generate.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import sys
import time
from pathlib import Path
from typing import Optional, Tuple
import itertools
import torch
import torch._dynamo
torch._dynamo.config.suppress_errors = True

import torch._inductor.config
import torch._dynamo.config
torch._inductor.config.coordinate_descent_tuning = True
torch._inductor.config.triton.unique_kernel_names = True

# support running without installing as a package
wd = Path(__file__).parent.parent.resolve()
sys.path.append(str(wd))

from model import Transformer
from sentencepiece import SentencePieceProcessor
from qwen_tokenizer import QwenTokenizer

def multinomial_sample_one_no_sync(probs_sort): # Does multinomial sampling without a cuda synchronization
    q = torch.empty_like(probs_sort).exponential_(1)
    return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)

def logits_to_probs(logits, temperature: float = 1.0, top_k: Optional[int] = None):
    logits = logits / max(temperature, 1e-5)

    if top_k is not None:
        v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
        pivot = v.select(-1, -1).unsqueeze(-1)
        logits = torch.where(logits < pivot, -float("Inf"), logits)
    probs = torch.nn.functional.softmax(logits, dim=-1)
    return probs

def sample(logits, temperature: float = 1.0, top_k: Optional[int] = None):
    probs = logits_to_probs(logits[0, -1], temperature, top_k)
    idx_next = multinomial_sample_one_no_sync(probs)
    return idx_next, probs

def prefill(model: Transformer, x: torch.Tensor, input_pos: torch.Tensor, **sampling_kwargs) -> torch.Tensor:
    # input_pos: [B, S]
    logits = model(x, input_pos)
    return sample(logits, **sampling_kwargs)[0]

def decode_one_token(model: Transformer, x: torch.Tensor, input_pos: torch.Tensor, **sampling_kwargs) -> Tuple[torch.Tensor, torch.Tensor]:
    # input_pos: [B, 1]
    assert input_pos.shape[-1] == 1
    logits = model(x, input_pos)
    return sample(logits, **sampling_kwargs)

def decode_n_tokens(model: Transformer, cur_token: torch.Tensor, input_pos: torch.Tensor, num_new_tokens: int, callback=lambda _: _, **sampling_kwargs):
    new_tokens, new_probs = [], []
    for i in range(num_new_tokens):
        with torch.backends.cuda.sdp_kernel(enable_flash=False, enable_mem_efficient=False, enable_math=True): # Actually better for Inductor to codegen attention here
            next_token, next_prob = decode_one_token(
                model, cur_token, input_pos, **sampling_kwargs
            )
        input_pos += 1
        new_tokens.append(next_token.clone())
        callback(new_tokens[-1])
        new_probs.append(next_prob.clone())
        cur_token = next_token.view(1, -1)
    return new_tokens, new_probs


def model_forward(model, x, input_pos):
    return model(x, input_pos)

@torch.no_grad()
def generate(
    model: Transformer,
    prompt: torch.Tensor,
    max_new_tokens: int,
    *,
    interactive: bool,
    callback = lambda x: x,
    **sampling_kwargs
) -> torch.Tensor:
    """
    Takes a conditioning sequence (prompt) as input and continues to generate as many tokens as requested.
    """

    # create an empty tensor of the expected final shape and fill in the current tokens
    T = prompt.size(0)
    T_new = T + max_new_tokens
    if interactive:
        max_seq_length = 350
    else:
        max_seq_length = min(T_new, model.config.block_size)

    device, dtype = prompt.device, prompt.dtype
    with torch.device(device):
        model.setup_caches(max_batch_size=1, max_seq_length=max_seq_length)

    # create an empty tensor of the expected final shape and fill in the current tokens
    empty = torch.empty(T_new, dtype=dtype, device=device)
    empty[:T] = prompt
    seq = empty
    input_pos = torch.arange(0, T, device=device)

    next_token = prefill(model, prompt.view(1, -1), input_pos, **sampling_kwargs)
    seq[T] = next_token
    callback(next_token)

    input_pos = torch.tensor([T], device=device, dtype=torch.int)

    generated_tokens, _ = decode_n_tokens(model, next_token.view(1, -1), input_pos, max_new_tokens - 1, callback=callback, **sampling_kwargs)
    seq[T + 1:] = torch.cat(generated_tokens)

    return seq

def encode_tokens(tokenizer, string, bos=True, device='cuda'):
    tokens = tokenizer.encode(string)
    if bos:
        tokens = [tokenizer.bos_id()] + tokens
    return torch.tensor(tokens, dtype=torch.int, device=device)

def _load_model(checkpoint_path, device, precision):
    with torch.device('meta'):
        model = Transformer.from_name(checkpoint_path.parent.name)

    if "int2" in str(checkpoint_path):
        print("Using quip quantization!")
        from quantize import WeightOnlyQuipQuantHandler
        use_rand = ("rand" in str(checkpoint_path))
        simple_quantizer = WeightOnlyQuipQuantHandler(model, use_rand)
        model = simple_quantizer.convert_for_runtime()

    checkpoint = torch.load(str(checkpoint_path), mmap=True, weights_only=True)
    model.load_state_dict(checkpoint, strict=True, assign=True)

    model = model.to(device=device, dtype=precision)
    return model.eval()

