[Misc] Load FP8 kv-cache scaling factors from checkpoints (#4893)

The 2nd PR for #4532.

This PR supports loading FP8 kv-cache scaling factors from a FP8 checkpoint (with .kv_scale parameter).

vllm/model_executor/models/baichuan.py

@@ -154,7 +154,8 @@ class BaiChuanAttention(nn.Module):
             self.attn = Attention(self.num_heads,
                                   self.head_dim,
                                   scaling,
-                                  alibi_slopes=alibi_slopes)
+                                  alibi_slopes=alibi_slopes,
+                                  quant_config=quant_config)
         else:
             self.rotary_emb = get_rope(
                 self.head_dim,
@@ -166,7 +167,8 @@ class BaiChuanAttention(nn.Module):
             self.attn = Attention(self.num_heads,
                                   self.head_dim,
                                   self.scaling,
-                                  cache_config=cache_config)
+                                  cache_config=cache_config,
+                                  quant_config=quant_config)
 
     def forward(
         self,