Instructions to use ecloudtech/Erk-32B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ecloudtech/Erk-32B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ecloudtech/Erk-32B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ecloudtech/Erk-32B")
model = AutoModelForCausalLM.from_pretrained("ecloudtech/Erk-32B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ecloudtech/Erk-32B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ecloudtech/Erk-32B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ecloudtech/Erk-32B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ecloudtech/Erk-32B

SGLang

How to use ecloudtech/Erk-32B 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 "ecloudtech/Erk-32B" \
    --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": "ecloudtech/Erk-32B",
		"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 "ecloudtech/Erk-32B" \
        --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": "ecloudtech/Erk-32B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ecloudtech/Erk-32B with Docker Model Runner:
```
docker model run hf.co/ecloudtech/Erk-32B
```

Erk-32B / betikler /eval32b_heldout.py

aigencydev

Erk-32B: model karti, olcum protokolu, betikler, sonuc vektorleri

9c90564 verified about 1 month ago

Raw History Blame Contribute Delete

4.48 kB

"""Erk-32B ara degerlendirme — HELD-OUT 750 (probe'un 150 sorusu HARIC).
Mevcut LoRA checkpoint'i taban modele takip gercek sayiyi olcer."""
import os, json, math, torch, torch.nn as nn
os.environ["HF_DATASETS_OFFLINE"]="1"
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset

MID="/ari/users/ytanis/qwen3-32b"
# Olculecek checkpoint dizini disaridan verilebilir (CPT / SFT ayni betikle
# ayni protokolde olculsun diye).
OUT=os.environ.get("CK_DIR", "/ari/users/ytanis/erk32b-cpt")
R=256; ALPHA=512
tok=AutoTokenizer.from_pretrained(MID)
model=AutoModelForCausalLM.from_pretrained(MID, torch_dtype=torch.bfloat16, device_map="auto").eval()

class LoRA(nn.Module):
    def __init__(self, base, r=R, alpha=ALPHA):
        super().__init__(); self.base=base
        d=base.weight.device
        self.A=nn.Parameter(torch.zeros(r, base.in_features, device=d, dtype=torch.bfloat16))
        self.B=nn.Parameter(torch.zeros(base.out_features, r, device=d, dtype=torch.bfloat16))
        self.s=alpha/r
    def forward(self,x):
        return self.base(x)+((x@self.A.t())@self.B.t())*self.s
for layer in model.model.layers:
    for mod,names in ((layer.self_attn,["q_proj","k_proj","v_proj","o_proj"]),
                      (layer.mlp,["gate_proj","up_proj","down_proj"])):
        for nm in names: setattr(mod,nm,LoRA(getattr(mod,nm)))

st=json.load(open(f"{OUT}/state.json"))
lsd=torch.load(f"{OUT}/lora_ck.pt", map_location="cpu", weights_only=True)
named=dict(model.named_parameters()); n=0
for k,v in lsd.items():
    if k in named:
        with torch.no_grad(): named[k].copy_(v.to(named[k].device))
        n+=1
print(f"LoRA yuklendi: {n} tensor | adim {st['step']} ({st['tokens']/1e9:.2f}B token)", flush=True)

ds=load_dataset("AYueksel/TurkishMMLU","All",split="test")
letters="ABCDE"; lid=[tok(" "+c, add_special_tokens=False).input_ids[-1] for c in letters]
probe=set(list(range(1,len(ds),6))[:150])
held=[i for i in range(min(900,len(ds))) if i not in probe]
# --- Olcum protokolu (adim 23500'de sabitlendi, olculdu):
#   1) probe'un 150 sorusu haric  -> secim sizintisi yok
#   2) kirli sorular haric        -> kulliyatta 13-kelime birebir gecenler
#      (held-out 750'nin 98'i; bkz. kirlilik_CPT.json)
#   3) sik sirasi dondurulur      -> harf/konum onyargisi elenir
# Bu ucu birden uygulanmadan olculen fark ~2 puan sisiyor (6.27 -> 4.14).
KIR=set()
try:
    KIR=set(json.load(open("/ari/users/ytanis/kirlilik_CPT.json"))["kirli_13"])
except Exception as e:
    print("UYARI: kirlilik listesi okunamadi (%s) — temiz alt kume raporlanmayacak" % e, flush=True)
temiz=[i for i in held if i not in KIR]

