File size: 35,974 Bytes
41c5b75 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 |
import math
import os
from dataclasses import dataclass
from typing import Optional
from huggingface_hub import hf_hub_download
import lm_eval as evaluator
import torch
import torch.nn as nn
import torch.nn.functional as F
from safetensors.torch import load_file
from torchtune.modules import RotaryPositionalEmbeddings
from transformers import (
AutoConfig,
AutoModel,
AutoModelForCausalLM,
PreTrainedModel,
PretrainedConfig,
)
from transformers.modeling_outputs import CausalLMOutput
try:
from flashfftconv import FlashFFTConv
flash_fft_available = True
except ImportError as e:
print(f"Unable to import FlashFFTConv: {e}. Falling back to PyTorch implementation.")
flash_fft_available = False
try:
from flash_attn import flash_attn_func
except ImportError as e:
print(f"Unable to import Triton-based flash attention: {e}. No alternative currently available.")
os.environ["HF_ALLOW_CODE_EVAL"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "false"
loss_fn = nn.CrossEntropyLoss()
def nearest_power_of_two(n: int, round_up: bool = False) -> int:
if n <= 1:
return 1
return 1 << ((n - 1).bit_length() if round_up else (n).bit_length() - 1)
def find_multiple(n: int, k: int) -> int:
if n % k == 0:
return n
return n + k - (n % k)
def get_hankel(seq_len: int, use_hankel_L: bool = False) -> torch.Tensor:
entries = torch.arange(1, seq_len + 1, dtype=torch.float64)
i_plus_j = entries.reshape(-1, 1) + entries.reshape(1, -1)
if use_hankel_L:
sgn = (-1.0) ** (i_plus_j - 2.0) + 1.0
denom = (i_plus_j + 3.0) * (i_plus_j - 1.0) * (i_plus_j + 1.0)
Z = sgn * (8.0 / denom)
elif not use_hankel_L:
Z = 2.0 / (i_plus_j**3 - i_plus_j)
else:
raise ValueError("use_hankel_L must be a boolean")
return Z
def get_spectral_filters(
seq_len: int,
K: int,
use_hankel_L: bool = False,
device: torch.device = None,
dtype: torch.dtype = torch.float64,
) -> torch.Tensor:
Z = get_hankel(seq_len, use_hankel_L).to(device)
sigma, phi = torch.linalg.eigh(Z)
sigma_k, phi_k = sigma[-K:], phi[:, -K:]
phi_k *= sigma_k**0.25
return phi_k.to(device=device, dtype=dtype)
class BaseConfigForCausalLM(PretrainedConfig):
"""Base PretrainedConfig class to be decorated with dataclass"""
model_type = "base_model"
def __init__(self, **kwargs):
super().__init__(**kwargs)
@dataclass
class FlashSTUConfig(BaseConfigForCausalLM):
model_type = "FlashSTU"
# Define fields with defaults (as before)
bsz: int = 1
dim: int = 1024
r: int = 1024
num_heads: int = 12
num_local_heads: Optional[int] = -1
num_layers: int = 12
seq_len: int = 4096
n: int = 8191
window_size: int = 2048
vocab_size: int = 200064
inter_dim: Optional[int] = 3072
mlp_scale: Optional[float] = 12.0
weight_tying: Optional[bool] = True
bias: Optional[bool] = False
rope_theta: Optional[float] = 10000.0
softcap: Optional[float] = 50.0
num_eigh: Optional[int] = 24
use_hankel_L: Optional[bool] = False
use_flash_fft: Optional[bool] = True
use_tensordot: Optional[bool] = True
use_attn: Optional[bool] = True
use_alibi: Optional[bool] = False
torch_dtype: torch.dtype = torch.bfloat16
device: torch.device = None
# Explicit __init__ to handle **kwargs for PretrainedConfig compatibility
def __init__(
self,
bsz: int = 1,
dim: int = 1024,
r: int = 1024,
num_heads: int = 12,
num_local_heads: Optional[int] = -1,
num_layers: int = 12,
seq_len: int = 4096,
n: int = 8191,
window_size: int = 2048,
vocab_size: int = 200064,
inter_dim: Optional[int] = 3072,
mlp_scale: Optional[float] = 12.0,
weight_tying: Optional[bool] = True,
