GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
User01110
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
User01110
Model Tree
Input Modalities
Output Modalities
Supported Functionality
<|im_start|> / <|im_end|> markersStreaming-only training mix:
Documents are tokenized with <bos> at the start and <eos> at the end. Fixed training blocks additionally use <|im_start|> and <|im_end|> as sequence-level markers so artificial block boundaries are distinct from document boundaries.
Final validation is computed after training on Salesforce/wikitext, config wikitext-103-raw-v1, split validation, reporting NLL, perplexity, and UTF-8 bits per byte.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer HF_token = "" # optional for public inference; paste token for higher HF rate limitsrepo = "User01110/tinyML-1.1-3M-exp"tok = AutoTokenizer.from_pretrained(repo, token=HF_token or None, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto", token=HF_token or None, trust_remote_code=True,)model.eval() im_start = tok.convert_tokens_to_ids("<|im_start|>")assert im_start != tok.unk_token_id, "missing <|im_start|> token"prompt = "The future of Artificial Intelligence is"device = next(model.parameters()).deviceids = torch.tensor( [[im_start, tok.bos_token_id] + tok.encode(prompt, add_special_tokens=False)], device=device,)mask = torch.ones_like(ids)with torch.no_grad(): out = model.generate( ids, attention_mask=mask, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.9, top_k=50, repetition_penalty=1.2, pad_token_id=tok.eos_token_id, eos_token_id=tok.eos_token_id, use_cache=False, )print(tok.decode(out[0], skip_special_tokens=True))import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer HF_token = "" # optional for public inference; paste token for higher HF rate limitsrepo = "User01110/tinyML-1.1-3M-exp"tok = AutoTokenizer.from_pretrained(repo, token=HF_token or None, trust_remote_code=True)model = AutoModelForCausalLM.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto", token=HF_token or None, trust_remote_code=True,)model.eval() im_start = tok.convert_tokens_to_ids("<|im_start|>")assert im_start != tok.unk_token_id, "missing <|im_start|> token"prompt = "The future of Artificial Intelligence is"device = next(model.parameters()).deviceids = torch.tensor( [[im_start, tok.bos_token_id] + tok.encode(prompt, add_special_tokens=False)], device=device,)mask = torch.ones_like(ids)with torch.no_grad(): out = model.generate( ids, attention_mask=mask, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.9, top_k=50, repetition_penalty=1.2, pad_token_id=tok.eos_token_id, eos_token_id=tok.eos_token_id, use_cache=False, )print(tok.decode(out[0], skip_special_tokens=True))