Add phi-3 checkpoint

[rasbt]

• Verify Phi-3-mini-4k-instruct configs
• Add prompt style
• Add other config files
• Add test_model.py
• Add to test_prompts.py
• Update 2 tables in README
• Update download_model_weights.md

[Andrei-Aksionov]

There is a modeling_*.py file.
Good luck 🙂.

[rasbt]

There is a modeling_*.py file.
Good luck 🙂.

Haha, I finally get the weights loaded but of course it's never easy ... of course it's generating gibberish

⚡ phi-3-checkpoint ~/litgpt litgpt chat --checkpoint_dir checkpoints/microsoft/Phi-3-mini-4k-instruct
Now chatting with Phi-3-mini-4k-instruct.
To exit, press 'Enter' on an empty prompt.

Seed set to 1234
>> Prompt: What do llamas eat?
>> Reply: epsonniformes }).selves }).SSIONunicívo }). EverythingFormsћassaiejalphutureievediennesenticaciónicaciónMilMinigh ninassaselvesselves exhaustselvesonnselvesktionΗracheracheionedΗ Avenoted Bij_+versionsmastevosepsselvesmobileselvesilleryassaucealphasseestoreselvesférFormsiej Mu Kaiser oppienngnatteversionsionedionedversionsSSIONectionaccoossFormassaselves_+uminatesonoSSIONológissancecenteecause_+ienn选uraleʋ Stepalphigosionaliilonverte }).ienn }).ativo Sternsonoiejuralassawnkademselves│uraleativaionedvos_+utschversionsponiej_+icacióniejiewerológvoasonverte shoutioned位ionedIdentmobi

Let the easter egg hunt begin 😭

[rasbt]

Some more tidbits via Daniel Han:

Phi 3 (3.8B) got released! The paper said it was just a Llama arch, but I found some quirks while adding this to
@UnslothAI
:

1. Sliding window of 2047? Mistral v1 4096. So does Phi mini have SWA? (And odd num?) Max RoPE position is 4096?
2. Upcasted RoPE? Like Gemma?
3. Dynamic RoPE for 128K context lengths
4. Fused MLP & QKV - need to unfuse
5. MMLU evals are very different betw the Phi team Llama-3 team - why?

[Andrei-Aksionov]

Ok, it's becoming more interesting.
Somewhat I expected from LlaMA 3, but it didn't deliver.

[carmocca]

It should be possible to not need this class at all and instead reshape the weights for LLaMAMLP in the checkpoint conversion.

[rasbt]

100% agree. I was thinking the same thing. Similar to OLMo, I was hoping to get it working first and then simplify from there.

[Andrei-Aksionov]

New models by Apple have if-else statements for this case: https://huggingface.co/apple/OpenELM-270M-Instruct/blob/main/modeling_openelm.py#L405-L462

For simplicity we definitely shouldn't do the same.

[rasbt]

Looks like the sliding window number was a typo: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/commit/b043e05a86cfc77f8d53eb0edf6a33e39afbcb5e

[Andrei-Aksionov]

Current code is an ugly state, but at least the model produces the same output as HF one.
The most notable change is that now Phi3 model doesn't use parallel_residual in contrast to Phi1.5 and Phi2.

The missing piece is the Tokenizer: it has a smaller vocab size (32k vs 50k) that was extended by 64 special tokens.
If I'm not mistaken, the current code doesn't add these tokens.

[rasbt]

The missing piece is the Tokenizer: it has a smaller vocab size (32k vs 50k) that was extended by 64 special tokens.
If I'm not mistaken, the current code doesn't add these tokens.

Yeah, that sounds about right based on the Phi-3 paper:

To best benefit the open source community, phi-3-mini is built upon a similar block structure as Llama-2 [TLI+23] and uses the same tokenizer with vocabulary size of 320641