bias: Optional[bool] = False,
rope_theta: Optional[float] = 10000.0,
softcap: Optional[float] = 50.0,
num_eigh: Optional[int] = 24,
use_hankel_L: Optional[bool] = False,
use_flash_fft: Optional[bool] = True,
use_tensordot: Optional[bool] = True,
use_attn: Optional[bool] = True,
use_alibi: Optional[bool] = False,
torch_dtype: torch.dtype = torch.bfloat16,
device: torch.device = None,
**kwargs, # Catch extra arguments like model_type
):
super().__init__(**kwargs) # Pass kwargs to parent __init__
# Assign fields from arguments
self.bsz = bsz
self.dim = dim
self.r = r
self.num_heads = num_heads
self.num_local_heads = num_local_heads
self.num_layers = num_layers
self.seq_len = seq_len
self.n = n
self.window_size = window_size
self.vocab_size = vocab_size
self.inter_dim = inter_dim
self.mlp_scale = mlp_scale
self.weight_tying = weight_tying
self.bias = bias
self.rope_theta = rope_theta
self.softcap = softcap
self.num_eigh = num_eigh
self.use_hankel_L = use_hankel_L
self.use_flash_fft = use_flash_fft
self.use_tensordot = use_tensordot
self.use_attn = use_attn
self.use_alibi = use_alibi
self.torch_dtype = torch_dtype
self.device = device
# Explicitly call __post_init__ if defined and needed
self.__post_init__()
def __post_init__(self):
# Ensure torch_dtype is a torch.dtype object, not a string
if isinstance(self.torch_dtype, str):
try:
self.torch_dtype = getattr(torch, self.torch_dtype)
except AttributeError:
raise ValueError(f"Invalid torch_dtype string: {self.torch_dtype}")
if self.num_local_heads == -1:
self.num_local_heads = self.num_heads
if self.inter_dim is None:
hidden_dim = self.mlp_scale * self.dim
num_hidden = int(2 * hidden_dim / 3)
self.inter_dim = find_multiple(num_hidden, 256)
self.head_dim = self.dim // self.num_heads
@classmethod
def from_name(cls, name: str):
# presets = {
# "tiny": dict(dim=128, num_heads=4, num_layers=2, vocab_size=10000),
# "small": dict(dim=256, num_heads=8, num_layers=4, vocab_size=20000),
# "gpt2-small": dict(dim=768, num_heads=12, num_layers=12, vocab_size=50257),
# # add more as needed
# }
# if name not in presets:
# raise ValueError(f"Unknown model config name: {name}")
# return cls(**presets[name])
print("Not yet implemented")
pass
class MLP(nn.Module):
def __init__(self, config: FlashSTUConfig) -> None:
super().__init__()
self.w1 = nn.Linear(config.dim, config.inter_dim)
self.w2 = nn.Linear(config.inter_dim, config.dim)
self.w2.SCALE_INIT = 1
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.w2(F.gelu(self.w1(x), approximate="tanh"))
class SlidingWindowAttention(nn.Module):
def __init__(self, config):
super().__init__()
self.wq = nn.Linear(config.dim, config.dim)
self.wk = nn.Linear(config.dim, config.dim)
self.wv = nn.Linear(config.dim, config.dim)
self.wo = nn.Linear(config.dim, config.dim)
self.wo.SCALE_INIT = 1
self.dim = config.dim
self.head_dim = config.head_dim
self.num_heads = config.num_heads
self.num_local_heads = config.num_local_heads
self.window_size = config.window_size
self.softcap = config.softcap
self.alibi_slopes = self._get_alibi_slopes(self.num_heads) if config.use_alibi else None
self.rotary_emb = RotaryPositionalEmbeddings(
dim=self.head_dim,
max_seq_len=config.seq_len,
base=config.rope_theta,
)
def forward(self, x):
bsz, seq_len, dim = x.shape
q, k, v = self.wq(x), self.wk(x), self.wv(x)
q = q.view(bsz, seq_len, self.num_heads, self.head_dim)
k = k.view(bsz, seq_len, self.num_local_heads, self.head_dim)
v = v.view(bsz, seq_len, self.num_local_heads, self.head_dim)
if self.alibi_slopes is None:
