tinygrad/examples/mamba.py

import os, sys, math, argparse, time
sys.path.append(os.getcwd())
from typing import Any, Optional, Dict
from dataclasses import dataclass, field

from tinygrad import Tensor, TinyJit, nn
from tinygrad.helpers import fetch
from tinygrad.nn.state import load_state_dict, torch_load
from extra.models.llama import RMSNorm

from tqdm import tqdm
from transformers import AutoTokenizer

MODELS = {
  "130m": {"dim":  768, "n_layers": 24, "vocab_size": 50277, "pad_vocab_size_multiple": 8},
  "370m": {"dim": 1024, "n_layers": 48, "vocab_size": 50277, "pad_vocab_size_multiple": 8},
  "790m": {"dim": 1536, "n_layers": 48, "vocab_size": 50277, "pad_vocab_size_multiple": 8},
  "1.4b": {"dim": 2048, "n_layers": 48, "vocab_size": 50277, "pad_vocab_size_multiple": 8},
  "2.8b": {"dim": 2560, "n_layers": 64, "vocab_size": 50277, "pad_vocab_size_multiple": 8},
}

def fetch_weights(model_name: str) -> Dict[str, Tensor]:
  if model_name not in MODELS:
    raise ValueError(f"Requested unknown mamba model: {model_name}")
  downloaded = fetch(f"https://huggingface.co/state-spaces/mamba-{model_name}/resolve/main/pytorch_model.bin?download=true")
  return torch_load(downloaded)

def selective_scan_ref(
  u,
  delta,
  A,
  B,
  C,
  D=None,
  z=None,
  delta_bias=None,
  delta_softplus=False,
  return_last_state=False,
):
  """
  u: r(B D L)
  delta: r(B D L)
  A: c(D N) or r(D N)
  B: c(D N) or r(B N L) or r(B N 2L) or r(B G N L) or (B G N L)
  C: c(D N) or r(B N L) or r(B N 2L) or r(B G N L) or (B G N L)
  D: r(D)
  z: r(B D L)
  delta_bias: r(D), fp32

  out: r(B D L)
  last_state (optional): r(B D dstate) or c(B D dstate)
  """
  u = u.float()
  delta = delta.float()
  if delta_bias is not None:
    delta = delta + delta_bias[..., None].float()
  if delta_softplus:
    delta = delta.softplus()
  batch, dim, dstate = u.shape[0], A.shape[0], A.shape[1]
  is_variable_B = len(B.shape) >= 3
  is_variable_C = len(C.shape) >= 3
  x = Tensor.zeros(batch, dim, dstate)
  ys = []
  deltaA = Tensor.einsum("bdl,dn->bdln", delta, A).exp()
  if not is_variable_B:
    deltaB_u = Tensor.einsum("bdl,dn,bdl->bdln", delta, B, u)
  else:
    if len(B.shape) == 3:
      deltaB_u = Tensor.einsum("bdl,bnl,bdl->bdln", delta, B, u)
    else:
      B = B.repeat((1, dim // B.shape[1], 1, 1))
      deltaB_u = Tensor.einsum("bdl,bdnl,bdl->bdln", delta, B, u)
  if is_variable_C and len(C.shape) == 4:
    C = C.repeat((1, dim // C.shape[1], 1, 1))
  last_state = None
  for i in range(u.shape[2]):
    x = deltaA[:, :, i] * x + deltaB_u[:, :, i]
    if not is_variable_C:
      y = Tensor.einsum("bdn,dn->bd", x, C)
    else:
      if len(C.shape) == 3:
        y = Tensor.einsum("bdn,bn->bd", x, C[:, :, i])
      else:
        y = Tensor.einsum("bdn,bdn->bd", x, C[:, :, :, i])
    if i == u.shape[2] - 1:
      last_state = x
    ys.append(y)
  y = Tensor.stack(ys, dim=2)  # (batch dim L)
  out = y if D is None else y + u * D.reshape((-1, 1))
  if z is not None:
    out = out * z.silu()
  return out if not return_last_state else (out, last_state)

