Qwen3้ซๆๅพฎ่ฐ๏ผไธ๏ผ
่พๅ
ฅโ/โๅฟซ้ๆๅ
ฅๅ
ๅฎน
Qwen3้ซๆๅพฎ่ฐ๏ผไธ๏ผโ
Qwen3้ซๆๅพฎ่ฐ๏ผไธ๏ผโ
่ฏพ็จ่ฏดๆ๏ผโ
โข
ไฝ้ช่ฏพๅ
ๅฎน่้่ชใ2025ๅคงๆจกๅAgentๆบ่ฝไฝๅผๅๅฎๆใ(5ๆ็ญ)ๅฎๆด็ไป่ดน่ฏพ็จโ
โโไฝ้ช่ฏพๆถ้ดๆ้๏ผ่ฅๆณๆทฑๅบฆๅญฆไน ๅคงๆจกๅๆๆฏ๏ผๆฌข่ฟๅคงๅฎถๆฅๅ็ฑๆไธป่ฎฒ็ใ2025ๅคงๆจกๅAgentๆบ่ฝไฝๅผๅๅฎๆใ(5ๆ็ญ)โ
ๆญคๅค๏ผๅ
ฌๅผ่ฏพๅ
จๅฅๅญฆไน ่ตๆ๏ผๅทฒไธไผ ่ณ็ฝ็โ
้่ฆๆด็ณป็ปๆทฑๅ
ฅๅญฆไน ๅคงๆจกๅๅฏๆซ็ โฌ๏ธๆทปๅ ๅฉๆๅจ่ฏขๅ๏ฝโ
Qwen3้ซๆๅพฎ่ฐๅฎๆ๏ผไธ๏ผโ
็บฏไปฃ็ ็โ
โข
ๅฏผๅ
ฅๆจกๅโ
ไปฃ็ ๅโ
from unsloth import FastLanguageModelโ
import torchโ
ไปฃ็ ๅโ
max_seq_length = 2048 โ
dtype = None โ
load_in_4bit = Trueโ
ไปฃ็ ๅโ
model, tokenizer = FastLanguageModel.from_pretrained(โ
model_name = "./Qwen3-32B-unsloth-bnb-4bit",โ
max_seq_length = max_seq_length,โ
dtype = dtype,โ
load_in_4bit = load_in_4bit,โ
)โ
ไปฃ็ ๅโ
==((====))== Unsloth 2025.4.7: Fast Qwen3 patching. Transformers: 4.51.3.โ
\\ /| NVIDIA H800 PCIe. Num GPUs = 1. Max memory: 79.205 GB. Platform: Linux.โ
O^O/ \_/ \ Torch: 2.7.0+cu126. CUDA: 9.0. CUDA Toolkit: 12.6. Triton: 3.3.0โ
\ / Bfloat16 = TRUE. FA [Xformers = 0.0.30. FA2 = False]โ
"-____-" Free license: http://github.com/unslothai/unslothโ
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!โ
โ
โ
Loading checkpoint shards: 100%|โโโโโโโโโโ| 9/9 [00:14<00:00, 1.63s/it]โ
ไปฃ็ ๅโ
modelโ
ไปฃ็ ๅโ
Qwen3ForCausalLM(โ
(model): Qwen3Model(โ
(embed_tokens): Embedding(151936, 5120, padding_idx=151654)โ
(layers): ModuleList(โ
(0-5): 6 x Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear4bit(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(6): Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(7): Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(8-22): 15 x Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear4bit(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(23-44): 22 x Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(45): Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(46-54): 9 x Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(55-61): 7 x Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear4bit(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(62): Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
(63): Qwen3DecoderLayer(โ
(self_attn): Qwen3Attention(โ
(q_proj): Linear4bit(in_features=5120, out_features=8192, bias=False)โ
(k_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(v_proj): Linear4bit(in_features=5120, out_features=1024, bias=False)โ
(o_proj): Linear4bit(in_features=8192, out_features=5120, bias=False)โ
(q_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(k_norm): Qwen3RMSNorm((128,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(mlp): Qwen3MLP(โ
(gate_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(up_proj): Linear4bit(in_features=5120, out_features=25600, bias=False)โ
(down_proj): Linear4bit(in_features=25600, out_features=5120, bias=False)โ
(act_fn): SiLU()โ
)โ
(input_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(post_attention_layernorm): Qwen3RMSNorm((5120,), eps=1e-06)โ
)โ
)โ
(norm): Qwen3RMSNorm((5120,), eps=1e-06)โ
(rotary_emb): LlamaRotaryEmbedding()โ
)โ
(lm_head): Linear(in_features=5120, out_features=151936, bias=False)โ
)โ