Instructions to use dappalumbo91/fluid-language-model with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use dappalumbo91/fluid-language-model with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="dappalumbo91/fluid-language-model", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("dappalumbo91/fluid-language-model", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use dappalumbo91/fluid-language-model with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "dappalumbo91/fluid-language-model" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dappalumbo91/fluid-language-model", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/dappalumbo91/fluid-language-model
- SGLang
How to use dappalumbo91/fluid-language-model with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "dappalumbo91/fluid-language-model" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dappalumbo91/fluid-language-model", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "dappalumbo91/fluid-language-model" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "dappalumbo91/fluid-language-model", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use dappalumbo91/fluid-language-model with Docker Model Runner:
docker model run hf.co/dappalumbo91/fluid-language-model
File size: 17,946 Bytes
fd43622 | 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 | """A small model trained from nothing, on the relation itself.
Each number is a sign and four digits, least place first, so a place is a
token. The slow stack learns the true sum, difference, product, exact
quotient, and comparison. A second head says whether a finished equation
holds. A short-term residual stays silent during that training. It moves
only when the second head rejects a draft, and the draft it then prefers
is one of its own alternatives. An outside answer is a score, not a target.
This run writes ``weights/scratch/`` only.
python scratch.py
"""
from __future__ import annotations
import json
import operator
from pathlib import Path
import torch
import torch.nn.functional as F
from torch import nn
ROOT = Path(__file__).resolve().parent
DEST = ROOT / "weights" / "scratch"
SEED = 20261109
WIDTH = 4
D_MODEL = 128
LAYERS = 4
HEADS = 4
RANK = 8
EPOCHS = 100
BATCH = 256
LR = 1e-3
FAR_LO = 31
FAR_HI = 45
PLUS_SIGN = 2
MINUS_SIGN = 3
OP_ADD = 4
OP_SUB = 5
OP_MUL = 6
OP_DIV = 7
OP_LARGER = 8
OP_SMALLER = 9
EQ = 10
DIGIT0 = 11
VOCAB = DIGIT0 + 10
PROMPT_LEN = 12
ANSWER_LEN = 5
LENGTH = PROMPT_LEN + ANSWER_LEN
YES = 0
NO = 1
OPS = (OP_ADD, OP_SUB, OP_MUL, OP_DIV, OP_LARGER, OP_SMALLER)
STIFF = ((OP_ADD, -5, 3), (OP_LARGER, 3, 9), (OP_SMALLER, 9, 3))
def _truth(op: int, left: int, right: int) -> int | None:
if op == OP_DIV:
if right == 0 or left % right != 0:
return None
return operator.floordiv(left, right)
return {
OP_ADD: operator.add,
OP_SUB: operator.sub,
OP_MUL: operator.mul,
OP_LARGER: max,
OP_SMALLER: min,
}[op](left, right)
def _expect(op: int, left: int, right: int) -> int | None:
if op == OP_ADD:
return left + right
if op == OP_SUB:
return left - right
if op == OP_MUL:
return left * right
if op == OP_DIV:
if right == 0 or left % right != 0:
return None
return left // right
if op == OP_LARGER:
return left if left >= right else right
if op == OP_SMALLER:
return left if left <= right else right
return None
def _span_held(left: int, right: int) -> bool:
"""Scattered holdout.
A mod-5 test collapses to the ones digits, because 10 ≡ 0 (mod 5).
That deletes the whole digit sum, so a later place can never learn it.