class MambaMixer:
  def __init__(
    self,
    dim,
    d_state=16,
    d_conv=4,
    expand=2,
    dt_rank="auto",
    dt_min=0.001,
    dt_max=0.1,
    dt_init="random",
    dt_scale=1.0,
    dt_init_floor=1e-4,
    conv_bias=True,
    bias=False,
    layer_idx=None,
  ):
    self.dim = dim
    self.d_state = d_state
    self.d_conv = d_conv
    self.expand = expand
    self.d_inner = self.expand * self.dim
    self.dt_rank = math.ceil(self.dim / 16) if dt_rank == "auto" else dt_rank
    self.layer_idx = layer_idx

    self.in_proj = nn.Linear(self.dim, self.d_inner * 2, bias=bias)

    self.conv1d = nn.Conv1d(in_channels=self.d_inner, out_channels=self.d_inner, bias=conv_bias,
                            kernel_size=d_conv, groups=self.d_inner, padding=d_conv-1)

    self.x_proj = nn.Linear(self.d_inner, self.dt_rank + self.d_state * 2, bias=False)
    self.dt_proj = nn.Linear(self.dt_rank, self.d_inner, bias=True)

    # Initialize special dt projection to preserve variance at initialization
    dt_init_std = self.dt_rank**-0.5 * dt_scale
    if dt_init == "constant":
      self.dt_proj.weight = Tensor.full(self.dt_proj.weight.shape, dt_init_std)
    elif dt_init == "random":
      self.dt_proj.weight = Tensor.uniform(self.dt_proj.weight.shape, low=-dt_init_std, high=dt_init_std)
    else:
      raise NotImplementedError

    dt = Tensor.uniform(self.d_inner, low=math.log(dt_min), high=math.log(dt_max)).exp().maximum(dt_init_floor)
    inv_dt = dt + (1 - (-dt).exp()).log()

    self.dt_proj.bias.assign(inv_dt)

    # S4D real initialization
    self.A_log = Tensor.arange(1, self.d_state+1).repeat([self.d_inner, 1]).log()

    # D "skip" parameter
    self.D = Tensor.ones(self.d_inner)  # Keep in fp32

    self.out_proj = nn.Linear(self.d_inner, self.dim, bias=bias)

  def __call__(self, hidden_states: Tensor, inference_params=None):
    batch, seqlen, dim = hidden_states.shape

    conv_state, ssm_state = None, None
    if inference_params is not None:
      conv_state, ssm_state = self._get_states_from_cache(inference_params, batch)
      if inference_params.seqlen_offset > 0:
        # The states are updated inplace
        out, _, _ = self.step(hidden_states[:, -1:, :], conv_state, ssm_state)
        return out

    xz = self.in_proj.weight @ hidden_states.permute(2,0,1).reshape(hidden_states.shape[2],hidden_states.shape[1]*hidden_states.shape[0])
    xz = xz.reshape(xz.shape[0],xz.shape[1]//seqlen, seqlen).permute(1,0,2)

    if self.in_proj.bias is not None:
      xz = xz + self.in_proj.bias.reshape((-1, 1))

    A = -self.A_log.exp()
    x, z = xz.chunk(2, dim=1)
    # Compute short convolution
    if conv_state is not None:
      conv_state.assign(x[:, :, -self.d_conv :])  # Update state (B D W)
      x = self.conv1d(x)[..., :seqlen].swish()

    x_dbl = self.x_proj(x.permute(0,2,1).reshape(x.shape[0]*x.shape[2], x.shape[1]))
    dt, B, C = Tensor.split(x_dbl, [self.dt_rank, self.d_state, self.d_state], dim=-1)
    dt = self.dt_proj.weight @ dt.T
    dt = dt.reshape(dt.shape[0], dt.shape[1]//seqlen, seqlen).permute(1,0,2)
    B = B.reshape(B.shape[0]//seqlen, seqlen, B.shape[1]).permute(0,2,1)
    C = C.reshape(C.shape[0]//seqlen, seqlen, C.shape[1]).permute(0,2,1)

