File size: 16,876 Bytes
0e3d4b8 | 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 | """SplitBit LLM — full autoregressive transformer model.
Pure NumPy implementation with:
- Token embedding + positional encoding (RoPE)
- Stack of transformer layers
- LM head for next-token prediction
- KV cache for fast autoregressive generation
- Streaming generation (token-by-token)
- SplitBit weight quantization support
"""
from __future__ import annotations
import logging
import math
import os
import time
from typing import Any, Iterator
import numpy as np
from .tokenizer import BPETokenizer, BOS_ID, EOS_ID, PAD_ID
from .layers import TransformerLayer, Embedding, Linear, layer_norm
from .quantization import SplitBitQuantizer
logger = logging.getLogger(__name__)
class SplitBitLLM:
"""Full autoregressive transformer LLM.
forward(tokens) → logits
generate(prompt, max_tokens, temperature) → text
generate_stream(prompt) → iterator yielding tokens
"""
def __init__(self, config: Any = None, tokenizer: BPETokenizer | None = None) -> None:
if config is None:
from ..config import Settings, get_model_config, detect_hardware
config = get_model_config(detect_hardware())
self.config = config
self.tokenizer = tokenizer
# Model architecture
self.embedding = Embedding(config.vocab_size, config.d_model)
self.layers = [
TransformerLayer(config.d_model, config.n_heads, config.d_ff, config.max_seq_len)
for _ in range(config.n_layers)
]
# Final layer norm
self.ln_f_gamma = np.ones(config.d_model, dtype=np.float32)
self.ln_f_beta = np.zeros(config.d_model, dtype=np.float32)
# LM head (tied with embedding)
self.lm_head = Linear(config.d_model, config.vocab_size, bias=False)
# Generation state
self._kv_cache_active = False
self._inference_count = 0
self._total_tokens_generated = 0
self._total_inference_time_s = 0.0
@property
def param_count(self) -> int:
"""Total parameter count."""
total = self.embedding.weight.size + self.lm_head.weight.size
total += self.ln_f_gamma.size + self.ln_f_beta.size
for layer in self.layers:
total += sum(p.size for p in layer.get_params().values())
return total
def forward(self, token_ids: np.ndarray, use_cache: bool = False, past_len: int = 0) -> np.ndarray:
"""
token_ids: [batch, seq_len]
Returns logits: [batch, seq_len, vocab_size]
"""
batch, seq_len = token_ids.shape
# Truncate to max_seq_len
if seq_len > self.config.max_seq_len:
token_ids = token_ids[:, -self.config.max_seq_len:]
seq_len = self.config.max_seq_len
if past_len > 0:
past_len = max(0, past_len - (seq_len - self.config.max_seq_len))
# Embedding
x = self.embedding.forward(token_ids) # [batch, seq_len, d_model]
# Transformer layers
for i, layer in enumerate(self.layers):
x = layer.forward(x, layer_idx=i, use_cache=use_cache, past_len=past_len)
# Final layer norm
x = layer_norm(x, self.ln_f_gamma, self.ln_f_beta)
# LM head
logits = self.lm_head.forward(x) # [batch, seq_len, vocab_size]
return logits
def reset_cache(self) -> None:
"""Reset KV cache for all layers."""
for layer in self.layers:
layer.attn.reset_cache()
self._kv_cache_active = False
def generate(
self,
prompt: str,
max_tokens: int = 128,
temperature: float = 0.7,
top_k: int = 40,
use_cache: bool = True,
) -> str:
"""Generate text from a prompt.
Args:
prompt: input text
max_tokens: max tokens to generate
temperature: sampling temperature (0 = greedy)
top_k: top-k sampling (0 = disabled)
use_cache: use KV cache for faster generation
Returns:
Generated text (prompt + completion)
"""
tokens = self._encode_prompt(prompt)
if not tokens:
return prompt
generated = list(tokens)
self.reset_cache()
# Initial forward pass
token_arr = np.array([generated], dtype=np.int64)
logits = self.forward(token_arr, use_cache=use_cache, past_len=0)
past_len = len(generated)
for _ in range(max_tokens):
# Get logits for last token
next_logits = logits[0, -1, :] # [vocab_size]
# Sample next token
next_token = self._sample(next_logits, temperature, top_k)
if next_token == EOS_ID:
break
generated.append(next_token)
self._total_tokens_generated += 1
# Forward just the new token with cache
if use_cache:
new_arr = np.array([[next_token]], dtype=np.int64)
logits = self.forward(new_arr, use_cache=True, past_len=past_len)
past_len += 1
else:
token_arr = np.array([generated[-self.config.max_seq_len:]], dtype=np.int64)
logits = self.forward(token_arr, use_cache=False, past_len=0)
self.reset_cache()
return self._decode(generated)
def generate_stream(
self,
prompt: str,
max_tokens: int = 128,
temperature: float = 0.7,
top_k: int = 40,
use_cache: bool = True,
) -> Iterator[str]:
"""Streaming generation — yields text chunks as they're generated.
