[docs] deploying

[stevhliu]

adds ecosystem integration docs for deploying with Candle, ExecuTorch, and MLX

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[stevhliu]

@ivarflakstad, would you mind reviewing the candle integration doc please? the goal is to explain and demonstrate how candle uses Transformers

[ivarflakstad]

Looks good to me!
The docs (https://huggingface.github.io/candle) are very outdated at this point though.
Much of the information is still correct, but still I'd rather we directed users to the candle readme.

[stevhliu]

@pcuenca, would you mind reviewing the MLX integration doc please? the goal is to explain and demonstrate how MLX uses Transformers

[pcuenca]

Took a quick look at the MLX section, made a few comments and suggested to add the MLX -> transformers integration, but found some problems while testing; will take a deeper look.

I'll review the rest of the sections later.

[pcuenca]

Suggested change

	Conversely, you can also load and run MLX-converted weights in Transformers, potentially on different platforms:
	```py
	from transformers import AutoModelForCausalLM, AutoTokenizer
	model_id = "pcuenq/tiny-llama-chat-mlx"
	tokenizer = AutoTokenizer.from_pretrained(model_id)
	model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda:0")
	messages = [
	{"role": "user", "content": "What is the capital of France?"},
	]
	inputs = tokenizer(prompt, return_tensors="pt")
	outputs = model.generate(inputs["input_ids"].to(model.device), do_sample=False, max_new_tokens=100)
	print(tokenizer.decode(outputs[0].to("cpu")))

Heads up: we may want to skip this for now as I'm finding friction finding checkpoints that work - incompatible quants or weight shapes. Will look into it a bit.

[pcuenca]

GitHub got confused with the nested quotes.

[stevhliu]

hey @pcuenca, lmk if its ok to skip this example for now so we can merge this. happy to follow up on this in a future PR :)

[pcuenca]

Hi @stevhliu, sorry I dropped the ball here! Yes, let's skip for now and get this out!

[LysandreJik]

Cool thank you! cc @ArthurZucker on the Executorch part