    # TODO: actually implement selective_scan_fn
    y = selective_scan_ref(x, dt, A, B, C, self.D, z=z, delta_bias=self.dt_proj.bias, delta_softplus=True,
                           return_last_state=ssm_state is not None)
    if ssm_state is not None:
      y, last_state = y
      ssm_state.assign(last_state)

    y = y.permute(0,2,1)
    out = self.out_proj(y)
    return out

  def step(self, hidden_states: Tensor, conv_state: Tensor, ssm_state: Tensor):
    assert hidden_states.shape[1] == 1, f"Only support decoding with 1 token at a time for now, attempted {hidden_states.shape[1]}"
    xz = self.in_proj(hidden_states.squeeze(1))  # (B 2D)
    x, z = xz.chunk(2, dim=-1)  # (B D)

    # Conv step
    conv_state.assign(conv_state[:, :, 1:].cat(x.unsqueeze(-1), dim=-1))
    x = (conv_state * self.conv1d.weight.squeeze(1)).sum(-1)
    if self.conv1d.bias is not None:
      x = x + self.conv1d.bias
    x = x.swish()

    x_db = self.x_proj(x)  # (B dt_rank+2*d_state)
    dt = x_db[:, : self.dt_rank]
    B = x_db[:, self.dt_rank : (self.dt_rank + self.d_state)]
    C = x_db[:, (self.dt_rank + self.d_state) :]
    # Don't add dt_bias here
    dt = self.dt_proj.weight @ dt.T
    A = -self.A_log.exp()

    # SSM step
    dt = (dt + self.dt_proj.bias.unsqueeze(-1)).softplus()
    dA = Tensor.einsum("db,dn->bdn", dt, A).exp()
    dB = Tensor.einsum("db,bn->bdn", dt, B)
    ssm_state.assign(ssm_state * dA + x.unsqueeze(-1) * dB)
    y = Tensor.einsum("bdn,bn->bd", ssm_state, C)
    y = y + self.D * x
    y = y * z.swish()  # (B D)

    out = self.out_proj(y)
    return out.unsqueeze(1), conv_state, ssm_state

  def _get_states_from_cache(self, inference_params, batch_size, initialize_states=False):
    assert self.layer_idx is not None
    if self.layer_idx not in inference_params.key_value_memory_dict:
      conv_state = Tensor.zeros(batch_size, self.dim * self.expand, self.d_conv).contiguous().realize()
      ssm_state = Tensor.zeros(batch_size, self.dim * self.expand, self.d_state).realize()
      inference_params.key_value_memory_dict[self.layer_idx] = (conv_state, ssm_state)
    else:
      conv_state, ssm_state = inference_params.key_value_memory_dict[self.layer_idx]
    return conv_state, ssm_state

class MambaBlock:
  def __init__(self, dim: int, norm_eps: float = 1e-5, rms_norm: bool = True, layer_idx: Optional[int] = None):
    self.mixer = MambaMixer(dim, layer_idx=layer_idx)
    if rms_norm:
      self.norm = RMSNorm(dim, norm_eps)
    else:
      raise NotImplementedError

  def __call__(self, hidden_states: Tensor, residual: Optional[Tensor] = None, inference_params=None):
    residual = (hidden_states + residual) if residual is not None else hidden_states
    hidden_states = self.norm(residual)
    hidden_states = self.mixer(hidden_states, inference_params=inference_params)
    return hidden_states, residual

class MambaBackbone:
  def __init__(self, dim: int, n_layers: int, vocab_size: int, rms_norm: bool = True, norm_eps: float = 1e-5):
    self.embedding = nn.Embedding(vocab_size, dim)
    self.layers = [MambaBlock(dim, rms_norm=rms_norm, layer_idx=i) for i in range(n_layers)]
    if rms_norm:
      self.norm_f = RMSNorm(dim, norm_eps)

  def __call__(self, input_ids: Tensor, inference_params=None) -> Any:
    hidden_states = self.embedding(input_ids)
    residual = None
    for layer in self.layers:
      hidden_states, residual = layer(hidden_states, residual, inference_params=inference_params)

