Text Generation
Transformers
Safetensors
PyTorch
English
perceiver_ar
mrcogito
meridian21
custom_code
concept-memory
size-140m
budget-11.5bt
long-context
base-model
research
e38
conversational
Eval Results (legacy)
Instructions to use Meridian21Lab/concept-memory-140m-11.5bt with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Meridian21Lab/concept-memory-140m-11.5bt with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Meridian21Lab/concept-memory-140m-11.5bt", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("Meridian21Lab/concept-memory-140m-11.5bt", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Meridian21Lab/concept-memory-140m-11.5bt with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Meridian21Lab/concept-memory-140m-11.5bt" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Meridian21Lab/concept-memory-140m-11.5bt", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/Meridian21Lab/concept-memory-140m-11.5bt
- SGLang
How to use Meridian21Lab/concept-memory-140m-11.5bt 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 "Meridian21Lab/concept-memory-140m-11.5bt" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Meridian21Lab/concept-memory-140m-11.5bt", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'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 "Meridian21Lab/concept-memory-140m-11.5bt" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Meridian21Lab/concept-memory-140m-11.5bt", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use Meridian21Lab/concept-memory-140m-11.5bt with Docker Model Runner:
docker model run hf.co/Meridian21Lab/concept-memory-140m-11.5bt
File size: 6,759 Bytes
e1ba4e9 | 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 | """Text samples for a model card: nucleus sampling and beam search, the way current base-model cards show text.
Greedy decoding loops on small base models ("The capital of France is Paris. The capital of France is ..."), so
greedy samples undersell a model and are not how anyone would run it. This script writes, for each prompt:
sample temperature 0.7, nucleus (top-p) 0.9, repetition penalty 1.1 over prompt + reply (as in 🤗 Transformers'
`repetition_penalty`), one reply per seed (seeds 0, 1, 2 by default), so a card can show the seed-0 reply
and say it was not picked;
beam 4 beams, no repeated 3-gram, the highest-scoring finished beam (length-normalised log-probability).
Works for our perceiver_ar checkpoints (loaded with `load_perceiver_ar_for_eval`, SDPA attention: prompts here are
short, and flex attention would recompile for every new length) and for any 🤗 causal LM (`--hf`). Every step re-runs
the whole row (no cache): fine for prompts and replies of a few hundred tokens.
uv run python evaluation/sample_text.py --model <our checkpoint folder> --out samples.json
uv run python evaluation/sample_text.py --model HuggingFaceTB/SmolLM2-135M --hf --out smollm2_samples.json
"""
from __future__ import annotations
import argparse
import json
import sys
import time
from pathlib import Path
import torch
import torch.nn.functional as F
PROMPTS = [
"The capital of France is",
"In 1905, Albert Einstein published",
"def fibonacci(n):",
]
def penalise(logits: torch.Tensor, seen: torch.Tensor, penalty: float) -> torch.Tensor:
"""🤗 Transformers' repetition penalty: divide positive logits of seen tokens by `penalty`, multiply negative."""
if penalty == 1.0:
return logits
g = logits.gather(-1, seen)
logits = logits.scatter(-1, seen, torch.where(g > 0, g / penalty, g * penalty))
return logits
def top_p_filter(logits: torch.Tensor, p: float) -> torch.Tensor:
srt, idx = logits.sort(-1, descending=True)
cum = F.softmax(srt, -1).cumsum(-1)
drop = cum - F.softmax(srt, -1) > p # keep the smallest set whose mass reaches p (the top token always stays)
srt = srt.masked_fill(drop, float("-inf"))
return torch.full_like(logits, float("-inf")).scatter(-1, idx, srt)
@torch.no_grad()
def next_logits(model, ids: torch.Tensor) -> torch.Tensor:
return model(input_ids=ids).logits[:, -1].float()
@torch.no_grad()
def sample(model, ids, new, eos, seed, temperature=0.7, top_p=0.9, penalty=1.1) -> list[int]:
g = torch.Generator(device=ids.device).manual_seed(seed)
out = ids
for _ in range(new):
logits = penalise(next_logits(model, out), out, penalty) / temperature
nxt = torch.multinomial(F.softmax(top_p_filter(logits, top_p), -1), 1, generator=g)
out = torch.cat([out, nxt], 1)
if eos is not None and int(nxt) == eos:
break
return out[0, ids.shape[1]:].tolist()
def _repeats(seq: list[int], tok: int, n: int) -> bool:
if n <= 0 or len(seq) < n - 1:
return False
tail = tuple(seq[len(seq) - (n - 1):] + [tok]) if n > 1 else (tok,)
return any(tuple(seq[i:i + n]) == tail for i in range(len(seq) - n + 1))
@torch.no_grad()
def beam(model, ids, new, eos, beams=4, no_repeat=3) -> list[int]:
"""Plain beam search: `beams` hypotheses, no repeated `no_repeat`-gram, best finished by mean log-probability."""
