[Model] Added GLM-4 series hf format model support vllm==0.6.4 (#10561)

Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
sixgod
2024-11-28 14:53:31 +00:00
committed by GitHub
co-authored by Isotr0py Cyrus Leung
parent 3ed5e73146
commit 5fc5ce0fe4
5 changed files with 30 additions and 1 deletions
+21
View File
@@ -0,0 +1,21 @@
"""Inference-only HF format GLM-4 model compatible with THUDM weights."""
from vllm.config import VllmConfig
from vllm.model_executor.models.llama import LlamaForCausalLM
from .utils import PPMissingLayer
class GlmForCausalLM(LlamaForCausalLM):
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__(vllm_config=vllm_config, prefix=prefix)
# Hack Llama model to fit HF format GLM implementation
# Attention difference between GLM and Llama:
# 1. Half partial rotary_dim and no Neox style.
# 2. There is no bias for o_proj in attention
for layer in self.model.layers:
if not isinstance(layer, PPMissingLayer):
layer.self_attn.rotary_emb.rotary_dim //= 2
layer.self_attn.rotary_emb.is_neox_style = False
layer.self_attn.o_proj.bias = None
layer.self_attn.o_proj.skip_bias_add = True