File size: 10,957 Bytes
0a626fd | 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 | """
TRM (Tiny Recursive Model) adapted for Causal Language Modeling (Chess).
Based on the official implementation: TinyRecursiveModels/models/recursive_reasoning/trm.py
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
# -----------------------------------------------------------------------------
# Configuration
# -----------------------------------------------------------------------------
class ChessConfig(PretrainedConfig):
model_type = "chess_transformer"
def __init__(
self,
vocab_size: int = 1200,
n_embd: int = 128,
n_head: int = 4,
n_ctx: int = 256,
h_cycles: int = 2, # Number of High-level reasoning cycles
l_cycles: int = 2, # Number of Low-level reasoning cycles per H-cycle
n_layers_per_block: int = 1, # Number of physical layers in the shared block
n_inner: Optional[int] = None,
n_layer: Optional[int] = None, # Not used directly; total layers = h_cycles * l_cycles
dropout: float = 0.0, # TRM usually uses 0 dropout for reasoning
layer_norm_epsilon: float = 1e-5,
tie_weights: bool = True,
rope_theta: float = 10000.0,
pad_token_id: int = 0, # Assuming 0 is padding based on your log
bos_token_id: int = 1,
eos_token_id: int = 2,
**kwargs,
):
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
**kwargs,
)
self.vocab_size = vocab_size
self.n_embd = n_embd
self.n_head = n_head
self.n_ctx = n_ctx
self.h_cycles = h_cycles
self.l_cycles = l_cycles
self.n_layers_per_block = n_layers_per_block
self.n_layers = n_layer
self.n_inner = n_inner if n_inner is not None else int(n_embd * 8/3) # SwiGLU convention
self.dropout = dropout
self.layer_norm_epsilon = layer_norm_epsilon
self.tie_weights = tie_weights
self.rope_theta = rope_theta
class RMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x):
var = torch.mean(x**2, dim=-1, keepdim=True)
x = x * torch.rsqrt(var + self.eps)
return self.weight * x
class RotaryEmbedding(nn.Module):
def __init__(self, dim, max_position_embeddings=2048, base=10000.0, device=None):
super().__init__()
self.dim = dim
self.base = base
self.max_position_embeddings = max_position_embeddings
self.register_buffer("inv_freq", None, persistent=False)
self.register_buffer("cos_cached", None, persistent=False)
self.register_buffer("sin_cached", None, persistent=False)
def _update_cos_sin_tables(self, x, seq_len):
if (self.cos_cached is None or self.cos_cached.shape[0] < seq_len):
self.inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=x.device).float() / self.dim))
t = torch.arange(max(seq_len, self.max_position_embeddings), device=x.device).float()
freqs = torch.outer(t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
self.cos_cached = emb.cos()
self.sin_cached = emb.sin()
def forward(self, x, seq_len=None):
if seq_len is None:
seq_len = x.shape[1]
self._update_cos_sin_tables(x, seq_len)
return self.cos_cached[:seq_len, ...], self.sin_cached[:seq_len, ...]
def rotate_half(x):
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def apply_rotary_pos_emb(q, k, cos, sin):
# q, k: [batch, seq, head, dim] (after transpose)
# cos, sin: [seq, dim] -> need broadcast
cos = cos.unsqueeze(0).unsqueeze(2) # [1, seq, 1, dim]
sin = sin.unsqueeze(0).unsqueeze(2)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
class MultiQueryAttention(nn.Module):
"""
Standard Attention with RoPE support.
Using Multi-Query (MQA) or standard MHA depending on config.
Adapted for Causal Masking.