def run(kaydir):
    v=[]
    with torch.no_grad():
        for i in held:
            ex=ds[i]
            try:
                a=ex["answer"]; ans=int(a) if isinstance(a,int) else letters.index(str(a).strip())
            except: v.append(None); continue
            ch=list(ex["choices"])[:5]; k=len(ch)
            if kaydir:
                ch=ch[-kaydir:]+ch[:-kaydir]; ans=(ans+kaydir)%k
            s="Asagidaki coktan secmeli soruyu cevapla.\n\nSoru: "+str(ex["question"])+"\n"
            for j,c in enumerate(ch): s+=f"{letters[j]}) {c}\n"
            ids=tok(s+"Cevap:", return_tensors="pt", truncation=True, max_length=1024).input_ids.to(model.device)
            lg=model(ids).logits[0,-1]
            v.append(int(int(torch.stack([lg[t] for t in lid[:k]]).argmax())==ans))
    return v

def oran(v, idx):
    z=[v[held.index(i)] for i in idx if v[held.index(i)] is not None]
    return 100*sum(z)/max(len(z),1), len(z)

sonuc={"step":st["step"],"tokens":st["tokens"]}
# Asama etiketi checkpoint dizininden turetilir; sabit "CPT" yazmak SFT
# olcumlerinde yaniltici log uretiyordu.
ASAMA = "SFT" if OUT.rstrip("/").endswith("sft") else "CPT"
print(f"\n=== ERK-32B ({ASAMA}, adim {st['step']}, {st['tokens']/1e9:.2f}B token) ===", flush=True)
for ad,kay in (("orijinal",0),("dondurulmus",1)):
    v=run(kay); sonuc["vec_"+ad]=v
    a750,n750=oran(v,held); atem,ntem=oran(v,temiz)
    sonuc[ad]={"held750":a750,"temiz":atem,"n_temiz":ntem}
    print(f"  {ad:12s} held-out {n750}: %{a750:.2f} | TEMIZ {ntem}: %{atem:.2f}", flush=True)
print(f"\nRAPORLANACAK SAYI (temiz + dondurulmus): %{sonuc['dondurulmus']['temiz']:.2f}", flush=True)
print(f"  stok Qwen3-32B ayni kosulda: %67.64 (adim 23500 olcumunden)", flush=True)
json.dump(sonuc, open(f"{OUT}/heldout_step{st['step']}.json","w"))
print("BITTI", flush=True)

	"""Erk-32B ara degerlendirme — HELD-OUT 750 (probe'un 150 sorusu HARIC).
	Mevcut LoRA checkpoint'i taban modele takip gercek sayiyi olcer."""
	import os, json, math, torch, torch.nn as nn
	os.environ["HF_DATASETS_OFFLINE"]="1"
	from transformers import AutoModelForCausalLM, AutoTokenizer
	from datasets import load_dataset

	MID="/ari/users/ytanis/qwen3-32b"
	# Olculecek checkpoint dizini disaridan verilebilir (CPT / SFT ayni betikle
	# ayni protokolde olculsun diye).
	OUT=os.environ.get("CK_DIR", "/ari/users/ytanis/erk32b-cpt")
	R=256; ALPHA=512
	tok=AutoTokenizer.from_pretrained(MID)
	model=AutoModelForCausalLM.from_pretrained(MID, torch_dtype=torch.bfloat16, device_map="auto").eval()

	class LoRA(nn.Module):
	def __init__(self, base, r=R, alpha=ALPHA):
	super().__init__(); self.base=base
	d=base.weight.device
	self.A=nn.Parameter(torch.zeros(r, base.in_features, device=d, dtype=torch.bfloat16))
	self.B=nn.Parameter(torch.zeros(base.out_features, r, device=d, dtype=torch.bfloat16))
	self.s=alpha/r
	def forward(self,x):
	return self.base(x)+((x@self.A.t())@self.B.t())*self.s
	for layer in model.model.layers:
	for mod,names in ((layer.self_attn,["q_proj","k_proj","v_proj","o_proj"]),
	(layer.mlp,["gate_proj","up_proj","down_proj"])):
	for nm in names: setattr(mod,nm,LoRA(getattr(mod,nm)))

