unrelated to the PR).
* Add a --seed argument to the stable-diffusion example.
* Make the case when no seed is specified, that it will not be set, but use the engine's default. This will make the CPU engine work again when no --seed is given, and will cause a bailout when a seed is there, as the engine does not currently support it.
---------
Co-authored-by: niklas <niklas@appli.se>
* Normalize loudness of the generated audio.
* Lints.
* One more lint.
* Avoid running the bs1770 tests.
* Another attempt at discarding doc comments.
* Also normalize the loudness in the encodec example.
* Add the metavoice transformer.
* Sketch the speaker-encoder module.
* Adding to the metavoice model.
* Start adding the metavoice example.
* Get some logits out.
* Load the second stage model.
* Get the second step to run.
* Tweak the example.
* Add encodec tilting.
* Glue the different bits together.
* Fix a shape issue.
* Use a constant.
* BPE tokenization.
* Fix the position index in metavoice.
* Add the metavoice transformer.
* Sketch the speaker-encoder module.
* Adding to the metavoice model.
* Start adding the metavoice example.
* Get some logits out.
* Load the second stage model.
* Get the second step to run.
* Tweak the example.
* Add encodec tilting.
* Glue the different bits together.
* Fix a shape issue.
* Use a constant.
* BPE tokenization.
* Add a warning.
* Encodec model.
* Fixes.
* Add the padding functions.
* Get the LSTM bit to work.
* Get the encodec model to generate some tokens (decoder only for now).
* Minor tweak.
* Minor tweak.
* and quantized rwkv v5 model
* Integrate the quantized rwkv model in the initial example.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* Boilerplate for the quantized cuda support.
* More basic cuda support.
* More cuda quantization (quantize on cpu for now).
* Add the dequantization bit.
* Start adding some dedicated cuda kernels from llama.cpp.
* Move the kernel code.
* Start interfacing with the kernel.
* Tweak the kernel launch params.
* Bugfix for quantized metal.
* Fix some clippy lints.
* Tweak the launch parameters.
* Tweak cuda basics to perform a quantized matmul.
* Perform the dequantization on the cpu + use cublas for matmul.
* Add the dequantization kernel.
* Test the qmatmul.
* More kernels.
* Matmul-vec kernel.
* Add a couple kernels.
* More dequantization kernels.
* Add the Gemma models.
* Add the gemma example.
* Adapt the RmsNorm.
* Get the 2b model to work.
* 7b support.
* Use the config head dim.
* Yet another fix.
* Make the matrixes contiguous.
* Also get the 7b model to work.
* And add to the readme.
* Start adding the RWKV model.
* More of the forward step.
* Handle rescaling.
* FeedForward.
* More work on RWKV.
* Better state tracking.
* Finish a first pass on forward.
* Fix the shape mismatches.
* Do not rescale in f32.
* Rename to rwkv-v5.
* Add the new models to the readme.
* feat: support microphone whisper streaming
* fix: cleanup print stmts and adjust how input is read
* fix: remove incorrect comment
* feat: split into new example and simplify
* fix: feature flag example file
* fix: fmt fixes
* feat: simplify and remove redundant files
* Sketch the mamba model for inference.
* Complete the forward pass.
* Add the mamba example.
* Optimize the selective-scan part.
* Fix a couple shape mismatches and get inference to work.
* Tweak the readmes.
* More readme tweaks.
* Use the repo config for trocr rather than hardcoding it + small tweaks.
* Add support for the printed models.
* Fail with an appropriate error message on missing position embeddings.
* Initial check-in for the qwen2 model.
* More qwen2 inference.
* Polish the qwen example.
* Fix the rope basis.
* Get the inference to work.
* Support different model sizes.
* Add the ChatGLM model.
* Rotary embeddings.
* Add to the forward pass.
* Add to the forward pass.
* Add the rotary embeddings.
* Add the KV cache.
* Add the chatglm example.
* Bugfix.
* More glm fixes.
* Fix some shape issues.
* Get the inference to work.
* Supports more audio formats
* Simplify the handling of the different buffer types.
* Check the sample rate.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* Metal quantized modifications proposal.
- Add a device param, wherever needed.
- Create new QMetal storage thing that implements QuantizedType.
- Update everywhere needed.
Fix Python.