"""
return (left * 3 + right * 5) % 7 == 0
def _holds(op: int, left: int, right: int) -> bool:
return (op, left, right) in STIFF or _span_held(left, right)
def _number(value: int) -> list[int]:
sign = PLUS_SIGN if value >= 0 else MINUS_SIGN
remain = abs(value)
digits = []
for _ in range(WIDTH):
digits.append(DIGIT0 + remain % 10)
remain //= 10
if remain:
raise RuntimeError(f"{value} does not fit in {WIDTH} digits")
return [sign, *digits]
def _read_number(tokens: list[int]) -> int | None:
if len(tokens) != ANSWER_LEN or tokens[0] not in (PLUS_SIGN, MINUS_SIGN):
return None
if any(token < DIGIT0 or token >= DIGIT0 + 10 for token in tokens[1:]):
return None
value = 0
for place, token in enumerate(tokens[1:]):
value += (token - DIGIT0) * 10**place
return value if tokens[0] == PLUS_SIGN else -value
def _sequence(op: int, left: int, right: int, value: int) -> list[int]:
row = _number(left) + [op] + _number(right) + [EQ] + _number(value)
if len(row) != LENGTH:
raise RuntimeError("sequence length drifted")
return row
def _rows(ops: tuple[int, ...], lo: int, hi: int, keep) -> list[tuple[int, int, int, int]]:
found = []
for op in ops:
for left in range(lo, hi + 1):
for right in range(lo, hi + 1):
if not keep(op, left, right):
continue
value = _truth(op, left, right)
if value is None or value != _expect(op, left, right):
continue
found.append((op, left, right, value))
return found
def _banks() -> dict[str, list[tuple[int, int, int, int]]]:
core = _rows(OPS, 0, 30, lambda op, left, right: not _holds(op, left, right))
signed = _rows((OP_ADD, OP_SUB), -12, 12, lambda op, left, right: (left < 0 or right < 0) and not _holds(op, left, right))
hold_in = _rows(OPS, 0, 30, lambda op, left, right: _span_held(left, right) and (op, left, right) not in STIFF)
hold_far = _rows(OPS, FAR_LO, FAR_HI, lambda op, left, right: True)
stiff = []
for op, left, right in STIFF:
value = _truth(op, left, right)
if value != _expect(op, left, right):
raise RuntimeError(f"stiff label drifted {op} {left} {right}")
stiff.append((op, left, right, value))
train_keys = {(op, left, right) for op, left, right, _value in core + signed}
for name, rows in (("in", hold_in), ("far", hold_far), ("stiff", stiff)):
leaked = [row for row in rows if (row[0], row[1], row[2]) in train_keys]
if leaked:
raise RuntimeError(f"{name} leaked: {leaked[:3]}")
return {"train": core + signed, "in": hold_in, "far": hold_far, "stiff": stiff}
def _audit(banks: dict[str, list[tuple[int, int, int, int]]]) -> None:
for name, rows in banks.items():
if not rows:
raise RuntimeError(f"{name} bank is empty")
for op, left, right, value in rows:
if _truth(op, left, right) != value or _expect(op, left, right) != value:
raise RuntimeError(f"bad label in {name}: {op} {left} {right} {value}")
if _read_number(_number(value)) != value:
raise RuntimeError(f"digit roundtrip failed for {value}")
print(
f"audit train {len(banks['train'])} in {len(banks['in'])} far {len(banks['far'])} stiff {len(banks['stiff'])}",
flush=True,
)
class Scratch(nn.Module):
"""Slow stack plus a silent short-term residual and a yes/no head."""
