Text Generation
Transformers
Safetensors
Turkish
qwen3
turkish
causal-lm
from-scratch
pretraining
text-generation-inference
Instructions to use eademir/TurkLM-scratch-0.5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use eademir/TurkLM-scratch-0.5B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="eademir/TurkLM-scratch-0.5B")# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("eademir/TurkLM-scratch-0.5B") model = AutoModelForCausalLM.from_pretrained("eademir/TurkLM-scratch-0.5B", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use eademir/TurkLM-scratch-0.5B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "eademir/TurkLM-scratch-0.5B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "eademir/TurkLM-scratch-0.5B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/eademir/TurkLM-scratch-0.5B
- SGLang
How to use eademir/TurkLM-scratch-0.5B 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 "eademir/TurkLM-scratch-0.5B" \ --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": "eademir/TurkLM-scratch-0.5B", "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 "eademir/TurkLM-scratch-0.5B" \ --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": "eademir/TurkLM-scratch-0.5B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use eademir/TurkLM-scratch-0.5B with Docker Model Runner:
docker model run hf.co/eademir/TurkLM-scratch-0.5B
File size: 6,187 Bytes
bfa0693 | 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 | """
Measure two efficiency metrics for the paper's efficiency analysis section:
1. Tokenizer fertility β avg tokens per word on Turkish text (lower = better compression)
2. Inference speed β tokens/second for autoregressive generation
Usage:
python eval/eval_efficiency.py --model-path ./ckpt_scratch_small
python eval/eval_efficiency.py --model-path Qwen/Qwen3-1.7B
python eval/eval_efficiency.py --model-path TURKCELL/Turkcell-LLM-7b-v1
"""
import argparse
import time
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset
# Fixed Turkish prompt for all models β kept short so it fits every tokenizer
PROMPT = (
"TΓΌrkiye, Avrupa ile Asya arasΔ±nda kΓΆprΓΌ gΓΆrevi gΓΆren eΕsiz bir coΔrafi konuma sahiptir. "
"Δ°stanbul, dΓΌnyanΔ±n iki kΔ±taya yayΔ±lan tek bΓΌyΓΌkΕehridir."
)
N_GENERATE = 100 # tokens to generate for speed test
N_WARMUP = 1 # warmup forward pass (discarded)
N_DOCS = 100 # Wikipedia docs for fertility measurement
SKIP_DOCS = 5_000
# ββ Tokenizer fertility ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def measure_fertility(tokenizer, texts):
total_tokens = 0
total_words = 0
for text in texts:
text = text.strip()
if not text:
continue
words = text.split()
tokens = tokenizer.encode(text, add_special_tokens=False)
total_words += len(words)
total_tokens += len(tokens)
return total_tokens / total_words if total_words else 0.0
# ββ Inference speed ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def measure_speed(model, tokenizer, prompt, device, n_generate=N_GENERATE, n_warmup=N_WARMUP):
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(device)
generate_kwargs = dict(
input_ids=input_ids,
max_new_tokens=n_generate,
do_sample=False, # greedy β deterministic, fair comparison
pad_token_id=tokenizer.eos_token_id,
)
# warmup
for _ in range(n_warmup):
with torch.no_grad():
model.generate(**generate_kwargs)
if device.type == "mps":
torch.mps.synchronize()
elif device.type == "cuda":
torch.cuda.synchronize()
start = time.perf_counter()
with torch.no_grad():
out = model.generate(**generate_kwargs)
if device.type == "mps":
torch.mps.synchronize()
elif device.type == "cuda":
torch.cuda.synchronize()
elapsed = time.perf_counter() - start
generated_tokens = out.shape[1] - input_ids.shape[1]
tokens_per_sec = generated_tokens / elapsed
ms_per_token = elapsed / generated_tokens * 1000
return tokens_per_sec, ms_per_token, generated_tokens
# ββ Main βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model-path", required=True)
parser.add_argument("--n-docs", type=int, default=N_DOCS)
parser.add_argument("--skip-docs", type=int, default=SKIP_DOCS)
parser.add_argument("--n-generate",type=int, default=N_GENERATE)
args = parser.parse_args()
device = torch.device(
"cuda" if torch.cuda.is_available() else
"mps" if torch.backends.mps.is_available() else
"cpu"
)
print(f"Device : {device}")
print(f"Loading model : {args.model_path}")
tokenizer = AutoTokenizer.from_pretrained(args.model_path)
model = AutoModelForCausalLM.from_pretrained(args.model_path, torch_dtype=torch.bfloat16)
model = model.to(device).eval()
n_params = sum(p.numel() for p in model.parameters())
print(f"Parameters : {n_params / 1e6:.0f}M")
# ββ Fertility ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\nLoading Turkish Wikipedia docs (skip={args.skip_docs}, n={args.n_docs}) ...")
ds = load_dataset("wikimedia/wikipedia", "20231101.tr", split="train", streaming=True)
texts = []
for i, ex in enumerate(ds):
if i < args.skip_docs:
continue
texts.append(ex["text"])
if len(texts) == args.n_docs:
break
fertility = measure_fertility(tokenizer, texts)
# ββ Speed ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"Measuring inference speed ({args.n_generate} tokens) ...")
tok_per_sec, ms_per_tok, n_gen = measure_speed(
model, tokenizer, PROMPT, device, n_generate=args.n_generate
)
# ββ Report βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
print(f"\n{'='*50}")
print(f"Model : {args.model_path}")
print(f"Parameters : {n_params / 1e6:.0f}M")
print(f"Vocab size : {len(tokenizer):,}")
print(f"β"*50)
print(f"Fertility : {fertility:.2f} tokens/word (lower = better)")
print(f"β"*50)
print(f"Inference speed : {tok_per_sec:.1f} tok/s")
print(f"Latency : {ms_per_tok:.1f} ms/token")
print(f"Tokens generated : {n_gen}")
print(f"{'='*50}")
if __name__ == "__main__":
main()
# The abandoned streaming dataset leaves pyarrow threads that deadlock at
# interpreter shutdown, which hangs run_all.sh. Results are already printed.
import os, sys
sys.stdout.flush()
sys.stderr.flush()
os._exit(0)
|