support mtp in v1_scheduler mode (#3695)

2025-10-04 16:22:57 +08:00 · 2025-09-04 17:39:59 +08:00
parent f265a26f8b
commit 88d44a2c93
11 changed files with 909 additions and 316 deletions
--- a/fastdeploy/config.py
+++ b/fastdeploy/config.py
@@ -891,7 +891,7 @@ class CacheConfig:
        else:
            self.kv_cache_ratio = 0.75
        self.enc_dec_block_num = 0 if current_platform.is_iluvatar() else 2
-        self.prealloc_dec_block_slot_num_threshold = 5
+        self.prealloc_dec_block_slot_num_threshold = 12
        self.cache_dtype = "bfloat16"
        self.model_cfg = None
        self.enable_chunked_prefill = False