q, k = self.rotary_emb(q), self.rotary_emb(k)
y = flash_attn_func(
q=q,
k=k,
v=v,
causal=True,
window_size=(self.window_size, 0),
# softcap=self.softcap,
alibi_slopes=self.alibi_slopes,
)
out = y.reshape(bsz, seq_len, -1)
out = self.wo(out)
return out
def _generate_slopes(self, n: int):
start = 2 ** (-(2 ** -(math.log2(n) - 3)))
return [start * (start**i) for i in range(n)]
def _get_alibi_slopes(self, num_heads: int, interpolation_factor: float = 0.25):
# If n_heads is a power of 2, generate slopes directly
if math.log2(num_heads).is_integer():
slopes = self._generate_slopes(num_heads)
else:
# Get slopes for the nearest power of two
n = nearest_power_of_two(num_heads, round_up=False)
slopes_power_of_two = self._generate_slopes(n)
# Generate extra slopes
extra_slopes = self._generate_slopes(2 * n)
extra_slopes_trunc = extra_slopes[0::2][: num_heads - n]
slopes = slopes_power_of_two + extra_slopes_trunc
slopes = torch.tensor(slopes, device=torch.device("cuda")) # FA ALiBi must be on CUDA
slopes = slopes * interpolation_factor # https://arxiv.org/pdf/2310.13017
return slopes
class STU(nn.Module):
def __init__(self, config):
super().__init__()
# Set at top-level post- model init
self.stu_filters = None
self.stu_filters_fft = None # TODO: Optimization: Precompute FFT of filters
self.n = config.n
self.num_eigh = config.num_eigh
self.d_in = config.dim
self.d_out = config.dim
self.r = config.r
self.use_hankel_L = config.use_hankel_L
self.use_tensordot = config.use_tensordot
self.flash_fft = (
FlashFFTConv(self.n, dtype=torch.bfloat16) if config.use_flash_fft and flash_fft_available else None
)
# TODO: Add dimensionality reduction `r` here.
if self.use_tensordot:
self.M_inputs = nn.Parameter(torch.zeros(self.d_in, self.d_out))
self.M_filters = nn.Parameter(torch.zeros(self.num_eigh, self.d_in))
else:
self.M_phi_plus = nn.Parameter(torch.zeros(self.num_eigh, self.d_in, self.d_out))
if not self.use_hankel_L:
self.M_phi_minus = nn.Parameter(torch.zeros(self.num_eigh, self.d_in, self.d_out))
def forward(self, x: torch.Tensor) -> torch.Tensor:
B, L, D = x.shape
if self.use_tensordot:
# Contract inputs and filters over (K, D) dims first, then convolve
x_proj = x @ self.M_inputs
phi_proj = self.stu_filters @ self.M_filters
if self.flash_fft:
spectral_plus, spectral_minus = self.flash_conv(x_proj, phi_proj, self.flash_fft, self.use_tensordot)
else:
spectral_plus, spectral_minus = self.conv(x_proj, phi_proj, self.n, self.use_tensordot)
else:
# Convolve inputs and filters first, then contract over (K, D) dims
if self.flash_fft:
U_plus, U_minus = self.flash_conv(x, self.stu_filters, self.flash_fft, self.use_tensordot)
else:
U_plus, U_minus = self.conv(x, self.stu_filters, self.n, self.use_tensordot)
B, L, K, D = U_plus.shape
spectral_plus = U_plus.reshape(B, L, K * D) @ self.M_phi_plus.reshape(K * D, self.d_out)
if not self.use_hankel_L:
spectral_minus = U_minus.reshape(B, L, K * D) @ self.M_phi_minus.reshape(K * D, self.d_out)
out = spectral_plus if self.use_hankel_L else spectral_plus + spectral_minus
return out
def conv(
self, u: torch.Tensor, v: torch.Tensor, n: int, use_tensordot: bool = True
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Performs convolution via FFT with causal alignment using a negative featurization.
The input tensor u is modulated by an alternating sign tensor (sgn) that multiplies every other
time step by -1. This "negative featurization" modulates the phase so that in this implementation
the correct causal output is obtained by simply slicing the first L elements (i.e. [:seq_len]).
