nexaai 1.0.19rc6__cp310-cp310-macosx_14_0_universal2.whl → 1.0.19rc7__cp310-cp310-macosx_14_0_universal2.whl

nexaai/binds/nexa_mlx/py-lib/vlm/generate.py

@@ -0,0 +1,572 @@
+import argparse
+import codecs
+import contextlib
+import functools
+import os
+import re
+import time
+from dataclasses import dataclass
+from typing import Any, Dict, Generator, List, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx_lm.generate import maybe_quantize_kv_cache
+from transformers import PreTrainedTokenizer
+
+from .modeling.models import cache
+from .modeling.prompt_utils import apply_chat_template
+from .modeling.sample_utils import top_p_sampling
+from .modeling.utils import (
+    StoppingCriteria,
+    apply_repetition_penalty,
+    load,
+    prepare_inputs,
+    tree_reduce,
+)
+
+DEFAULT_MODEL_PATH = "mlx-community/gemma-3-4b-it-8bit"
+
+
+def parse_media_from_input(user_input):
+    """Parse quoted media files from user input and return prompt and media paths"""
+    # Find all quoted strings (both single and double quotes)
+    quoted_pattern = r'["\']([^"\']*)["\']'
+    quoted_matches = re.findall(quoted_pattern, user_input)
+    
+    # Remove quoted strings from the input to get the actual prompt
+    prompt = re.sub(quoted_pattern, '', user_input).strip()
+    
+    # Separate image and audio files based on extensions
+    image_extensions = {'.png', '.jpg', '.jpeg', '.gif', '.bmp', '.tiff', '.webp'}
+    audio_extensions = {'.mp3', '.wav', '.flac', '.aac', '.ogg', '.m4a'}
+    
+    image_paths = []
+    audio_paths = []
+    
+    for quoted_file in quoted_matches:
+        if quoted_file:  # Skip empty quotes
+            # Expand user path if it starts with ~
+            if quoted_file.startswith('~'):
+                quoted_file = os.path.expanduser(quoted_file)
+            
+            # Check if file exists
+            if not os.path.exists(quoted_file):
+                print(f"Warning: File '{quoted_file}' not found")
+                continue
+                
+            file_ext = os.path.splitext(quoted_file.lower())[1]
+            if file_ext in image_extensions:
+                image_paths.append(quoted_file)
+            elif file_ext in audio_extensions:
+                audio_paths.append(quoted_file)
+    
+    return prompt, image_paths if image_paths else None, audio_paths if audio_paths else None
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description="Generate text from an image using a model."
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default=DEFAULT_MODEL_PATH,
+        help="The path to the local model directory or Hugging Face repo.",
+    )
+    return parser.parse_args()
+
+
+# A stream on the default device just for generation
+generation_stream = mx.new_stream(mx.default_device())
+
+
+@contextlib.contextmanager
+def wired_limit(model: nn.Module, streams: Optional[List[mx.Stream]] = None):
+    """
+    A context manager to temporarily change the wired limit.
+
+    Note, the wired limit should not be changed during an async eval.  If an
+    async eval could be running pass in the streams to synchronize with prior
+    to exiting the context manager.
+    """
+    model_bytes = tree_reduce(
+        lambda acc, x: acc + x.nbytes if isinstance(x, mx.array) else acc, model, 0
+    )
+    max_rec_size = mx.metal.device_info()["max_recommended_working_set_size"]
+    if model_bytes > 0.9 * max_rec_size:
+        model_mb = model_bytes // 2**20
+        max_rec_mb = max_rec_size // 2**20
+        print(
+            f"[WARNING] Generating with a model that requires {model_mb} MB "
+            f"which is close to the maximum recommended size of {max_rec_mb} "
+            "MB. This can be slow. See the documentation for possible work-arounds: "
+            "https://github.com/ml-explore/mlx-lm/tree/main#large-models"
+        )
+    old_limit = mx.set_wired_limit(max_rec_size)
+    try:
+        yield None
+    finally:
+        if streams is not None:
+            for s in streams:
+                mx.synchronize(s)
+        else:
+            mx.synchronize()
+        mx.set_wired_limit(old_limit)
+
+
+@dataclass
+class GenerationResult:
+    text: str
+    token: Optional[int]
+    logprobs: Optional[List[float]]
+    prompt_tokens: int
+    generation_tokens: int
+    prompt_tps: float
+    generation_tps: float
+    peak_memory: float
+
+
+def generate_step(
+    input_ids: mx.array,
+    model: nn.Module,
+    pixel_values,
+    mask,
+    *,
+    max_tokens: int = 256,
+    temperature: float = 0.0,
+    repetition_penalty: Optional[float] = None,
+    repetition_context_size: Optional[int] = 20,
+    top_p: float = 1.0,
+    logit_bias: Optional[Dict[int, float]] = None,
+    prompt_cache: Optional[List[Any]] = None,
+    max_kv_size: Optional[int] = None,
+    kv_bits: Optional[int] = None,
+    kv_group_size: int = 64,
+    quantized_kv_start: int = 0,
+    **kwargs,
+) -> Generator[Tuple[mx.array, mx.array], None, None]:
+    """
+    A generator producing token ids based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        temperature (float): The temperature for sampling, if 0 the argmax is used.