"""
def __init__(self, config: ChessConfig):
super().__init__()
self.n_head = config.n_head
self.n_embd = config.n_embd
self.head_dim = config.n_embd // config.n_head
self.c_q = nn.Linear(config.n_embd, config.n_embd, bias=False)
self.c_k = nn.Linear(config.n_embd, self.head_dim, bias=False)
self.c_v = nn.Linear(config.n_embd, self.head_dim, bias=False)
self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=False)
self.dropout = nn.Dropout(config.dropout)
def forward(self, x, cos, sin, attention_mask=None):
B, T, C = x.size()
q = self.c_q(x).view(B, T, self.n_head, self.head_dim)
k = self.c_k(x).view(B, T, 1, self.head_dim)
v = self.c_v(x).view(B, T, 1, self.head_dim)
q, k = apply_rotary_pos_emb(q, k, cos, sin)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
k = k.expand(-1, self.n_head, -1, -1)
v = v.expand(-1, self.n_head, -1, -1)
y = F.scaled_dot_product_attention(
q, k, v,
attn_mask=None,
dropout_p=self.dropout.p if self.training else 0.0,
is_causal=True
)
y = y.transpose(1, 2).contiguous().view(B, T, C)
y = self.c_proj(y)
return y
class SwiGLU(nn.Module):
def __init__(self, config: ChessConfig):
super().__init__()
self.w1 = nn.Linear(config.n_embd, config.n_inner, bias=False)
self.w2 = nn.Linear(config.n_embd, config.n_inner, bias=False)
self.w3 = nn.Linear(config.n_inner, config.n_embd, bias=False)
self.dropout = nn.Dropout(config.dropout)
def forward(self, x):
x1 = self.w1(x)
x2 = self.w2(x)
hidden = F.silu(x1) * x2
return self.dropout(self.w3(hidden))
class TRMBlock(nn.Module):
def __init__(self, config: ChessConfig):
super().__init__()
self.self_attn = MultiQueryAttention(config)
self.mlp = SwiGLU(config)
self.ln_1 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.ln_2 = RMSNorm(config.n_embd, eps=config.layer_norm_epsilon)
def forward(self, x, cos, sin):
attn_out = self.self_attn(x, cos, sin)
x = self.ln_1(x + attn_out)
mlp_out = self.mlp(x)
x = self.ln_2(x + mlp_out)
return x
class TRMReasoningModule(nn.Module):
"""
The reusable module containing shared layers.
Implements Input Injection: hidden_states = hidden_states + injection
"""
def __init__(self, config: ChessConfig):
super().__init__()
self.layers = nn.ModuleList([TRMBlock(config) for _ in range(config.n_layers_per_block)])
def forward(self, hidden_states, input_injection, cos, sin):
hidden_states = hidden_states + input_injection
for layer in self.layers:
hidden_states = layer(hidden_states, cos, sin)
return hidden_states
class ChessForCausalLM(PreTrainedModel):
config_class = ChessConfig
def __init__(self, config: ChessConfig):
super().__init__(config)
self.config = config
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
self.rotary = RotaryEmbedding(config.n_embd // config.n_head, max_position_embeddings=config.n_ctx, base=config.rope_theta)
self.reasoning_module = TRMReasoningModule(config)
self.z_H_init = nn.Parameter(torch.randn(1, 1, config.n_embd) * 0.02)
self.z_L_init = nn.Parameter(torch.randn(1, 1, config.n_embd) * 0.02)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
if config.tie_weights:
self.lm_head.weight = self.wte.weight
self.post_init()
def _init_weights(self, module):
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
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=0.02)
def forward(
self,
input_ids: torch.LongTensor,
labels: Optional[torch.LongTensor] = None,
return_dict: Optional[bool] = None,
**kwargs,
) -> Union[Tuple, CausalLMOutputWithPast]:
B, T = input_ids.size()
x_emb = self.wte(input_ids)
cos, sin = self.rotary(x_emb, seq_len=T)
z_H = self.z_H_init.expand(B, T, -1).contiguous()
z_L = self.z_L_init.expand(B, T, -1).contiguous()
with torch.no_grad():
for _h in range(self.config.h_cycles - 1):
# L-loop (updates z_L)
for _l in range(self.config.l_cycles):
z_L = self.reasoning_module(
hidden_states=z_L,
input_injection=(z_H + x_emb),
cos=cos, sin=sin
)
# H-loop step (updates z_H)
z_H = self.reasoning_module(
hidden_states=z_H,
input_injection=z_L,
cos=cos, sin=sin
)
for _l in range(self.config.l_cycles):
z_L = self.reasoning_module(
hidden_states=z_L,
input_injection=(z_H + x_emb),
cos=cos, sin=sin
)
z_H = self.reasoning_module(
hidden_states=z_H,
input_injection=z_L,
cos=cos, sin=sin
)
logits = self.lm_head(z_H)
loss = None
if labels is not None:
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
loss_fct = nn.CrossEntropyLoss(ignore_index=-100)
loss = loss_fct(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1))
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=None
)
from transformers import AutoConfig, AutoModelForCausalLM
AutoConfig.register("chess_transformer", ChessConfig)
AutoModelForCausalLM.register(ChessConfig, ChessForCausalLM) |