Note: Using a conventional slice [seq_len-1:2*seq_len-1] would yield a flipped alignment, resulting in leakage.
Args:
u: Input tensor of shape (bsz, seq_len, d_in).
v: Kernel tensor; expected shape is (seq_len, d_out) if use_tensordot is True.
n: FFT length (typically set to 2*seq_len - 1 for linear convolution with implicit right zero-padding).
use_tensordot: Boolean flag to control kernel reshaping.
Returns:
A tuple (U_plus, U_minus) where:
- U_plus is the primary convolution output.
- U_minus is the secondary output, corrected by the sign tensor.
"""
bsz, seq_len, d_in = u.shape
sgn = torch.full((1, seq_len, 1), 1, device=u.device)
sgn[:, 1::2] *= -1 # Apply negative featurization: multiply every other element by -1.
if use_tensordot:
_, d_out = v.shape
v = v.view(1, -1, d_out, 1).to(torch.float32).contiguous()
else:
_, K = v.shape
sgn = sgn.unsqueeze(-1)
v = v.view(1, -1, K, 1, 1).to(torch.float32).contiguous() # (bsz, seq_len, K, d_in, stack)
u = u.view(bsz, -1, 1, d_in).expand(bsz, -1, K, d_in)
# Cast kernel to float32 for FFT
v_fft = torch.fft.rfft(v.to(torch.float32), n=n, dim=1)
U = torch.stack([u, u * sgn], dim=-1).to(torch.float32).contiguous()
# Cast input stack to float32 for FFT
U_fft = torch.fft.rfft(U.to(torch.float32), n=n, dim=1)
# Slicing the first seq_len outputs yields the proper causal convolution given the negative modulation.
# Perform convolution in float32 and cast back
U_conv = torch.fft.irfft(v_fft * U_fft, n=n, dim=1)[:, :seq_len].to(u.dtype)
U_plus, U_minus = torch.unbind(U_conv, dim=-1)
U_minus = U_minus * sgn
return U_plus.type_as(u), U_minus.type_as(u)
def flash_conv(
self,
u: torch.Tensor,
v: torch.Tensor,
flash_fft: FlashFFTConv,
use_tensordot: bool = True,
) -> tuple[torch.Tensor, torch.Tensor]:
"""Flash FFT convolution.
Args:
u (torch.Tensor): Input tensor of shape `(B, L, d_in)`, where:
- `B` is the batch size,
- `L` is the sequence length,
- `d_in` is the input dimension.
v (torch.Tensor): Filter tensor of shape `(K, d_in)`, where:
- `K` is the number of filters,
- `d_in` is the input dimension.
flash_fft (FlashFFTConv): An instance of the FlashFFTConv module, used to perform the convolution.
use_tensordot (bool, optional): If `True`, performs the tensordot approximation (default is `True`).
Returns:
tuple[torch.Tensor, torch.Tensor]: A tuple `(U_plus, U_minus)`:
- `U_plus`: Convolved output tensor with positive eigenvalues.
- Shape depends on `use_tensordot`:
- If `use_tensordot=True`: `(B, L, d_in)`
- If `use_tensordot=False`: `(B, L, K, d_in)`
- `U_minus`: Convolved output tensor with negative eigenvalues.
- Shape depends on `use_tensordot`:
- If `use_tensordot=True`: `(B, L, d_in)`
- If `use_tensordot=False`: `(B, L, K, d_in)`
Raises:
ValueError: If the input tensor shapes do not conform to the expected dimensions.