L0 = ids.shape[1]
hyps = [(ids[0].tolist(), 0.0)]
done: list[tuple[list[int], float]] = []
for _ in range(new):
batch = torch.tensor([h for h, _ in hyps], device=ids.device)
logp = F.log_softmax(next_logits(model, batch), -1)
cand = []
for b, (h, s) in enumerate(hyps):
vals, toks = logp[b].topk(4 * beams)
for v, t in zip(vals.tolist(), toks.tolist()):
if _repeats(h, t, no_repeat):
continue
cand.append((h + [t], s + v))
cand.sort(key=lambda x: x[1] / (len(x[0]) - L0), reverse=True)
hyps = []
for h, s in cand:
(done if eos is not None and h[-1] == eos else hyps).append((h, s))
if len(hyps) == beams:
break
if not hyps:
break
done += hyps
best = max(done, key=lambda x: x[1] / max(1, len(x[0]) - L0))
return best[0][L0:]
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--model", required=True)
ap.add_argument("--hf", action="store_true", help="a 🤗 causal LM instead of a perceiver_ar checkpoint")
ap.add_argument("--prompts", default="", help="JSON list of prompts (default: the card's three)")
ap.add_argument("--new", type=int, default=60, help="new tokens per reply")
ap.add_argument("--seeds", default="0,1,2")
ap.add_argument("--start", default="auto", choices=("auto", "none"),
help="auto: our checkpoints start with <|endoftext|> as in training; HF models as their tokenizer")
ap.add_argument("--device", default="cuda")
ap.add_argument("--out", required=True)
a = ap.parse_args()
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(a.model)
if a.hf:
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(a.model, dtype=torch.bfloat16).to(a.device).eval()
else:
from evaluation.lm_eval_perceiver_ar import load_perceiver_ar_for_eval
model = load_perceiver_ar_for_eval(a.model, device=a.device, attn_backend="sdpa")
prompts = json.loads(a.prompts) if a.prompts else PROMPTS
eos = tok.eos_token_id
lead = [eos] if (not a.hf and a.start == "auto") else []
seeds = [int(s) for s in a.seeds.split(",") if s]
res = {"model": a.model, "new_tokens": a.new, "start_tokens": lead,
"sample": {"temperature": 0.7, "top_p": 0.9, "repetition_penalty": 1.1, "seeds": seeds},
"beam": {"beams": 4, "no_repeat_ngram": 3}, "date_utc": time.strftime("%Y-%m-%d %H:%M UTC", time.gmtime()),
"prompts": []}
for p in prompts:
ids = torch.tensor([lead + tok(p, add_special_tokens=a.hf).input_ids], device=a.device)
row = {"prompt": p, "sample": {}, "beam": tok.decode(beam(model, ids, a.new, eos))}
for s in seeds:
row["sample"][s] = tok.decode(sample(model, ids, a.new, eos, s))
res["prompts"].append(row)
print(json.dumps(row, indent=1, ensure_ascii=False), flush=True)
Path(a.out).write_text(json.dumps(res, indent=2, ensure_ascii=False))
if __name__ == "__main__":
main()
|