IQuestLab

/

IQuest-Coder-V1-40B-Loop-Instruct

like 289

Follow

IQuest 407

I'm trying to understand the caching behavior in modeling_iquestloopcoder.py and noticed a difference between prefill and generation:

Prefill (_forward_loop, lines 1072-1074):

hidden_states, gate_mean = decoder_layer.forward_loop2_mixed(...)
if use_cache and loop_idx == 2:
    hidden_states_normed = decoder_layer.input_layernorm(hidden_states)
    _, k2, v2 = decoder_layer.self_attn.get_qkv(hidden_states_normed, position_ids)

Here hidden_states is the layer OUTPUT (after attention + MLP).

Generation (forward_decode_loop2, line 630):

q2, k2, v2 = self.get_qkv(hidden_states, position_ids)

Here hidden_states is the layer INPUT (before attention).

Is the difference in prefill intentional, or should K2/V2 be computed from the same source in both paths? I ran some tests comparing both approaches against full recomputation and found that INPUT-based K2 matches exactly, while OUTPUT-based differs slightly. However, the practical impact seems minimal since the gates strongly favor global attention (~87%). I'm curious whether the difference is intentional or an oversight.