Example:
>>> u = torch.randn(4, 16, 32) # (B, L, d_in)
>>> v = torch.randn(8, 32) # (K, d_in)
>>> flash_fft = FlashFFTConv(n=16, dtype=torch.float32)
>>> U_plus, U_minus = flash_convolve(u, v, flash_fft, use_tensordot=True)
>>> print(U_plus.shape, U_minus.shape)
torch.Size([4, 16, 32]) torch.Size([4, 16, 32])
"""
bsz, seq_len, d_in = u.shape
_, K = v.shape
padded_len = nearest_power_of_two(seq_len, round_up=True)
pad_len = padded_len - seq_len
sgn = torch.full((1, 1, padded_len), 1, device=u.device)
sgn[:, :, 1::2] = -1
if use_tensordot:
u_padded = F.pad(u.transpose(1, 2), (0, pad_len)).to(torch.bfloat16)
v_padded = F.pad(v.transpose(0, 1), (0, pad_len)).to(torch.float32)
u_conv = torch.stack([u_padded, u_padded * sgn], dim=0).reshape(2 * bsz, d_in, padded_len)
else:
u_k_padded = F.pad(u.transpose(1, 2), (0, pad_len)).repeat_interleave(K, dim=1)
v_padded = F.pad(v.transpose(0, 1), (0, pad_len)).to(torch.float32).repeat(d_in, 1)
u_conv = torch.stack([u_k_padded, u_k_padded * sgn], dim=0).reshape(2 * bsz, K * d_in, padded_len)
# Ensure inputs to flash_fft are bfloat16 (input) and float32 (kernel)
U_conv = flash_fft(u_conv.to(torch.bfloat16), v_padded.to(torch.float32))
# Trim the output back to the original sequence length
U_conv = U_conv[..., :seq_len]
u_plus, u_minus = torch.chunk(U_conv, 2, dim=0)
if use_tensordot:
u_minus = u_minus * sgn[:, :, :seq_len]
U_plus, U_minus = u_plus.transpose(1, 2), u_minus.transpose(1, 2)
else:
sgn = sgn[:, :, :seq_len].unsqueeze(-1).transpose(1, 2)
U_plus = u_plus.view(bsz, d_in, K, seq_len).permute(0, 3, 2, 1).contiguous()
U_minus = u_minus.view(bsz, d_in, K, seq_len).permute(0, 3, 2, 1).contiguous() * sgn
return U_plus, U_minus
class SlidingWindowAttentionLayer(nn.Module):
def __init__(self, config):
super().__init__()
self.swa_norm = nn.LayerNorm(config.dim)
self.swa = SlidingWindowAttention(config)
self.mlp_norm = nn.LayerNorm(config.dim)
self.mlp = MLP(config)
def forward(self, x):
x = x + self.swa(self.swa_norm(x))
x = x + self.mlp(self.mlp_norm(x))
return x
class STULayer(nn.Module):
def __init__(self, config):
super().__init__()
self.stu_norm = nn.LayerNorm(config.dim)
self.stu = STU(config)
self.mlp_norm = nn.LayerNorm(config.dim)
self.mlp = MLP(config)
def forward(self, x):
x = x + self.stu(self.stu_norm(x))
x = x + self.mlp(self.mlp_norm(x))
return x
class FlashSTU(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.tok_emb = nn.Embedding(config.vocab_size, config.dim)
self.layers = nn.ModuleList()
for layer_idx in range(config.num_layers):
# For more complex %-split arrangements, see https://arxiv.org/pdf/2406.07887
if layer_idx % 2 == 0:
self.layers.append(STULayer(config))
else:
self.layers.append(SlidingWindowAttentionLayer(config)) if config.use_attn else self.layers.append(
STULayer(config)
)
self.norm_f = nn.LayerNorm(config.dim)
self.lm_head = nn.Linear(config.dim, config.vocab_size, bias=False)
if self.config.weight_tying:
self.tok_emb.weight = self.lm_head.weight
self.std = self.config.dim**-0.5
def init_weights(self, module):
std = self.std
if isinstance(module, nn.Linear):
if hasattr(module, "SCALE_INIT"):
std *= (2 * self.config.num_layers) ** -0.5
torch.nn.init.normal_(module.weight, mean=0.0, std=std)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std=std)
def forward(self, input_ids: torch.Tensor, labels: torch.Tensor = None, **kwargs) -> CausalLMOutput:
x = self.tok_emb(input_ids)
for layer in self.layers:
x = layer(x)
x = self.norm_f(x)
logits = self.lm_head(x)
loss = None
if labels is not None:
loss = loss_fn(logits.flatten(0, 1), labels.flatten(0, 1))
return CausalLMOutput(
loss=loss,
logits=logits,
)
def setup_filters(
self,
spectral_filters: torch.Tensor,
spectral_filters_fft: torch.Tensor,
):
for layer in self.layers:
if isinstance(layer, STULayer):
layer.stu.stu_filters = spectral_filters
layer.stu.stu_filters_fft = spectral_filters_fft
def get_num_params(self):
"""
Return the number of parameters in the model.