+          Default: ``0``.
+        repetition_penalty (float, optional): The penalty factor for repeating
+          tokens.
+        repetition_context_size (int, optional): The number of tokens to
+          consider for repetition penalty. Default: ``20``.
+        top_p (float, optional): Nulceus sampling, higher means model considers
+          more less likely words.
+        logit_bias (dictionary, optional): Additive logit bias.
+
+    Yields:
+        Generator[Tuple[mx.array, mx.array], None, None]: A generator producing
+          one token and a vector of log probabilities.
+    """
+
+    quantize_cache_fn = functools.partial(
+        maybe_quantize_kv_cache,
+        quantized_kv_start=quantized_kv_start,
+        kv_group_size=kv_group_size,
+        kv_bits=kv_bits,
+    )
+
+    def sample(logits: mx.array) -> Tuple[mx.array, float]:
+        if logit_bias:
+            indices = mx.array(list(logit_bias.keys()))
+            values = mx.array(list(logit_bias.values()))
+            logits[:, indices] += values
+        logprobs = logits - mx.logsumexp(logits)
+
+        if temperature == 0:
+            token = mx.argmax(logits, axis=-1)
+        else:
+            if top_p > 0 and top_p < 1.0:
+                token = top_p_sampling(logits, top_p, temperature)
+            else:
+                token = mx.random.categorical(logits * (1 / temperature))
+
+        return token, logprobs
+
+    if repetition_penalty and (
+        repetition_penalty < 0 or not isinstance(repetition_penalty, float)
+    ):
+        raise ValueError(
+            f"repetition_penalty must be a non-negative float, got {repetition_penalty}"
+        )
+
+    y = input_ids
+    # Create the KV cache for generation
+    if prompt_cache is None:
+        prompt_cache = cache.make_prompt_cache(
+            model.language_model,
+            max_kv_size=max_kv_size,
+        )
+
+    repetition_context = input_ids.reshape(-1).tolist()
+
+    if repetition_context_size:
+        repetition_context = repetition_context[-repetition_context_size:]
+
+    def _step(y, **kwargs):
+        with mx.stream(generation_stream):
+            nonlocal repetition_context
+            if "decoder_input_ids" in kwargs:
+                outputs = model.language_model(
+                    cache=prompt_cache,
+                    **kwargs,
+                )
+            else:
+                outputs = model.language_model(
+                    y[None],
+                    cache=prompt_cache,
+                    **kwargs,
+                )
+
+            logits = outputs.logits[:, -1, :]
+
+            if repetition_penalty:
+                logits = apply_repetition_penalty(
+                    logits, repetition_context, repetition_penalty
+                )
+                y, logprobs = sample(logits)
+                repetition_context.append(y.item())
+            else:
+                y, logprobs = sample(logits)
+
+            if repetition_context_size:
+                if len(repetition_context) > repetition_context_size:
+                    repetition_context = repetition_context[-repetition_context_size:]
+
+            quantize_cache_fn(prompt_cache)
+            return y, logprobs.squeeze(0)
+
+    outputs = model(input_ids, pixel_values, cache=prompt_cache, mask=mask, **kwargs)
+
+    logits = outputs.logits[:, -1, :]
+    quantize_cache_fn(prompt_cache)
+    y, logprobs = sample(logits)
+    mx.async_eval(y)
+
+    if outputs.cross_attention_states is not None:
+        kwargs = {
+            k: v
+            for k, v in zip(
+                ["cross_attention_states"], [outputs.cross_attention_states]
+            )
+        }
+    elif outputs.encoder_outputs is not None:
+        kwargs = {
+            "decoder_input_ids": y[None],
+            "encoder_outputs": outputs.encoder_outputs,
+        }
+    else:
+        kwargs = {}
+
+    n = 0
+    while True:
+        if n != max_tokens:
+            next_y, next_logprobs = _step(y, **kwargs)
+            mx.async_eval(next_y)
+            if "decoder_input_ids" in kwargs:
+                kwargs["decoder_input_ids"] = next_y[None]
+            yield y.item(), logprobs
+            y, logprobs = next_y, next_logprobs
+        if n == max_tokens:
+            break
+
+        n += 1
+
+        # Periodically clear cache to prevent memory accumulation
+        if n % 256 == 0:  # Clear cache every 256 tokens
+            mx.clear_cache()
+
+
+def stream_generate(
+    model: nn.Module,
+    processor: PreTrainedTokenizer,
+    prompt: str,
+    image: Union[str, List[str]] = None,
+    audio: Union[str, List[str]] = None,
+    **kwargs,
+) -> Union[str, Generator[str, None, None]]:
+    """
+    A generator producing text based on the given prompt from the model.
