Fix bias='lora_only' state-dict collection in train_lora

[Chessing234]

Bug

In fastchat/train/train_lora.py, get_peft_state_maybe_zero_3 crashes
or silently stores bias params under the wrong key whenever bias == \"lora_only\" is used and the model has any bias parameter.

Root cause

elif bias == \"lora_only\":
    to_return = {}
    maybe_lora_bias = {}
    lora_bias_names = set()
    for k, t in named_params:
        if \"lora_\" in k:
            to_return[k] = t
            bias_name = k.split(\"lora_\")[0] + \"bias\"
            lora_bias_names.add(bias_name)
        elif \"bias\" in k:
            maybe_lora_bias[k] = t
    for k, t in maybe_lora_bias:                 # (1)
        if bias_name in lora_bias_names:         # (2)
            to_return[bias_name] = t             # (3)

1. for k, t in maybe_lora_bias: iterates dict keys, not items.
   Unpacking a string key into k, t raises
   ValueError: too many values to unpack the first time the dict is
   non-empty.
2. bias_name leaks from the previous loop – it's always the last
   lora-derived bias name – so the membership check is independent of
   the current bias key k.
3. to_return[bias_name] = t writes the tensor t under that leaked
   name, not k, so the saved bias param has the wrong key (or
   overwrites whatever was saved there).

Why the fix is correct

peft's reference implementation (the comment says this helper was
"borrowed from peft.utils.get_peft_model_state_dict") matches bias
keys directly against the current loop variable, not a leaked one.
Using maybe_lora_bias.items() and k makes the three lines self
consistent and produces the same behaviour as peft's lora_only branch.

Change

fastchat/train/train_lora.py:

-        for k, t in maybe_lora_bias:
-            if bias_name in lora_bias_names:
-                to_return[bias_name] = t
+        for k, t in maybe_lora_bias.items():
+            if k in lora_bias_names:
+                to_return[k] = t