For non-embedding count (default), the position embeddings get subtracted.
"""
n_params = sum(p.numel() for p in self.parameters())
return n_params
def create_base_model_components(model_name_or_path=None, **kwargs):
"""Create config and filters needed for model initialization"""
if model_name_or_path is not None:
config = FlashSTUConfig.from_pretrained(model_name_or_path, **kwargs)
else:
config = FlashSTUConfig(**kwargs)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
filters = get_spectral_filters(
seq_len=config.seq_len,
K=config.num_eigh,
use_hankel_L=config.use_hankel_L,
device=device,
dtype=config.torch_dtype,
)
assert filters.dtype == config.torch_dtype, f"filters dtype is {filters.dtype}, expected {config.torch_dtype}"
return config, filters
class FlashSTUForCausalLM(PreTrainedModel):
"""Thin wrapper to comply with HuggingFace's expected interface"""
config_class = FlashSTUConfig
base_model_prefix = "FlashSTU"
def __init__(self, config):
super().__init__(config)
self.flash_stu = FlashSTU(config)
self.flash_stu.apply(self.flash_stu.init_weights)
device = (
config.device
if config.device is not None
else torch.device("cuda" if torch.cuda.is_available() else "cpu")
)
torch_dtype = config.torch_dtype # Assumes __post_init__ already converted it to torch.dtype
spectral_filters = get_spectral_filters(
seq_len=config.seq_len,
K=config.num_eigh,
use_hankel_L=config.use_hankel_L,
device=device,
# Note: get_spectral_filters returns float64, cast later
)
spectral_filters_fft = torch.fft.rfft(spectral_filters, n=config.n, dim=1)
# Setup filters in the model, casting to the target dtype
self.flash_stu.setup_filters(
spectral_filters.to(dtype=torch_dtype), spectral_filters_fft.to(dtype=torch_dtype)
)
# Note: Moving the entire model to device happens later, after loading weights.
def forward(
self, input_ids: torch.Tensor, labels: torch.Tensor = None, attention_mask: torch.Tensor = None, **kwargs
) -> CausalLMOutput:
outputs = self.flash_stu(input_ids, labels=labels, **kwargs)
return outputs
def generate(
self,
input_ids: torch.Tensor,
max_length: int = 32,
num_return_sequences: int = 4,
temperature: float = 0.8,
top_k: int = 50,
top_p: float = 0.95,
repetition_penalty: float = 1.2,
seed: int = 42,
) -> torch.Tensor:
"""Generate text using top-k and nucleus sampling with temperature and repetition penalty.
Args:
input_ids: Input token ids of shape (batch_size, seq_len)
max_length: Maximum length of generated sequence
num_return_sequences: Number of sequences to generate per input
temperature: Sampling temperature. Higher = more random, lower = more focused
top_k: Number of highest probability tokens to keep for top-k sampling
top_p: Cumulative probability cutoff for nucleus sampling
repetition_penalty: Penalty factor for repeating tokens. 1.0 = no penalty
seed: Random seed for reproducibility
Returns:
Generated token ids of shape (num_return_sequences, max_length)
"""
self.eval() # Set to eval mode
device = input_ids.device
# Expand input for multiple sequences
input_ids = input_ids.repeat(num_return_sequences, 1)
generated = input_ids
# Set up generator for reproducible sampling
sample_rng = torch.Generator(device=device)
sample_rng.manual_seed(seed)
# Generate tokens until we reach max_length
with torch.no_grad():
while generated.size(1) < max_length:
# Get logits for next token
outputs = self.flash_stu(generated)
next_token_logits = outputs.logits[:, -1, :]
# Apply repetition penalty
if repetition_penalty != 1.0:
for i in range(generated.shape[0]):
for token in generated[i]:
if token in next_token_logits[i]:
next_token_logits[i, token] /= repetition_penalty
# Apply temperature
if temperature != 1.0:
next_token_logits = next_token_logits / temperature
# Get probabilities
probs = torch.nn.functional.softmax(next_token_logits, dim=-1)
# Top-k sampling
if top_k > 0:
indices_to_remove = probs < torch.topk(probs, top_k)[0][..., -1, None]
probs[indices_to_remove] = 0
# Nucleus (top-p) sampling
if top_p < 1.0:
sorted_probs, sorted_indices = torch.sort(probs, descending=True)
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
# Remove tokens with cumulative probability above the threshold
sorted_indices_to_remove = cumulative_probs > top_p
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
probs[indices_to_remove] = 0
# Renormalize probabilities
probs = probs / probs.sum(dim=-1, keepdim=True).clamp(min=1e-8)
# Sample next token
next_token = torch.multinomial(probs, num_samples=1, generator=sample_rng)
# Append to generated sequence
generated = torch.cat([generated, next_token], dim=1)
return generated
def get_num_params(self):
return self.flash_stu.get_num_params()
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
# Get config and create model
config, _ = create_base_model_components(pretrained_model_name_or_path, **kwargs)
model = cls(config)
# Download safetensors file from hub
weights_path = hf_hub_download(
repo_id=pretrained_model_name_or_path,
filename="model.safetensors",
cache_dir=kwargs.get("cache_dir"),
force_download=kwargs.get("force_download", False),
proxies=kwargs.get("proxies", None),
local_files_only=kwargs.get("local_files_only", False),
use_auth_token=kwargs.get("use_auth_token", None),
revision=kwargs.get("revision", None),
subfolder=kwargs.get("subfolder", ""),
)
state_dict = load_file(weights_path)
# Reconstruct weight tying for tok_emb and lm_head
tok_emb_key = "tok_emb.weight"
lm_head_key = "lm_head.weight"
tok_emb_present = tok_emb_key in state_dict
lm_head_present = lm_head_key in state_dict
if tok_emb_present and not lm_head_present:
print(f"Reconstructing weight tying: Linking missing '{lm_head_key}' to existing '{tok_emb_key}'")
state_dict[lm_head_key] = state_dict[tok_emb_key]
elif lm_head_present and not tok_emb_present:
print(f"Reconstructing weight tying: Linking missing '{tok_emb_key}' to existing '{lm_head_key}'")
state_dict[tok_emb_key] = state_dict[lm_head_key]
elif not tok_emb_present and not lm_head_present:
# This case should ideally not happen if the file is valid
print(
f"Warning: Neither '{tok_emb_key}' nor '{lm_head_key}' found in state_dict. Weight tying cannot be reconstructed."
)
# If both are present, assume they are loaded correctly (or were never tied)
# Prepend 'flash_stu.' to all keys to match wrapper's state dict
final_state_dict = {f"flash_stu.{k}": v for k, v in state_dict.items()}
model.load_state_dict(final_state_dict)
# Move to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device=device, dtype=torch.bfloat16)
model.eval()
# Print parameter count as a sanity check
num_params = model.get_num_params()
print(f"\nModel loaded: {pretrained_model_name_or_path}")
print(f"Parameter count: {num_params / 1e6:.2f}M")
return model
# Create initial config and filters for registration
config, filters = create_base_model_components()
# Register models
AutoConfig.register("FlashSTU", FlashSTUConfig)
AutoModel.register(FlashSTUConfig, FlashSTU)
AutoModelForCausalLM.register(FlashSTUConfig, FlashSTUForCausalLM)
print("Registered FlashSTU model and configuration.")
def run_model_diagnostics(model, tokenizer, device):
"""Run detailed diagnostics to analyze model behavior."""
print("\nRunning model diagnostics...")