Yields decoded text chunks (may be partial words).
"""
tokens = self._encode_prompt(prompt)
if not tokens:
return
# Yield prompt first
yield self.tokenizer.decode(tokens) if self.tokenizer else ""
generated = list(tokens)
self.reset_cache()
# Initial forward pass
token_arr = np.array([generated], dtype=np.int64)
logits = self.forward(token_arr, use_cache=use_cache, past_len=0)
past_len = len(generated)
for _ in range(max_tokens):
next_logits = logits[0, -1, :]
next_token = self._sample(next_logits, temperature, top_k)
if next_token == EOS_ID:
break
generated.append(next_token)
self._total_tokens_generated += 1
# Decode just this token
if self.tokenizer:
chunk = self.tokenizer.decode([next_token])
if chunk:
yield chunk
if use_cache:
new_arr = np.array([[next_token]], dtype=np.int64)
logits = self.forward(new_arr, use_cache=True, past_len=past_len)
past_len += 1
else:
token_arr = np.array([generated[-self.config.max_seq_len:]], dtype=np.int64)
logits = self.forward(token_arr, use_cache=False, past_len=0)
self.reset_cache()
def generate_stream_sentences(
self,
prompt: str,
max_tokens: int = 128,
temperature: float = 0.7,
top_k: int = 40,
) -> Iterator[str]:
"""Streaming generation that yields complete sentences.
Used for voice/TTS — first sentence comes out ASAP.
"""
buffer = ""
for chunk in self.generate_stream(prompt, max_tokens, temperature, top_k):
buffer += chunk
# Check for sentence boundaries
while buffer:
# Find sentence end
end_idx = -1
for delim in [". ", "! ", "? ", ".\n", "!\n", "?\n"]:
idx = buffer.find(delim)
if idx >= 0 and (end_idx < 0 or idx < end_idx):
end_idx = idx + len(delim)
if end_idx > 0:
yield buffer[:end_idx]
buffer = buffer[end_idx:]
else:
break
if buffer:
yield buffer
def _encode_prompt(self, prompt: str) -> list[int]:
"""Encode prompt to token IDs."""
if self.tokenizer:
return self.tokenizer.encode(prompt, add_bos=True)
# Fallback: simple char-level encoding
return [BOS_ID] + [min(ord(c), self.config.vocab_size - 1) for c in prompt[:self.config.max_seq_len - 1]]
def _decode(self, tokens: list[int]) -> str:
"""Decode tokens to text."""
if self.tokenizer:
return self.tokenizer.decode(tokens)
return "".join(chr(t) for t in tokens if t < 128 and t not in (PAD_ID, BOS_ID, EOS_ID))
def _sample(self, logits: np.ndarray, temperature: float, top_k: int) -> int:
"""Sample next token from logits."""
if temperature <= 0:
return int(np.argmax(logits))
# Apply temperature
logits = logits / max(temperature, 1e-8)
# Top-k filtering
if top_k > 0 and top_k < len(logits):
top_indices = np.argpartition(logits, -top_k)[-top_k:]
mask = np.full_like(logits, -1e9)
mask[top_indices] = logits[top_indices]
logits = mask
# Softmax and sample
probs = np.exp(logits - np.max(logits))
probs = probs / np.sum(probs)
return int(np.random.choice(len(probs), p=probs))
def save(self, path: str, quantizer: SplitBitQuantizer | None = None) -> None:
"""Save model to disk. If quantizer provided, weights are quantized."""
os.makedirs(os.path.dirname(path) or ".", exist_ok=True)
data = {
"config": {
"n_layers": self.config.n_layers,
"n_heads": self.config.n_heads,
"d_model": self.config.d_model,
"d_ff": self.config.d_ff,
"vocab_size": self.config.vocab_size,
"max_seq_len": self.config.max_seq_len,
},
"embedding": self.embedding.weight,
"lm_head": self.lm_head.weight,
"ln_f_gamma": self.ln_f_gamma,
"ln_f_beta": self.ln_f_beta,
"layers": [],
"quantized": quantizer is not None,
}
for layer in self.layers:
params = layer.get_params()
if quantizer:
layer_data = {}
for k, v in params.items():
if "ln" in k:
layer_data[k] = v # Don't quantize layer norm
else:
packed = quantizer.quantize(v)
layer_data[k] = packed
else:
layer_data = params
data["layers"].append(layer_data)
if quantizer:
data["embedding"] = quantizer.quantize(self.embedding.weight)
data["lm_head"] = quantizer.quantize(self.lm_head.weight)
np.savez(path, **self._flatten_save_dict(data))
logger.info("Model saved to %s (quantized=%s)", path, quantizer is not None)
def _flatten_save_dict(self, data: dict, prefix: str = "") -> dict:
"""Flatten nested dict for np.savez."""
