GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SPAISS6F1
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
SPAISS6F1
Model Tree
Input Modalities
Output Modalities
Supported Functionality
ตัด decoder layer ตรงกลางทิ้ง (มักทำงานซ้ำซ้อน) เก็บเฉพาะ layer หัว (เข้าใจ input) และ layer ท้าย (สร้าง output) — embedding / lm_head / norm คงเดิม จึงไม่พัง dimension
python
import torchfrom transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B", torch_dtype=torch.bfloat16) # หา text-decoder layer list (เลี่ยง vision encoder กรณี multimodal)holder, layers = None, Nonefor _, mod in model.named_modules(): L = getattr(mod, "layers", None) if isinstance(L, torch.nn.ModuleList) and len(L) and hasattr(L[0], "self_attn"): holder, layers = mod, L if "language" in _.lower() or "text" in _.lower(): breakN = len(layers) # เก็บ 18 จาก 36 layers: หัว + ท้ายkeep = [0,1,2,3,4,5,6,7,8, 27,28,29,30,31,32,33,34,35]holder.layers = torch.nn.ModuleList([layers[i] for i in keep]) # อัปเดต config (รองรับ nested text_config ของ Gemma3)for c in {model.config, getattr(model.config, "text_config", model.config)}: if getattr(c, "num_hidden_layers", None) is not None: c.num_hidden_layers = len(keep) lt = getattr(c, "layer_types", None) if isinstance(lt, list) and len(lt) == N: c.layer_types = [lt[i] for i in keep] # reindex layer_idx ของแต่ละ block (สำคัญต่อ KV cache)for i, lyr in enumerate(holder.layers): if hasattr(lyr, "self_attn") and hasattr(lyr.self_attn, "layer_idx"): lyr.self_attn.layer_idx = iผลลัพธ์: 3.09B -> 1.70B (ยังไม่ถึง 1B เป๊ะ เพราะ embedding+lm_head+vocab ไม่ลดตาม layer)
หลัง prune โมเดลจะพ่น gibberish ทันที (เส้นประสาทถูกตัดขาด) -> ต้อง Healing ต่อ
เทรนต่อด้วย causal-LM บน Thai corpus เพื่อให้ layer ที่เหลือกลับมาทำงานร่วมกัน
python
from transformers import Trainer, TrainingArguments, DataCollatorForLanguageModelingfrom datasets import load_dataset ds = load_dataset("SPAISS6F1/spai-ss6-llm-1b-thai-corpus", split="train") # หรือ SEA-PILE v2 'th'tds = ds.map(lambda e: tok(e["text"], truncation=True, max_length=1024), batched=True, remove_columns=ds.column_names) model.gradient_checkpointing_enable(); model.config.use_cache = Falseargs = TrainingArguments(output_dir="out", num_train_epochs=2, per_device_train_batch_size=4, gradient_accumulation_steps=4, learning_rate=1e-4, lr_scheduler_type="cosine", warmup_ratio=0.03, bf16=True)Trainer(model=model, args=args, train_dataset=tds, data_collator=DataCollatorForLanguageModeling(tok, mlm=False)).train()Hyperparameters:
1e-4 (สูงกว่าปกติเพื่อสมานแผล) | Epochs: 2python
model.config.use_cache = Truetry: model.save_pretrained("out", safe_serialization=True)except RuntimeError: # Gemma3: tied embeddings -> fallback .bin model.save_pretrained("out", safe_serialization=False)โมเดลนี้ save เป็น: model.safetensors
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerm = "SPAISS6F1/qwen-1b-pruned-th"tok = AutoTokenizer.from_pretrained(m)model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.bfloat16, device_map="cuda")ids = tok("ปัญญาประดิษฐ์ คือ", return_tensors="pt").to(model.device)out = model.generate(**ids, max_new_tokens=120, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.3)print(tok.decode(out[0], skip_special_tokens=True))repetition_penalty >= 1.2 กันการวนซ้ำimport torchfrom transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B", torch_dtype=torch.bfloat16) # หา text-decoder layer list (เลี่ยง vision encoder กรณี multimodal)holder, layers = None, Nonefor _, mod in model.named_modules(): L = getattr(mod, "layers", None) if isinstance(L, torch.nn.ModuleList) and len(L) and hasattr(L[0], "self_attn"): holder, layers = mod, L if "language" in _.lower() or "text" in _.lower(): breakN = len(layers) # เก็บ 18 จาก 36 layers: หัว + ท้ายkeep = [0,1,2,3,4,5,6,7,8, 27,28,29,30,31,32,33,34,35]holder.layers = torch.nn.ModuleList([layers[i] for i in keep]) # อัปเดต config (รองรับ nested text_config ของ Gemma3)for c in {model.config, getattr(model.config, "text_config", model.config)}: if getattr(c, "num_hidden_layers", None) is not None: c.num_hidden_layers = len(keep) lt = getattr(c, "layer_types", None) if isinstance(lt, list) and len(lt) == N: c.layer_types = [lt[i] for i in keep] # reindex layer_idx ของแต่ละ block (สำคัญต่อ KV cache)for i, lyr in enumerate(holder.layers): if hasattr(lyr, "self_attn") and hasattr(lyr.self_attn, "layer_idx"): lyr.self_attn.layer_idx = ifrom transformers import Trainer, TrainingArguments, DataCollatorForLanguageModelingfrom datasets import load_dataset ds = load_dataset("SPAISS6F1/spai-ss6-llm-1b-thai-corpus", split="train") # หรือ SEA-PILE v2 'th'tds = ds.map(lambda e: tok(e["text"], truncation=True, max_length=1024), batched=True, remove_columns=ds.column_names) model.gradient_checkpointing_enable(); model.config.use_cache = Falseargs = TrainingArguments(output_dir="out", num_train_epochs=2, per_device_train_batch_size=4, gradient_accumulation_steps=4, learning_rate=1e-4, lr_scheduler_type="cosine", warmup_ratio=0.03, bf16=True)Trainer(model=model, args=args, train_dataset=tds, data_collator=DataCollatorForLanguageModeling(tok, mlm=False)).train()model.config.use_cache = Truetry: model.save_pretrained("out", safe_serialization=True)except RuntimeError: # Gemma3: tied embeddings -> fallback .bin model.save_pretrained("out", safe_serialization=False)import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerm = "SPAISS6F1/qwen-1b-pruned-th"tok = AutoTokenizer.from_pretrained(m)model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.bfloat16, device_map="cuda")ids = tok("ปัญญาประดิษฐ์ คือ", return_tensors="pt").to(model.device)out = model.generate(**ids, max_new_tokens=120, do_sample=True, temperature=0.7, top_p=0.9, repetition_penalty=1.3)print(tok.decode(out[0], skip_special_tokens=True))