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 15, 13:30 UTC
[Bugfix] Fix GLM rotary_dim issue and support v1 (#16912)
Signed-off-by: isotr0py <2037008807@qq.com>
Isotr0py committed
Apr 21, 2025 at 22:26 UTC
55d6d3fdb8a08aec7c0b09be95c4982705921ab3
1 file changed
+2
-3
vllm/model_executor/models/glm.py
+2
-3
@@ -3,13 +3,13 @@
3
from vllm.config import VllmConfig
4
from vllm.model_executor.models.llama import LlamaForCausalLM
5
6
-from .interfaces import SupportsV0Only
6
from .utils import PPMissingLayer
7
8
10
-class GlmForCausalLM(LlamaForCausalLM, SupportsV0Only):
9
+class GlmForCausalLM(LlamaForCausalLM):
10
11
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
12
+ vllm_config.model_config.hf_config.partial_rotary_factor = 0.5
13
super().__init__(vllm_config=vllm_config, prefix=prefix)
14
# Hack Llama model to fit HF format GLM implementation
15
# Attention difference between GLM and Llama:
@@ -17,7 +17,6 @@ class GlmForCausalLM(LlamaForCausalLM, SupportsV0Only):
17
# 2. There is no bias for o_proj in attention
18
for layer in self.model.layers:
19
if not isinstance(layer, PPMissingLayer):
20
- layer.self_attn.rotary_emb.rotary_dim //= 2
20
layer.self_attn.rotary_emb.is_neox_style = False
21
layer.self_attn.o_proj.bias = None
22
layer.self_attn.o_proj.skip_bias_add = True