    residual = (hidden_states + residual) if residual is not None else hidden_states
    hidden_states = self.norm_f(residual)

    return hidden_states

class Mamba:
  def __init__(self, dim: int, n_layers: int, vocab_size: int, pad_vocab_size_multiple: int = 1):
    if vocab_size % pad_vocab_size_multiple != 0:
      vocab_size += pad_vocab_size_multiple - (vocab_size % pad_vocab_size_multiple)

    self.backbone = MambaBackbone(dim, n_layers, vocab_size)
    self.lm_head = nn.Linear(dim, vocab_size, bias=False)

    self.forward_jit = TinyJit(self.forward)

  def forward(self, input_ids, inference_params, num_last_tokens):
    hidden_states = self.backbone(input_ids, inference_params=inference_params)
    if num_last_tokens > 0:
      hidden_states = hidden_states[:, -num_last_tokens:]
    return self.lm_head(hidden_states).realize()

  def __call__(self, input_ids, inference_params=None, num_last_tokens=0, jit=True):
    if inference_params is None:
      return self.forward(input_ids, inference_params, num_last_tokens)
    if jit and inference_params.seqlen_offset > 0:
      return self.forward_jit(input_ids, inference_params, num_last_tokens)
    else:
      return self.forward(input_ids, inference_params, num_last_tokens)
  @staticmethod
  def from_pretrained(model_name: str):
    weights = fetch_weights(model_name)
    model = Mamba(**MODELS[model_name])
    load_state_dict(model, weights)

    return model

@dataclass
class InferenceParams:
  """Inference parameters that are passed to the main model in order
  to efficienly calculate and store the context during inference."""
  max_seqlen: int
  max_batch_size: int
  seqlen_offset: int = 0
  batch_size_offset: int = 0
  key_value_memory_dict: dict = field(default_factory=dict)
  lengths_per_sample: Optional[Tensor] = None

  def reset(self, max_seqlen, max_batch_size):
    self.max_seqlen = max_seqlen
    self.max_batch_size = max_batch_size
    self.seqlen_offset = 0
    if self.lengths_per_sample is not None:
      self.lengths_per_sample.zero_()

def generate(model, tokenizer, prompt: str, n_tokens_to_gen: int = 10, sample: bool = False, top_k: int = None):
  tks = tokenizer(prompt)["input_ids"]
  while len(tks) < 4:
    tks = [50279] + tks
  # TODO: sampling
  temperature = 0.5
  start_pos = 0
  inference_params = InferenceParams(max_seqlen=1, max_batch_size=1, seqlen_offset=0)
  for _ in tqdm(range(n_tokens_to_gen), desc="Speed Gen"):
    logits = model(Tensor([tks[start_pos:]]), inference_params, start_pos, jit=False)
    inference_params.seqlen_offset = len(tks)
    tok = logits[:, -1, :].argmax(axis=-1).item()
    start_pos = len(tks)
    tks.append(tok)
  output_completions = ''.join([tokenizer.decode(output) for output in tks])
  return output_completions

if __name__ == "__main__":
  parser = argparse.ArgumentParser(description="Run Mamba in tinygrad", formatter_class=argparse.ArgumentDefaultsHelpFormatter)
  parser.add_argument("--prompt", type=str, default="Why is gravity ", help="Prompt for LLM completion")
  parser.add_argument("--size", type=str, default="370m",
                      help=f"Size of model to use [{', '.join([k for k in MODELS.keys()])}]")
  parser.add_argument("--n_tokens", type=int, default=10, help="Number of tokens to generate")
  args = parser.parse_args()

  tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neox-20b")
  model = Mamba.from_pretrained(args.size)
  prompt = args.prompt
  num_toks = args.n_tokens
  s = time.time()
  tinyoutput = generate(model, tokenizer, prompt, n_tokens_to_gen=num_toks)
  print(tinyoutput)
  print('TIME: ', time.time() - s)
  TORCHOUTPUT = "Why is gravity \nso important?\nBecause it's the only"
  print('Outputs Match:', tinyoutput == TORCHOUTPUT)