Fixing examples.
Fix: fmt + clippy + stub.
Moving everything around.
Only missing the actual implems.
Fixing everything + adding dequantized kernels.
More work.
Fixing matmul.
Fmt + Clippy
Some clippy fixes.
Working state.
Q2K Metal -> Bugged (also present in GGML).
Q4K CPU -> Bugged (present previously, new test catch it).
Q5K CPU -> Bugged (present previously).
Q8_1 Both -> Never really implemented it seems
Q8K metal -> Never implemented in metal
Fixing Q2K bug (present in ggml).
* Cleanup.
* Fix the rebase.
* Removing the fences speeds everything up and *is* correct this time...
* Cleanup the fence.
* After rebase.
* Bad code removal.
* Rebase after phi2 merge + fix replit default to CPU.
* Making the CI happy.
* More happy tests.
---------
Co-authored-by: Nicolas Patry <nicolas@Nicolass-MacBook-Pro.local>
* Update the Phi model to use the updated architecture.
* Add more of the phi model.
* Repeat KV + caching.
* Apply the rotary embeddings.
* Add support for the new phi model in the phi example.
* Fix a couple glitches.
* Fix a couple more glitches.
* Use cfg to seperate benchmark results based on features
* Add metal where_cond for f16 and bf16. Add benchmark
* Remove allow pragma
* Avoid some unnecessary returns.
* Improve benchmarks layout
* Updated feature separated benchmarks
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* Add RepVGG model.
* Add RepVGG README
* Extract var to top level
* Replace hashmap with a match
* Add a variant for the model kind + avoid some unnecessary config cloning.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* added policy_gradient, modified main, ddpg and README
* fixed typo in README
* removed unnecessary imports
* small refactor
* Use clap for picking up the subcommand to run.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
* Add the Mixtral model.
* Add more of the mixtral layers.
* Add the final layers for mixtral.
* Sketch the expert selection.
* Add some expert routing logic.
* Hopefully finish the routing logic for mixtral.
* Add the mixtral example.
* Fix the weight filenames.
* Bugfix.
* Another fix.
* Yet another fix + remove the unused pragma.
* Shape fix.
* Support for quantized mixtral.
* Support mixtral in the quantized example.
* Mlp or moe type.
* Fix the expert field namings.
* Refactor the mlp bit.
* More MoE logic.
* Add the MoE quantized logic.
* Fix the experts length.
* Add the Mixtral model.
* Add more of the mixtral layers.
* Add the final layers for mixtral.
* Sketch the expert selection.
* Add some expert routing logic.
* Hopefully finish the routing logic for mixtral.
* Add the mixtral example.
* Fix the weight filenames.
* Bugfix.
* Another fix.
* Yet another fix + remove the unused pragma.
* Shape fix.
* Add a readme.
* Add support for SD Turbo
* Set Leading as default in euler_ancestral discrete
* Use the appropriate default values for n_steps and guidance_scale.
---------
Co-authored-by: Laurent <laurent.mazare@gmail.com>
Few fixes.
Going back on remote metal-rs.
Reusing a single buffer (for now) to speed things up.
Adding some half kernels.
All tests are panicking instead of random failure.
Putting back f16 index select.
Add erf.
Working version for llama2-c.
Fixes + cache compute_pipeline_state.
BF16 metal fix.
Remove some prints.
new_owned -> new()..to_owned().
Better batched matmul.
Metal operational.
Reuse buffers on our own reference counts.
Tmp gemm.
Revert "Tmp gemm."
This reverts commit c65f68e988.
Interleave committing.
Speeding up copies using blit.
Fmt.
Fmt.
Remove the assert!
Fmt all.
Fixes after big rebase.
Add softmax for half and bfloat + tests
Fixing Llama example + accumulate softmax in float.
* add bce with logit loss
* add bce with logit loss
* remove imports
* fix tiny bug
* add test documentation and refactor function
* fix test cases and formatting
* distilbet files
* Apply various cleanups.
* More cleanups.
* More polish.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
* Fix linspace implementation
`steps` should be strictly greater than 1 to make it consistent with the context.
* Handle steps == 0 and steps == 1.
* Fix rustfmt.
---------
Co-authored-by: laurent <laurent.mazare@gmail.com>
Laurent Mazare