	st=json.load(open(f"{OUT}/state.json"))
	lsd=torch.load(f"{OUT}/lora_ck.pt", map_location="cpu", weights_only=True)
	named=dict(model.named_parameters()); n=0
	for k,v in lsd.items():
	if k in named:
	with torch.no_grad(): named[k].copy_(v.to(named[k].device))
	n+=1
	print(f"LoRA yuklendi: {n} tensor \| adim {st['step']} ({st['tokens']/1e9:.2f}B token)", flush=True)

	ds=load_dataset("AYueksel/TurkishMMLU","All",split="test")
	letters="ABCDE"; lid=[tok(" "+c, add_special_tokens=False).input_ids[-1] for c in letters]
	probe=set(list(range(1,len(ds),6))[:150])
	held=[i for i in range(min(900,len(ds))) if i not in probe]
	# --- Olcum protokolu (adim 23500'de sabitlendi, olculdu):
	# 1) probe'un 150 sorusu haric -> secim sizintisi yok
	# 2) kirli sorular haric -> kulliyatta 13-kelime birebir gecenler
	# (held-out 750'nin 98'i; bkz. kirlilik_CPT.json)
	# 3) sik sirasi dondurulur -> harf/konum onyargisi elenir
	# Bu ucu birden uygulanmadan olculen fark ~2 puan sisiyor (6.27 -> 4.14).
	KIR=set()
	try:
	KIR=set(json.load(open("/ari/users/ytanis/kirlilik_CPT.json"))["kirli_13"])
	except Exception as e:
	print("UYARI: kirlilik listesi okunamadi (%s) — temiz alt kume raporlanmayacak" % e, flush=True)
	temiz=[i for i in held if i not in KIR]

	def run(kaydir):
	v=[]
	with torch.no_grad():
	for i in held:
	ex=ds[i]
	try:
	a=ex["answer"]; ans=int(a) if isinstance(a,int) else letters.index(str(a).strip())
	except: v.append(None); continue
	ch=list(ex["choices"])[:5]; k=len(ch)
	if kaydir:
	ch=ch[-kaydir:]+ch[:-kaydir]; ans=(ans+kaydir)%k
	s="Asagidaki coktan secmeli soruyu cevapla.\n\nSoru: "+str(ex["question"])+"\n"
	for j,c in enumerate(ch): s+=f"{letters[j]}) {c}\n"
	ids=tok(s+"Cevap:", return_tensors="pt", truncation=True, max_length=1024).input_ids.to(model.device)
	lg=model(ids).logits[0,-1]
	v.append(int(int(torch.stack([lg[t] for t in lid[:k]]).argmax())==ans))
	return v

	def oran(v, idx):
	z=[v[held.index(i)] for i in idx if v[held.index(i)] is not None]
	return 100*sum(z)/max(len(z),1), len(z)

	sonuc={"step":st["step"],"tokens":st["tokens"]}
	# Asama etiketi checkpoint dizininden turetilir; sabit "CPT" yazmak SFT
	# olcumlerinde yaniltici log uretiyordu.
	ASAMA = "SFT" if OUT.rstrip("/").endswith("sft") else "CPT"
	print(f"\n=== ERK-32B ({ASAMA}, adim {st['step']}, {st['tokens']/1e9:.2f}B token) ===", flush=True)
	for ad,kay in (("orijinal",0),("dondurulmus",1)):
	v=run(kay); sonuc["vec_"+ad]=v
	a750,n750=oran(v,held); atem,ntem=oran(v,temiz)
	sonuc[ad]={"held750":a750,"temiz":atem,"n_temiz":ntem}
	print(f" {ad:12s} held-out {n750}: %{a750:.2f} \| TEMIZ {ntem}: %{atem:.2f}", flush=True)
	print(f"\nRAPORLANACAK SAYI (temiz + dondurulmus): %{sonuc['dondurulmus']['temiz']:.2f}", flush=True)
	print(f" stok Qwen3-32B ayni kosulda: %67.64 (adim 23500 olcumunden)", flush=True)
	json.dump(sonuc, open(f"{OUT}/heldout_step{st['step']}.json","w"))
	print("BITTI", flush=True)