@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / a3e8a05d4c
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 23:15 UTC

[Bugfix] Fix KV head calculation for MPT models when using GQA (#5142)

Bruce Fontaine committed Jun 17, 2024 at 15:26 UTC a3e8a05d4c1b79dd44eb92bb6f57eb40c3fbdb21
1 file changed +5 -1
vllm/config.py
+5 -1
@@ -302,7 +302,11 @@ class ModelConfig:
302 return 1
303
304 # For DBRX and MPT
305 - if self.hf_config.model_type in ["dbrx", "mpt"]:
305 + if self.hf_config.model_type == "mpt":
306 + if "kv_n_heads" in self.hf_config.attn_config:
307 + return self.hf_config.attn_config["kv_n_heads"]
308 + return self.hf_config.num_attention_heads
309 + if self.hf_config.model_type == "dbrx":
310 return getattr(self.hf_config.attn_config, "kv_n_heads",
311 self.hf_config.num_attention_heads)
312