B_INST, E_INST = "[INST]", "[/INST]"

def main(
    prompt: str = "Hello, my name is",
    interactive: bool = False,
    num_samples: int = 5,
    max_new_tokens: int = 100,
    top_k: int = 200,
    temperature: float = 0.8,
    checkpoint_path: Path = Path("checkpoints/meta-Transformer/Transformer-2-7b-chat-hf/model.pth"),
    compile: bool = True,
    compile_prefill: bool = False,
    profile: Optional[Path] = None,
) -> None:
    """Generates text samples based on a pre-trained Transformer model and tokenizer.
    """
    assert checkpoint_path.is_file(), checkpoint_path

    device = 'cuda'
    precision = torch.float16
    is_chat = "chat" in str(checkpoint_path).lower()
    is_qwen = "qwen" in str(checkpoint_path).lower()

    print("Loading model ...")
    t0 = time.time()
    model = _load_model(checkpoint_path, device, precision)

    torch.cuda.synchronize()
    print(f"Time to load model: {time.time() - t0:.02f} seconds")

    tokenizer_path = checkpoint_path.parent / "tokenizer.model"
    # llama
    if not is_qwen:
        tokenizer_path = checkpoint_path.parent / "tokenizer.model"
        tokenizer = SentencePieceProcessor(model_file=str(tokenizer_path))
        bos = True
    # qwen
    else:
        tokenizer_path = checkpoint_path.parent / "qwen.tiktoken"
        tokenizer = QwenTokenizer(str(tokenizer_path))
        bos = False
    encoded = encode_tokens(tokenizer, prompt, bos=bos, device=device)
    prompt_length = encoded.size(0)

    torch.manual_seed(1234)
    model_size = sum([p.numel() * p.dtype.itemsize for p in itertools.chain(model.parameters(), model.buffers())])
    if compile:
        global decode_one_token, prefill
        decode_one_token = torch.compile(decode_one_token, mode="reduce-overhead", fullgraph=False)

        # Uncomment to squeeze more perf out of prefill
        if args.compile_prefill:
            prefill = torch.compile(prefill, fullgraph=False, dynamic=True)


    aggregate_metrics = {
        'tokens_per_sec': [],
    }
    start = -1 if compile else 0

    for i in range(start, num_samples):
        torch.cuda.synchronize()
        if i >= 0 and interactive:
            prompt = input("What is your prompt? ")
            if is_chat:
                if is_qwen:
                    prompt = f"<|im_start|>user\n{prompt.strip()}<|im_end|>\n<|im_start|>assistant\n"
                else:
                    prompt = f"{B_INST} {prompt.strip()} {E_INST}"
            encoded = encode_tokens(tokenizer, prompt, bos=(not is_qwen), device=device)

        if interactive and i >= 0:
            buffer = []
            period_id = tokenizer.encode('.')[0]
            done_generating = False
            def callback(x):
                nonlocal done_generating
                if done_generating:
                    return
                buffer.append(tokenizer.decode([period_id] + x.tolist())[1:])
                if x.item() == tokenizer.eos_id():
                    done_generating = True
                if len(buffer) == 4 or done_generating:
                    print(''.join(buffer), end='', flush=True)
                    buffer.clear()
                # print(, end='', flush=True)
        else:
            callback = lambda x : x
        t0 = time.perf_counter()
        import contextlib
        if i != num_samples - 1 or not profile:
            prof = contextlib.nullcontext()
        else:
            torch.profiler._utils._init_for_cuda_graphs()
            prof = torch.profiler.profile()
        with prof:
            y = generate(
                model,
                encoded,
                max_new_tokens,
                interactive=interactive,
                callback=callback,
                temperature=temperature,
                top_k=top_k,
            )
        if i == -1:
            print(f"Compilation time: {time.perf_counter() - t0:.2f} seconds")
            continue
        if hasattr(prof, "export_chrome_trace"):
            prof.export_chrome_trace(f"{profile}.json")
        torch.cuda.synchronize()
        t = time.perf_counter() - t0

        if not interactive:
            print(tokenizer.decode(y.tolist()))
        else:
            print()
        tokens_generated = y.size(0) - prompt_length
        tokens_sec = tokens_generated / t
        aggregate_metrics['tokens_per_sec'].append(tokens_sec)
        print(f"Time for inference {i + 1}: {t:.02f} sec total, {tokens_sec:.02f} tokens/sec")
        print(f"Bandwidth achieved: {model_size * tokens_sec / 1e9:.02f} GB/s")
    print("==========")
    print(f"Average tokens/sec: {torch.mean(torch.tensor(aggregate_metrics['tokens_per_sec'])).item():.2f}")
    print(f"Memory used: {torch.cuda.max_memory_reserved() / 1e9:.02f} GB")


if __name__ == '__main__':
    import argparse
    parser = argparse.ArgumentParser(description='Your CLI description.')

    parser.add_argument('--prompt', type=str, default="Hello, my name is", help='Input prompt.')
    parser.add_argument('--interactive', action='store_true', help='Whether to launch in interactive mode')
    parser.add_argument('--num_samples', type=int, default=5, help='Number of samples.')
    parser.add_argument('--max_new_tokens', type=int, default=128, help='Maximum number of new tokens.')
    parser.add_argument('--top_k', type=int, default=200, help='Top-k for sampling.')
    parser.add_argument('--temperature', type=float, default=1, help='Temperature for sampling.')
    parser.add_argument('--checkpoint_path', type=Path, default=Path("checkpoints/meta-Transformer/Transformer-2-7b-chat-hf/model.pth"), help='Model checkpoint path.')
    parser.add_argument('--compile', action='store_true', help='Whether to compile the model.')
    parser.add_argument('--compile_prefill', action='store_true', help='Whether to compile the prefill (improves prefill perf, but higher compile times)')
    parser.add_argument('--profile', type=Path, default=None, help='Profile path.')

    args = parser.parse_args()
    main(
        args.prompt, args.interactive, args.num_samples, args.max_new_tokens, args.top_k,
        args.temperature, args.checkpoint_path, args.compile, args.compile_prefill, args.profile,
    )