def __init__(self) -> None:
super().__init__()
self.embed = nn.Embedding(VOCAB, D_MODEL)
block = nn.TransformerEncoderLayer(
d_model=D_MODEL,
nhead=HEADS,
dim_feedforward=D_MODEL * 4,
dropout=0.0,
activation="gelu",
batch_first=True,
norm_first=True,
)
self.stack = nn.TransformerEncoder(block, num_layers=LAYERS, enable_nested_tensor=False)
self.read = nn.Linear(D_MODEL, VOCAB)
self.verdict = nn.Linear(D_MODEL, 2)
self.fast_down = nn.Parameter(torch.empty(D_MODEL, RANK))
self.fast_up = nn.Parameter(torch.zeros(RANK, D_MODEL))
nn.init.normal_(self.fast_down, std=0.02)
self.register_buffer("fast_down0", self.fast_down.detach().clone())
def reset_fast(self) -> None:
self.fast_down.data.copy_(self.fast_down0)
self.fast_up.data.zero_()
def forward(self, tokens: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
hidden = self.embed(tokens)
hidden = hidden + (hidden @ self.fast_down @ self.fast_up)
mask = nn.Transformer.generate_square_subsequent_mask(
hidden.shape[1], device=hidden.device, dtype=hidden.dtype
)
hidden = self.stack(hidden, mask=mask, is_causal=True)
return self.read(hidden), self.verdict(hidden[:, -1])
def _tensors(rows: list[tuple[int, int, int, int]], device: torch.device) -> torch.Tensor:
data = [_sequence(*row) for row in rows]
return torch.tensor(data, dtype=torch.long, device=device)
def _corrupt(batch: torch.Tensor, generator: torch.Generator) -> torch.Tensor:
forged = batch.clone()
for row in forged:
value = _read_number(row[PROMPT_LEN:].tolist())
if value is None:
raise RuntimeError("corrupt saw a non-number")
for _ in range(6):
trial = row.clone()
if value != 0 and int(torch.randint(0, 2, (1,), generator=generator)) == 0:
trial[PROMPT_LEN] = MINUS_SIGN if trial[PROMPT_LEN] == PLUS_SIGN else PLUS_SIGN
else:
slot = PROMPT_LEN + 1 + int(torch.randint(0, WIDTH, (1,), generator=generator))
trial[slot] = DIGIT0 + (trial[slot] - DIGIT0 + 1 + int(torch.randint(0, 9, (1,), generator=generator))) % 10
if _read_number(trial[PROMPT_LEN:].tolist()) != value:
row.copy_(trial)
break
else:
raise RuntimeError("could not forge a wrong answer")
return forged
def _answer_loss(logits: torch.Tensor, batch: torch.Tensor) -> torch.Tensor:
pred = logits[:, :-1, :]
labels = batch[:, 1:].clone()
labels[:, : PROMPT_LEN - 1] = -100
return F.cross_entropy(pred.reshape(-1, VOCAB), labels.reshape(-1), ignore_index=-100)
def _slow_parameters(model: Scratch) -> list[nn.Parameter]:
return [param for name, param in model.named_parameters() if not name.startswith("fast_")]
@torch.no_grad()
def _greedy(model: Scratch, prompt: torch.Tensor) -> torch.Tensor:
tokens = prompt.clone()
for index in range(ANSWER_LEN):
logits, _verdict = model(tokens)
nxt = logits[:, -1, :]
if index == 0:
choice = nxt[:, PLUS_SIGN : MINUS_SIGN + 1].argmax(dim=-1) + PLUS_SIGN
else:
choice = nxt[:, DIGIT0 : DIGIT0 + 10].argmax(dim=-1) + DIGIT0
tokens = torch.cat((tokens, choice.unsqueeze(1)), dim=1)
return tokens
@torch.no_grad()
def _exact(model: Scratch, rows: list[tuple[int, int, int, int]], device: torch.device) -> tuple[int, int]:
if not rows:
return 0, 0
hits = 0
prompts = _tensors(rows, device)[:, :PROMPT_LEN]
for start in range(0, len(rows), BATCH):
drafted = _greedy(model, prompts[start : start + BATCH])
for row, draft in zip(rows[start : start + BATCH], drafted.tolist()):
if _read_number(draft[PROMPT_LEN:]) == row[3]:
hits += 1
return hits, len(rows)
def _span(rows: list[tuple[int, int, int, int]], count: int = 512) -> list[tuple[int, int, int, int]]:
if len(rows) <= count:
return rows
step = max(1, len(rows) // count)
return rows[::step][:count]
def _copy_left(rows: list[tuple[int, int, int, int]]) -> int:
return sum(value == left for _op, left, _right, value in rows)
@torch.no_grad()
def _verdict_hits(model: Scratch, rows: list[tuple[int, int, int, int]], device: torch.device) -> tuple[int, int]:
hits = 0
prompts = _tensors(rows, device)[:, :PROMPT_LEN]
for start in range(0, len(rows), BATCH):
chunk = rows[start : start + BATCH]
drafted = _greedy(model, prompts[start : start + BATCH])
_logits, verdict = model(drafted)
called = verdict.argmax(dim=-1).tolist()
for row, draft, call in zip(chunk, drafted.tolist(), called):
right = _read_number(draft[PROMPT_LEN:]) == row[3]
if (call == YES and right) or (call == NO and not right):
hits += 1
return hits, len(rows)
def _alternate(model: Scratch, draft: torch.Tensor) -> torch.Tensor:
"""Second token at the answer place where the margin is thinnest."""