# Test cases of varying difficulty and length
test_cases = [
# Simple completion
"2 + 2 =",
# Medium difficulty
"The capital of France is Paris. The capital of Germany is",
# Complex reasoning
"If a train travels 120 kilometers in 2 hours, its average speed is",
# Pattern completion
"1, 2, 3, 4,",
# Long context
"The following is a detailed explanation of photosynthesis: Plants use sunlight to",
]
with torch.no_grad():
for prompt in test_cases:
print(f"\nAnalyzing prompt: {prompt}")
# Tokenize
tokens = tokenizer(prompt, return_tensors="pt")
input_ids = tokens["input_ids"].to(device)
outputs = model.flash_stu(input_ids, labels=input_ids)
labels = input_ids.clone()
shift_logits = outputs.logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss_fct = nn.CrossEntropyLoss(reduction="none")
token_losses = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)).view(
shift_labels.size()
)
# Print token-by-token analysis
input_tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
print("\nToken-by-token loss:")
for i, (token, loss) in enumerate(zip(input_tokens[1:], token_losses[0])):
print(f"{token}: {loss.item():.3f}")
print(f"Average loss: {token_losses.mean().item():.3f}")
# Generate with different temperatures
temps = [0.5, 0.7, 1.0]
print("\nGeneration temperature comparison:")
for temp in temps:
gen_ids = model.generate(
input_ids,
max_length=25,
num_return_sequences=1,
temperature=temp,
top_p=0.9,
repetition_penalty=1.5,
seed=42,
)
gen_text = tokenizer.decode(gen_ids[0], skip_special_tokens=True)
print(f"\nTemp {temp}: {gen_text}")
def validate_model_generation():
print("\nRunning generation validation test...")
try:
from transformers import AutoTokenizer
# Load model and tokenizer
# model_id = "Hazan-Lab/Flash_STU_550M"
model_id = "Hazan-Lab/FlashSTU-340M-0428"
model = FlashSTUForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
# Move to GPU if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device=device, dtype=torch.bfloat16)
model.eval()
# Print parameter count as a sanity check
num_params = model.get_num_params()
print(f"\nModel loaded: {model_id}")
print(f"Parameter count: {num_params / 1e6:.2f}M")
# Run additional diagnostics
run_model_diagnostics(model, tokenizer, device)
except Exception as e:
print(f"\nError during validation: {str(e)}")
raise
# Run evaluation tasks
tasks = [
# "mmlu",
"hellaswag",
# "piqa",
# "siqa",
# "boolq",
# "winogrande",
# "commonsense_qa",
# "openbookqa",
# "arc",
# "arc_easy",
# "arc_challenge",
# "triviaqa",
# "nq_open",
# "humaneval",
# "mbpp",
# "gms8k",
# "hendrycks_math",
# "mathqa",
# "minerva_math",
# "score",
# "asdiv",
# "agieval",
# "bigbench",
]
tasks_fewshot = {
"hellaswag": 0,
# "mmlu": 5,
# "piqa": 0,
# "siqa": 0,
# "boolq": 0,
# "winogrande": -1,
# "commonsense_qa": 7,
# "openbookqa": -1,
# "arc": -1,
# "arc_easy": -1,
# "arc_challenge": -1,
# "triviaqa": 5,
# "nq_open": 5,
# "humaneval": -1,
# "mbpp": 3,
# "gms8k": -1,
# "hendrycks_math": 4,
# "mathqa": -1,
# "minerva_math": -1,
# "score": -1,
# "asdiv": -1,
# "agieval": -1,
# "bigbench": -1,
}
all_results = {}
# First validate generation works
validate_model_generation()
print("\nStarting evaluation tasks...")
for task in tasks:
print(f"\nEvaluating task: {task}")
eval_kwargs = dict(
model="hf",
model_args=(
# "pretrained=Hazan-Lab/Flash_STU_550M,"
"pretrained=Hazan-Lab/FlashSTU-340M-0428,"
"trust_remote_code=True,"
"dtype=bfloat16,"
"cache_dir=/scratch/gpfs/mn4560/hazan-lab/tensorized_filters/tensorized_filters/eval/cache"
),
tasks=[task],
batch_size="auto",
device="cuda:0",
)
few_shot_value = tasks_fewshot.get(task, -1)
if few_shot_value != -1:
eval_kwargs["num_fewshot"] = few_shot_value
results = evaluator.simple_evaluate(**eval_kwargs)
task_result = results["results"].get(task, {})
all_results[task] = task_result
print(f"Results for {task}:")
print(task_result)
print("\n" + "=" * 50 + "\n")
print("All Evaluation Results:")
for task, result in all_results.items():
print(f"{task}: {result}")
|