@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 382b6a4852
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, 07:00 UTC

[Misc] Avoid misleading warning messages (#10438)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

Jee Jee Li committed Nov 19, 2024 at 16:54 UTC 382b6a4852b9afc9a740b02736688e20f7d58446
2 files changed +7 -8
vllm/model_executor/models/chatglm.py
+2 -3
@@ -575,8 +575,7 @@ class ChatGLMModel(nn.Module):
575 return hidden_states
576
577
578 -class ChatGLMBaseModel(nn.Module, SupportsLoRA, SupportsPP,
579 - SupportsMultiModal):
578 +class ChatGLMBaseModel(nn.Module, SupportsLoRA, SupportsPP):
579
580 def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
581 super().__init__()
@@ -695,7 +694,7 @@ class ChatGLM(ChatGLMBaseModel):
694 embedding_padding_modules = []
695
696
698 -class ChatGLMV(ChatGLMBaseModel):
697 +class ChatGLMV(ChatGLMBaseModel, SupportsMultiModal):
698 packed_modules_mapping = {
699 "query_key_value": ["query_key_value"],
700 "dense_h_to_4h": ["dense_h_to_4h"],
vllm/model_executor/models/qwen.py
+5 -5
@@ -870,7 +870,7 @@ def dummy_data_for_qwen(
870 return DummyData(seq_data, mm_data)
871
872
873 -class QWenBaseModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
873 +class QWenBaseModel(nn.Module, SupportsPP, SupportsLoRA):
874
875 def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
876 super().__init__()
@@ -1024,7 +1024,7 @@ class QWenLLM(QWenBaseModel):
1024 embedding_padding_modules = []
1025
1026
1027 -class QWenVL(QWenBaseModel):
1027 +class QWenVL(QWenBaseModel, SupportsMultiModal):
1028 packed_modules_mapping = {
1029 "c_attn": ["c_attn"],
1030 "gate_up_proj": [
@@ -1062,7 +1062,7 @@ class QWenVL(QWenBaseModel):
1062 @MULTIMODAL_REGISTRY.register_max_image_tokens(MAX_QWEN_IMG_TOKENS)
1063 @INPUT_REGISTRY.register_dummy_data(dummy_data_for_qwen)
1064 @INPUT_REGISTRY.register_input_processor(input_processor_for_qwen)
1065 -class QWenLMHeadModel(QWenBaseModel, SupportsLoRA):
1065 +class QWenLMHeadModel(QWenBaseModel, SupportsMultiModal, SupportsLoRA):
1066 """
1067 QWenLMHeadModel is not only applicable to LLM but also to VL, which is not
1068 conducive to the current integration logic of LoRA in vLLM. Therefore, it
@@ -1083,7 +1083,7 @@ class QWenLMHeadModel(QWenBaseModel, SupportsLoRA):
1083 config = vllm_config.model_config.hf_config
1084 # Initialize VL
1085 if hasattr(config, "visual"):
1086 - return QWenVL(vllm_config=vllm_config)
1086 + return QWenVL(vllm_config=vllm_config, prefix=prefix)
1087 # Initialize LLM
1088 else:
1089 - return QWenLLM(vllm_config=vllm_config)
1089 + return QWenLLM(vllm_config=vllm_config, prefix=prefix)