Skip to content

Fix loading Qwen VL fp4 ckpt #7301

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
wants to merge 4 commits into
base: main
Choose a base branch
from
Draft

Fix loading Qwen VL fp4 ckpt #7301

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
904bdd8
war
pamelap-nvidia Aug 27, 2025
b45aee6
fix
pamelap-nvidia Sep 6, 2025
e96dda5
fix
pamelap-nvidia Sep 7, 2025
3375b4c
fix test
pamelap-nvidia Sep 8, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
37 changes: 25 additions & 12 deletions tensorrt_llm/_torch/models/modeling_qwen2vl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,9 +4,12 @@

import torch
import torch.nn as nn
from transformers import (AutoProcessor, AutoTokenizer, PretrainedConfig,
PreTrainedModel, Qwen2_5_VLForConditionalGeneration,
Qwen2VLForConditionalGeneration)
from transformers import (AutoConfig, AutoProcessor, AutoTokenizer,
PretrainedConfig, PreTrainedModel)
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import \
Qwen2_5_VisionTransformerPretrainedModel
from transformers.models.qwen2_vl.modeling_qwen2_vl import \
Qwen2VisionTransformerPretrainedModel

from tensorrt_llm.inputs.multimodal import MultimodalParams

Expand Down Expand Up @@ -328,14 +331,22 @@ def __init__(self, model_config: ModelConfig[PretrainedConfig],
# TODO: Change the model class to TRT-LLM's Qwen2VisionModel
# Currently, copying vision encoder on all devices.
# NOTE: Using attn_implementation='flash_attention_2' to avoid the issue of vision model's GPU OOM.
model = model_class.from_pretrained(
model_path,
hf_model_config = AutoConfig.from_pretrained(model_path)
vision_model = model_class._from_config(
hf_model_config.vision_config,
torch_dtype=pretrained_config.torch_dtype,
attn_implementation='flash_attention_2').eval()
attn_implementation='flash_attention_2')
# TODO: Make vision model compatible with meta init mode and load_weights at the same place
self.visual = model.visual.to(self.device)
self.visual = vision_model.to(self.device)
self.post_config()

def load_weights(self, weights):
filtered_weights = {
k.replace('visual.', ''): v
for k, v in weights.items() if k.startswith('visual.')
}
self.visual.load_state_dict(filtered_weights)

def post_config(self):
self.config = self.visual.config

Expand Down Expand Up @@ -473,6 +484,7 @@ def init_rotary_cos_sin_ori(self):

def load_weights(self, weights):
self.llm.load_weights(weights)
self.mm_encoder.load_weights(weights)
self.init_rotary_cos_sin_ori()

def infer_max_seq_len(self) -> int:
Expand Down Expand Up @@ -626,7 +638,7 @@ def forward(


@register_vision_encoder(Qwen2VisionModelBase,
vlm_base_model=Qwen2VLForConditionalGeneration)
vlm_base_model=Qwen2VisionTransformerPretrainedModel)
@register_auto_model("Qwen2VLForConditionalGeneration")
@register_input_processor(
Qwen2VLInputProcessorBase,
Expand All @@ -645,11 +657,12 @@ def __init__(self, model_config: ModelConfig[PretrainedConfig], *args,
super().__init__(model_config, *args, **kwargs)
if not DISAGG:
self.mm_encoder = Qwen2VisionModelBase(
model_config, Qwen2VLForConditionalGeneration)
model_config, Qwen2VisionTransformerPretrainedModel)


@register_vision_encoder(Qwen2VisionModelBase,
vlm_base_model=Qwen2_5_VLForConditionalGeneration)
@register_vision_encoder(
Qwen2VisionModelBase,
vlm_base_model=Qwen2_5_VisionTransformerPretrainedModel)
@register_auto_model("Qwen2_5_VLForConditionalGeneration")
@register_input_processor(
Qwen2VLInputProcessorBase,
Expand All @@ -666,4 +679,4 @@ def __init__(self, model_config: ModelConfig[PretrainedConfig], *args,
super().__init__(model_config, *args, **kwargs)
if not DISAGG:
self.mm_encoder = Qwen2VisionModelBase(
model_config, Qwen2_5_VLForConditionalGeneration)
model_config, Qwen2_5_VisionTransformerPretrainedModel)
Loading