C10X/6.4m-falcon-tokenizer-hybrid-sw768-v2

Qwen3 6.4M Model with Falcon-H1-0.5B-Instruct Tokenizer

Model Description

This is a 6.4M parameter Qwen3 model architecture combined with the Falcon-H1-0.5B-Instruct tokenizer (32K vocabulary).

• Architecture: Qwen3 (Grouped Query Attention, RMS Normalization, Q/K Normalization, RoPE)
• Tokenizer: Falcon-H1-0.5B-Instruct (32K vocab)
• Parameters: 6,393,440
• Precision: BF16
• Format: SafeTensors
• Vocabulary Size: 32768
• Use Case: Lightweight model with hybrid sliding window

Configuration

• vocab_size: 32768
• hidden_size: 96
• num_attention_heads: 8
• num_key_value_heads: 4
• num_hidden_layers: 8
• intermediate_size: 384
• head_dim: 128
• max_position_embeddings: 8192

Special Tokens

• BOS: <|begin_of_text|> (id: 17)
• EOS: <|end_of_text|> (id: 11)
• PAD: <|pad|> (id: 32767)

Usage

from transformers import Qwen3ForCausalLM, AutoTokenizer

model = Qwen3ForCausalLM.from_pretrained("./workspace/6.4m-falcon-tokenizer")
tokenizer = AutoTokenizer.from_pretrained("./workspace/6.4m-falcon-tokenizer")

# Generate text
inputs = tokenizer("Hello, world!", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Batch processing (start small)
texts = ["Hello", "How are you", "Good morning"]
inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=20)

Important Notes

• Model uses Qwen3 architecture with Falcon tokenizer (32K vocabulary)
• All token IDs must be < 32768 to avoid CUDA errors
• Start with small batch sizes (1-4) and gradually increase
• Use proper padding to prevent dimension mismatches
• Model initialized with random weights - requires fine-tuning
• Compatible with Qwen3 APIs but uses Falcon vocabulary