tokens = draft[:PROMPT_LEN].clone().unsqueeze(0)
margins = []
choices = []
for index in range(ANSWER_LEN):
logits, _verdict = model(tokens)
nxt = logits[0, -1]
if index == 0:
pool = nxt[PLUS_SIGN : MINUS_SIGN + 1]
base = PLUS_SIGN
else:
pool = nxt[DIGIT0 : DIGIT0 + 10]
base = DIGIT0
order = torch.topk(pool, k=2)
margins.append(float(order.values[0] - order.values[1]))
choices.append(base + int(order.indices[1]))
taken = draft[PROMPT_LEN + index].view(1, 1)
tokens = torch.cat((tokens, taken), dim=1)
slot = min(range(ANSWER_LEN), key=margins.__getitem__)
revised = draft.clone()
revised[PROMPT_LEN + slot] = choices[slot]
return revised
def _revise(model: Scratch, rows: list[tuple[int, int, int, int]], device: torch.device) -> tuple[int, int, int]:
"""On a rejected draft, keep the alternative the verdict prefers."""
improved = changed = rejected = 0
was_right = 0
prompts = _tensors(rows, device)
fast = [model.fast_down, model.fast_up]
for row, full in zip(rows, prompts):
model.reset_fast()
draft = _greedy(model, full[:PROMPT_LEN].unsqueeze(0))[0]
_logits, verdict = model(draft.unsqueeze(0))
if int(verdict.argmax()) == YES:
continue
rejected += 1
other = _alternate(model, draft)
_logits, other_verdict = model(other.unsqueeze(0))
if float(other_verdict[0, YES]) <= float(verdict[0, YES]):
continue
optimizer = torch.optim.SGD(fast, lr=0.1)
target = other[PROMPT_LEN:].unsqueeze(0)
prompt = full[:PROMPT_LEN].unsqueeze(0)
optimizer.zero_grad(set_to_none=True)
made = _greedy_train(model, prompt, target)
made.backward()
optimizer.step()
revised = _greedy(model, prompt)[0]
before = _read_number(draft[PROMPT_LEN:].tolist()) == row[3]
after = _read_number(revised[PROMPT_LEN:].tolist()) == row[3]
if revised[PROMPT_LEN:].tolist() != draft[PROMPT_LEN:].tolist():
changed += 1
if after and not before:
improved += 1
if before:
was_right += 1
return rejected, changed, improved
def _greedy_train(model: Scratch, prompt: torch.Tensor, target: torch.Tensor) -> torch.Tensor:
tokens = prompt
losses = []
for index in range(ANSWER_LEN):
logits, _verdict = model(tokens)
losses.append(F.cross_entropy(logits[:, -1, :], target[:, index]))
tokens = torch.cat((tokens, target[:, index].unsqueeze(1)), dim=1)
return torch.stack(losses).mean()
def _fit(model: Scratch, train: list[tuple[int, int, int, int]], banks: dict[str, list[tuple[int, int, int, int]]], device: torch.device) -> None:
generator = torch.Generator(device="cpu")
generator.manual_seed(SEED)
optimizer = torch.optim.AdamW(_slow_parameters(model), lr=LR, weight_decay=0.01)
data = _tensors(train, device)
best_far = -1
best_state = None
for epoch in range(1, EPOCHS + 1):
model.train()
order = torch.randperm(len(train), generator=generator)
total = 0.0
seen = 0
for start in range(0, len(train), BATCH):
batch = data[order[start : start + BATCH]]
forged = _corrupt(batch, generator)