+
+    Args:
+        prompt (mx.array): The input prompt.
+        model (nn.Module): The model to use for generation.
+        max_tokens (int): The ma
+        kwargs: The remaining options get passed to :func:`generate_step`.
+          See :func:`generate_step` for more details.
+
+    Yields:
+        Generator[Tuple[mx.array, mx.array]]: A generator producing text.
+    """
+    tokenizer = processor.tokenizer if hasattr(processor, "tokenizer") else processor
+
+    # Skip special tokens
+    skip_special_tokens = kwargs.pop("skip_special_tokens", False)
+    skip_special_token_ids = (
+        set(tokenizer.all_special_ids)
+        if skip_special_tokens and hasattr(tokenizer, "all_special_ids")
+        else []
+    )
+
+    add_special_tokens = (
+        not hasattr(processor, "chat_template")
+        if model.config.model_type in ["gemma3", "gemma3n"]
+        else True
+    )
+
+    resize_shape = kwargs.pop("resize_shape", None)
+    image_token_index = getattr(model.config, "image_token_index", None)
+
+    if kwargs.get("input_ids", None) is not None:
+        input_ids = kwargs.pop("input_ids")
+        pixel_values = kwargs.pop("pixel_values", None)
+        mask = kwargs.pop("mask", None)
+    else:
+        inputs = prepare_inputs(
+            processor,
+            images=image,
+            audio=audio,
+            prompts=prompt,
+            image_token_index=image_token_index,
+            resize_shape=resize_shape,
+            add_special_tokens=add_special_tokens,
+        )
+        input_ids = inputs.get("input_ids", None)
+        pixel_values = inputs.get("pixel_values", None)
+        mask = inputs.get("attention_mask", None)
+        data_kwargs = {
+            k: v
+            for k, v in inputs.items()
+            if k not in ["input_ids", "pixel_values", "attention_mask"]
+        }
+        kwargs.update(data_kwargs)
+
+    with wired_limit(model, [generation_stream]):
+        detokenizer = processor.detokenizer
+        detokenizer.reset()
+        tic = time.perf_counter()
+        for n, (token, logprobs) in enumerate(
+            generate_step(input_ids, model, pixel_values, mask, **kwargs)
+        ):
+            if n == 0:
+                prompt_time = time.perf_counter() - tic
+                prompt_tps = input_ids.size / prompt_time
+                tic = time.perf_counter()
+
+            # Stop generation if the token is in the eos_token_ids
+            if tokenizer.stopping_criteria(token):
+                break
+
+            detokenizer.add_token(token, skip_special_token_ids=skip_special_token_ids)
+
+            # Yield the last segment if streaming
+            yield GenerationResult(
+                text=detokenizer.last_segment,
+                token=token,
+                logprobs=logprobs,
+                prompt_tokens=input_ids.size,
+                generation_tokens=n + 1,
+                prompt_tps=prompt_tps,
+                generation_tps=(n + 1) / (time.perf_counter() - tic),
+                peak_memory=mx.get_peak_memory() / 1e9,
+            )
+
+        detokenizer.finalize()
+        yield GenerationResult(
+            text=detokenizer.last_segment,
+            token=token,
+            logprobs=logprobs,
+            prompt_tokens=input_ids.size,
+            generation_tokens=n + 1,
+            prompt_tps=prompt_tps,
+            generation_tps=(n + 1) / (time.perf_counter() - tic),
+            peak_memory=mx.get_peak_memory() / 1e9,
+        )
+
+        # Cleanup after generation
+        mx.clear_cache()
+
+
+def generate(
+    model: nn.Module,
+    processor: PreTrainedTokenizer,
+    prompt: str,
+    image: Union[str, List[str]] = None,
+    audio: Union[str, List[str]] = None,
+    verbose: bool = False,
+    **kwargs,
+) -> str:
+    """
+    Generate text from the model.
+
+    Args:
+       model (nn.Module): The language model.
+       tokenizer (PreTrainedTokenizer): The tokenizer.
+       prompt (str): The string prompt.
+       temperature (float): The temperature for sampling (default 0).
+       max_tokens (int): The maximum number of tokens (default 100).
+       verbose (bool): If ``True``, print tokens and timing information
+           (default ``False``).
+       formatter (Optional[Callable]): A function which takes a token and a
+           probability and displays it.