flat = {}
for k, v in data.items():
key = f"{prefix}_{k}" if prefix else k
if isinstance(v, dict) and "data" not in v:
flat.update(self._flatten_save_dict(v, key))
elif isinstance(v, list):
for i, item in enumerate(v):
flat.update(self._flatten_save_dict(item, f"{key}_{i}"))
elif isinstance(v, np.ndarray):
flat[key] = v
elif isinstance(v, dict):
# Packed quantized data
for pk, pv in v.items():
if isinstance(pv, np.ndarray):
flat[f"{key}_{pk}"] = pv
elif pv is not None:
flat[f"{key}_{pk}"] = np.array(pv)
elif v is not None:
flat[key] = np.array(v)
return flat
@classmethod
def load(cls, path: str, tokenizer: BPETokenizer | None = None,
quantizer: SplitBitQuantizer | None = None) -> "SplitBitLLM":
"""Load model from disk."""
from ..config import ModelConfig
npz = np.load(path, allow_pickle=True)
config = ModelConfig(
n_layers=int(npz["config_n_layers"]),
n_heads=int(npz["config_n_heads"]),
d_model=int(npz["config_d_model"]),
d_ff=int(npz["config_d_ff"]),
vocab_size=int(npz["config_vocab_size"]),
max_seq_len=int(npz["config_max_seq_len"]),
)
model = cls(config=config, tokenizer=tokenizer)
# Load embedding and LM head
if quantizer and "embedding_data" in npz:
model.embedding.weight = quantizer.dequantize({
"data": npz["embedding_data"],
"scale": npz["embedding_scale"] if "embedding_scale" in npz else None,
"shape": npz["embedding_shape"],
"format": str(npz["embedding_format"]) if "embedding_format" in npz else "q4_k_m",
"bits": int(npz["embedding_bits"]) if "embedding_bits" in npz else 4,
"n_blocks": int(npz["embedding_n_blocks"]) if "embedding_n_blocks" in npz else 0,
"block_size": int(npz["embedding_block_size"]) if "embedding_block_size" in npz else 32,
"pad_len": int(npz["embedding_pad_len"]) if "embedding_pad_len" in npz else 0,
})
model.lm_head.weight = quantizer.dequantize({
"data": npz["lm_head_data"],
"scale": npz["lm_head_scale"] if "lm_head_scale" in npz else None,
"shape": npz["lm_head_shape"],
"format": str(npz["lm_head_format"]) if "lm_head_format" in npz else "q4_k_m",
"bits": int(npz["lm_head_bits"]) if "lm_head_bits" in npz else 4,
"n_blocks": int(npz["lm_head_n_blocks"]) if "lm_head_n_blocks" in npz else 0,
"block_size": int(npz["lm_head_block_size"]) if "lm_head_block_size" in npz else 32,
"pad_len": int(npz["lm_head_pad_len"]) if "lm_head_pad_len" in npz else 0,
})
else:
model.embedding.weight = npz["embedding"]
model.lm_head.weight = npz["lm_head"]
model.ln_f_gamma = npz["ln_f_gamma"]
model.ln_f_beta = npz["ln_f_beta"]
# Load layers
for i, layer in enumerate(model.layers):
params = {}
for key in ["wq", "wk", "wv", "wo", "w1", "w2"]:
full_key = f"layers_{i}_{key}"
if quantizer and f"{full_key}_data" in npz:
params[key] = quantizer.dequantize({
"data": npz[f"{full_key}_data"],
"scale": npz[f"{full_key}_scale"] if f"{full_key}_scale" in npz else None,
"shape": npz[f"{full_key}_shape"],
"format": str(npz[f"{full_key}_format"]) if f"{full_key}_format" in npz else "q4_k_m",
"bits": int(npz[f"{full_key}_bits"]) if f"{full_key}_bits" in npz else 4,
"n_blocks": int(npz[f"{full_key}_n_blocks"]) if f"{full_key}_n_blocks" in npz else 0,
"block_size": int(npz[f"{full_key}_block_size"]) if f"{full_key}_block_size" in npz else 32,
"pad_len": int(npz[f"{full_key}_pad_len"]) if f"{full_key}_pad_len" in npz else 0,
})
elif full_key in npz:
params[key] = npz[full_key]
for key in ["ln1_gamma", "ln1_beta", "ln2_gamma", "ln2_beta"]:
full_key = f"layers_{i}_{key}"
if full_key in npz:
params[key] = npz[full_key]
layer.set_params(params)
logger.info("Model loaded from %s (%d params)", path, model.param_count)
return model
def get_stats(self) -> dict[str, Any]:
"""Get model statistics."""
avg_time = self._total_inference_time_s / max(self._inference_count, 1)
return {
"param_count": self.param_count,
"config": {
"n_layers": self.config.n_layers,
"n_heads": self.config.n_heads,
"d_model": self.config.d_model,
"d_ff": self.config.d_ff,
"vocab_size": self.config.vocab_size,
"max_seq_len": self.config.max_seq_len,
},
"inference_count": self._inference_count,
"total_tokens_generated": self._total_tokens_generated,
"avg_inference_time_s": round(avg_time, 4),
"tokens_per_second": round(self._total_tokens_generated / max(self._total_inference_time_s, 0.001), 2),
}
|