data flow

input_ids = tokenizer.build_chat_input(query, history=history, role=role)

input_ids -> [1, 6] inputs_embeds -> [6, 1, 4096] 4096:hidden_size rotary_pos_emb -> [6, 1, 32, 2] 32:pos的编码维度 2:cos+sin

hidden_states = inputs_embeds for layers : GLMBlock(hidden_states, rotary_pos_emb) hidden_states = self.final_layernorm(hidden_states) hidden_states = hidden_states[-1:] lm_logits = self.output_layer(hidden_states) lm_logits = lm_logits.transpose(0, 1).contiguous() -> [1, 1, 65024]

probs = softmax(lm_logits) -> [1, 65024] next_tokens = torch.multinomial(probs, num_samples=1) 采样 -> [1] input_ids = torch.cat([input_ids, next_tokens) -> [1, 7]

response = tokenizer.decode(outputs)