David Koski
ac273a14ea
fix float types in Phi (use float16) ( #25 )
...
- per suggestions in #23 ensure that the values that go into the cache are float16
2024-03-14 13:18:40 -07:00
David Koski
0fb74cbfdc
adopt MLXFast.scaledDotProductAttention ( #23 )
2024-03-12 14:04:43 -07:00
John Mai
a94bf79d7e
feat: Support Starcoder2 ( #20 )
...
* feat: Support Starcoder2
2024-03-07 21:28:37 -08:00
Madroid Ma
e876e18605
update qwen2 chat template ( #18 )
2024-03-07 07:51:54 -08:00
David Koski
dfc9f2fc01
apply swift-format
2024-03-03 18:40:49 -08:00
John Mai
66d9202360
feat: Qwen2 support
2024-03-03 22:26:28 +08:00
David Koski
7b746cb89c
allow alternate location for tokenizer
2024-03-01 23:27:03 -08:00
David Koski
ff7a615db7
improve phi prompt -- partial fix for #9
2024-03-01 22:45:01 -08:00
David Koski
c49dd73c28
swift-format, circleci setup
2024-03-01 16:10:34 -08:00
David Koski
b41f14fba7
add LLM evaluator example
...
- runs on iOS and macOS
- downloads a model / tokenizer from hugging face
- evaluates the given prompt
2024-03-01 16:10:00 -08:00
David Koski
2157333905
swift-format!
2024-03-01 14:47:43 -08:00
David Koski
82f6a969d4
llm improvements
...
- document the tokenizer used (https://github.com/huggingface/swift-transformers )
- provide a hook for tokenizer configuration, prompt augmentation
- this isn't as rich as the python equivalents but it helps a little
2024-03-01 14:46:32 -08:00
David Koski
3f02fcc1cb
expose eosToken
2024-02-26 14:58:51 -08:00
David Koski
4fad86d84b
split tokenizer code out into new file
2024-02-26 14:42:40 -08:00
David Koski
c7919cf7fe
fix rmsnorm for gemma
2024-02-26 14:09:48 -08:00
David Koski
a2ff291608
add reference to filed issue
2024-02-26 13:31:33 -08:00
David Koski
c86d1c195e
partial fix for #1
...
- handle loading models with different names for the safetensors files (gemma)
- handle merge tokens that can't be split
- organize code into Load/Evaluate
2024-02-26 13:23:21 -08:00
David Koski
bb7bacc077
fix for #2 -- CodeLlama crashes
...
- add replacement tokenizer class for unknown tokenizers
- fix quantization for models that don't have lm_head quantized
Requires https://github.com/ml-explore/mlx-swift/pull/28
2024-02-26 10:38:05 -08:00
David Koski
5a83d7d92b
fix broken links, clarify documentation
2024-02-22 12:46:44 -08:00
David Koski
b6d1e14465
initial commit
2024-02-22 10:41:02 -08:00
