Explore os LLMs (grandes modelos de linguagem) no Apple Silicon com o MLX

1:12 - Running DeepSeek AI's model with MLX LM

mlx_lm.chat --model mlx-community/DeepSeek-V3-0324-4bit

3:51 - Text generation with MLX LM

mlx_lm.generate --model "mlx-community/Mistral-7B-Instruct-v0.3-4bit" \
                --prompt "Write a quick sort in Swift"

4:35 - Changing the model's behavior with flags

mlx_lm.generate --model "mlx-community/Mistral-7B-Instruct-v0.3-4bit" \
                --prompt "Write a quick sort in Swift" \
                --top-p 0.5 \
                --temp 0.2 \
                --max-tokens 1024

4:48 - Getting help for MLX LM

mlx_lm.generate --help

5:26 - MLX LM Python API

# Using MLX LM from Python

from mlx_lm import load, generate

# Load the model and tokenizer directly from HF
model, tokenizer = load("mlx-community/Mistral-7B-Instruct-v0.3-4bit")

# Prepare the prompt for the model
prompt = "Write a quick sort in Swift"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

# Generate the text
text = generate(model, tokenizer, prompt=prompt, verbose=True)

6:24 - Inspecting model architecture

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Mistral-7B-Instruct-v0.3-4bit")

print(model)
print(model.parameters())
print(model.layers[0].self_attn)

8:01 - Generation with KV cache

from mlx_lm import load, generate
from mlx_lm.models.cache import make_prompt_cache

# Load the model and tokenizer directly from HF
model, tokenizer = load("mlx-community/Mistral-7B-Instruct-v0.3-4bit")

# Prepare the prompt for the model
prompt = "Write a quick sort in Swift"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

cache = make_prompt_cache(model)

# Generate the text
text = generate(model, tokenizer, prompt=prompt, prompt_cache=cache, verbose=True)

9:37 - Quantization

mlx_lm.convert --hf-path "mistralai/Mistral-7B-Instruct-v0.3" \
               --mlx-path "./mistral-7b-v0.3-4bit" \
               --dtype float16 \
               --quantize --q-bits 4 --q-group-size 64

10:33 - Model quantization with MLX LM in Python

from mlx_lm.convert import convert

# We can choose a different quantization per layer
def mixed_quantization(layer_path, layer, model_config):
    if "lm_head" in layer_path or "embed_tokens" in layer_path:
        return {"bits": 6, "group_size": 64}
    elif hasattr(layer, "to_quantized"):
        return {"bits": 4, "group_size": 64}
    else:
        return False

# Convert can be used to change precision, quantize and upload models to HF
convert(
    hf_path="mistralai/Mistral-7B-Instruct-v0.3",
    mlx_path="./mistral-7b-v0.3-mixed-4-6-bit",
    quantize=True,
    quant_predicate=mixed_quantization
)

13:37 - Model fine-tuning

mlx_lm.lora --model "mlx-community/Mistral-7B-Instruct-v0.3-4bit" 
						--train 
            --data /path/to/our/data/folder
            --iters 300 
            --batch-size 16

15:06 - Prompting before fine-tuning

mlx_lm.generate --model "./mistral-7b-v0.3-4bit" \
    --prompt "Who won the latest super bowl?"

15:34 - Fine-tuning to learn new knowledge

mlx_lm.lora --model "./mistral-7b-v0.3-4bit" 
						--train 
            --data ./data 
            --iters 300 
            --batch-size 8 
            --mask-prompt 
            --learning-rate 1e-5

15:48 - Prompting after fine-tuning

mlx_lm.generate --model "mlx-community/Mistral-7B-Instruct-v0.3-4bit" \
                --prompt "Who won the latest super bowl?" \
                --adapter "adapters"

16:29 - Fusing models

mlx_lm.fuse --model "mlx-community/Mistral-7B-Instruct-v0.3-4bit"
            --adapter-path "path/to/trained/adapters" \
            --save-path "fused-mistral-7b-v0.3-4bit" \
            --upload-repo "my-name/fused-mistral-7b-v0.3-4bit"
            
# Fusing our fine-tuned model adapters
mlx_lm.fuse --model "./mistral-7b-v0.3-4bit" \
            --adapter-path "adapters" \
            --save-path "fused-mistral-7b-v0.3-4bit"

17:14 - LLMs in MLX Swift

import Foundation
import MLX
import MLXLMCommon
import MLXLLM

@main
struct LLM {
    static func main() async throws {
        // Load the model and tokenizer directly from HF
        let modelId = "mlx-community/Mistral-7B-Instruct-v0.3-4bit"
        let modelFactory = LLMModelFactory.shared
        let configuration = ModelConfiguration(id: modelId)
        let model = try await modelFactory.loadContainer(configuration: configuration)
        
        try await model.perform({context in
            // Prepare the prompt for the model
            let prompt = "Write a quicksort in Swift"
            let input = try await context.processor.prepare(input: UserInput(prompt: prompt))
            
            // Generate the text
            let params = GenerateParameters(temperature: 0.0)
            let tokenStream = try generate(input: input, parameters: params, context: context)
            for await part in tokenStream {
                print(part.chunk ?? "", terminator: "")
            }
        })
    }
}

18:00 - Generation with KV cache in MLX Swift

import Foundation
import MLX
import MLXLMCommon
import MLXLLM

@main
struct LLM {
    static func main() async throws {
        // Load the model and tokenizer directly from HF
        let modelId = "mlx-community/Mistral-7B-Instruct-v0.3-4bit"
        let modelFactory = LLMModelFactory.shared
        let configuration = ModelConfiguration(id: modelId)
        let model = try await modelFactory.loadContainer(configuration: configuration)
        
        try await model.perform({context in
            // Prepare the prompt for the model
            let prompt = "Write a quicksort in Swift"
            let input = try await context.processor.prepare(input: UserInput(prompt: prompt))

            // Create the key-value cache
            let generateParameters = GenerateParameters()
            let cache = context.model.newCache(parameters: generateParameters)

            // Low level token iterator
            let tokenIter = try TokenIterator(input: input,
                                              model: context.model,
                                              cache: cache,
                                              parameters: generateParameters)
            let tokenStream = generate(input: input, context: context, iterator: tokenIter)
            for await part in tokenStream {
                print(part.chunk ?? "", terminator: "")
            }
        })
    }
}