optimizer.zero_grad(set_to_none=True)
logits, verdict = model(batch)
_forged_logits, forged_verdict = model(forged)
loss = _answer_loss(logits, batch)
loss = loss + 0.5 * F.cross_entropy(verdict, torch.full((len(batch),), YES, device=device, dtype=torch.long))
loss = loss + 0.5 * F.cross_entropy(forged_verdict, torch.full((len(batch),), NO, device=device, dtype=torch.long))
loss.backward()
torch.nn.utils.clip_grad_norm_(_slow_parameters(model), 1.0)
optimizer.step()
model.reset_fast()
total += float(loss.detach()) * len(batch)
seen += len(batch)
if epoch == 1 or epoch % 10 == 0 or epoch == EPOCHS:
model.eval()
model.reset_fast()
train_hits, train_n = _exact(model, _span(train), device)
far_hits, far_n = _exact(model, banks["far"], device)
print(
f"epoch {epoch} loss {total / seen:.4f} train {train_hits}/{train_n} far {far_hits}/{far_n} gpu_gib {torch.cuda.memory_allocated() / 1024**3:.2f}",
flush=True,
)
if far_hits > best_far:
best_far = far_hits
best_state = {key: value.detach().clone() for key, value in model.state_dict().items()}
if best_state is None:
raise RuntimeError("no epoch was scored")
model.load_state_dict(best_state)
print(f"kept far {best_far}/{len(banks['far'])}", flush=True)
def main() -> None:
if not torch.cuda.is_available():
raise SystemExit("CUDA is not available. Refusing to build the model in system RAM.")
print(f"device {torch.cuda.get_device_name(0)}", flush=True)
banks = _banks()
_audit(banks)
torch.manual_seed(SEED)
device = torch.device("cuda")
model = Scratch().to(device)
if any(param.device.type != "cuda" for param in model.parameters()):
raise SystemExit("a parameter is off the GPU")
print(f"gpu_mib {torch.cuda.memory_allocated() / 1024**2:.1f} params {sum(param.numel() for param in model.parameters())}", flush=True)
if float(model.fast_up.detach().norm()) != 0.0:
raise SystemExit("short-term layer is not silent")
_fit(model, banks["train"], banks, device)
model.eval()
model.reset_fast()
report = {}
for name in ("train", "in", "far", "stiff"):
sample = banks[name] if name != "train" else _span(banks["train"])
hits, count = _exact(model, sample, device)
copied = _copy_left(sample)
verdict_hits, verdict_n = _verdict_hits(model, sample, device)
report[name] = {"exact": hits, "count": count, "copy_left": copied, "verdict": verdict_hits, "verdict_n": verdict_n}
print(
f"{name} {hits}/{count} copy_left {copied}/{count} verdict {verdict_hits}/{verdict_n}",
flush=True,
)
rejected, changed, improved = _revise(model, banks["stiff"] + banks["far"][:64], device)
print(f"revise rejected {rejected} changed {changed} improved {improved}", flush=True)
DEST.mkdir(parents=True, exist_ok=True)
torch.save(model.state_dict(), DEST / "model.pt")
(DEST / "scores.json").write_text(json.dumps(report, indent=2), encoding="utf-8")
print(f"wrote {DEST / 'model.pt'}", flush=True)
if __name__ == "__main__":
main()
|