+       repetition_penalty (float, optional): The penalty factor for repeating tokens.
+       repetition_context_size (int, optional): The number of tokens to consider for repetition penalty.
+    """
+
+    if verbose:
+        print("=" * 10)
+        files = []
+        if image is not None:
+            files.extend(image)
+        if audio is not None:
+            files.extend(audio)
+        if kwargs.get("video") is not None:
+            files.extend(kwargs.get("video"))
+
+        print(f"Files: {files}", "\n")
+
+        print("Prompt:", prompt)
+
+    text = ""
+    last_response = None
+
+    eos_tokens = kwargs.get("eos_tokens", None)
+    stopping_criteria = kwargs.get("stopping_criteria", None)
+
+    # Get the tokenizer
+    tokenizer = processor.tokenizer if hasattr(processor, "tokenizer") else processor
+
+    # Add custom EOS tokens to the stopping criteria
+    if eos_tokens is not None:
+        tokenizer.stopping_criteria.add_eos_token_ids(eos_tokens)
+
+    # Use custom stopping criteria
+    elif stopping_criteria is not None:
+        if isinstance(stopping_criteria, StoppingCriteria) or callable(
+            stopping_criteria
+        ):
+            tokenizer.stopping_criteria = stopping_criteria
+        else:
+            raise ValueError(
+                "stopping_criteria must be an instance of StoppingCriteria or a callable"
+            )
+    else:
+        tokenizer.stopping_criteria.reset(model.config.eos_token_id)
+
+    for response in stream_generate(model, processor, prompt, image, audio, **kwargs):
+        if verbose:
+            print(response.text, end="", flush=True)
+        text += response.text
+        last_response = response
+
+    if verbose:
+        print("\n" + "=" * 10)
+        if len(text) == 0:
+            print("No text generated for this prompt")
+            return
+        print(
+            f"Prompt: {last_response.prompt_tokens} tokens, "
+            f"{last_response.prompt_tps:.3f} tokens-per-sec"
+        )
+        print(
+            f"Generation: {last_response.generation_tokens} tokens, "
+            f"{last_response.generation_tps:.3f} tokens-per-sec"
+        )
+        print(f"Peak memory: {last_response.peak_memory:.3f} GB")
+
+    usage_stats = {
+        "input_tokens": last_response.prompt_tokens,
+        "output_tokens": last_response.generation_tokens,
+        "total_tokens": last_response.prompt_tokens + last_response.generation_tokens,
+        "prompt_tps": last_response.prompt_tps,
+        "generation_tps": last_response.generation_tps,
+        "peak_memory": last_response.peak_memory,
+    }
+
+    return text, usage_stats
+
+
+def main():
+    args = parse_arguments()
+    
+    # Load model and processor
+    model, processor = load(args.model, None)
+    config = model.config
+
+    # Initialize chat history
+    chat = []
+    
+    print("Multi-round conversation started. Type 'exit' or 'quit' to stop.")
+    print("You can include image/audio files in quotes, e.g.: 'what does this image mean \"/path/to/image.jpg\"'")
+    print("=" * 50)
+    
+    # Main chat loop
+    while True:
+        try:
+            user_input = input("User: ").strip()
+            
+            # Exit conditions
+            if user_input.lower() in ['exit', 'quit', '']:
+                break
+            
+            # Parse media files from user input
+            prompt_text, image_paths, audio_paths = parse_media_from_input(user_input)
+            
+            # If no text prompt after parsing, use the original input
+            if not prompt_text.strip():
+                prompt_text = user_input
+                image_paths = None
+                audio_paths = None
+            
+            # Add user message to chat history
+            chat.append({"role": "user", "content": prompt_text})
+            
+            # Calculate number of images for chat template
+            num_images = len(image_paths) if image_paths else 0
+            num_audios = len(audio_paths) if audio_paths else 0
+            
+            # Apply chat template
+            formatted_prompt = apply_chat_template(
+                processor, config, chat, num_images=num_images, num_audios=num_audios
+            )
+            
+            # Generate response
+            response = ""
+            print("Assistant: ", end="", flush=True)
+            
+            for chunk in stream_generate(
+                model,
+                processor,
+                formatted_prompt,
+                image_paths,
+                audio_paths,
+                max_tokens=100,
+                temperature=0.7,
+                top_p=0.9,
+                verbose=True,
+            ):
+                response += chunk.text
+                print(chunk.text, end="", flush=True)
+            
+            print()  # New line after response
+            
+            # Add assistant response to chat history
+            chat.append({"role": "assistant", "content": response})
+            
+        except KeyboardInterrupt:
+            print("\nConversation interrupted by user.")
+            break
+        except Exception as e:
+            print(f"Error: {e}")
+            continue
+
+
+if __name__ == "__main__":
+    main()