Qwen/Qwen3-VL-4B-Instruct-FP8 AssertionError: Attempted to load weight (torch.Size([4096, 2560])) into parameter (torch.Size([6144, 2560]))

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/qwen3_vl.py", line 1602, in load_weights
[rank5]: return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 294, in load_weights
[rank5]: autoloaded_weights = set(self._load_module("", self.module, weights))
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 252, in _load_module
[rank5]: yield from self._load_module(prefix,
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 225, in _load_module
[rank5]: loaded_params = module_load_weights(weights)
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/qwen3.py", line 341, in load_weights
[rank5]: return loader.load_weights(weights)
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 294, in load_weights
[rank5]: autoloaded_weights = set(self._load_module("", self.module, weights))
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 252, in _load_module
[rank5]: yield from self._load_module(prefix,
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/utils.py", line 225, in _load_module
[rank5]: loaded_params = module_load_weights(weights)
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/models/qwen2.py", line 423, in load_weights
[rank5]: weight_loader(param, loaded_weight)
[rank5]: File "/home/team_cv/miniconda3/envs/megatron/lib/python3.11/site-packages/vllm/model_executor/model_loader/weight_utils.py", line 810, in default_weight_loader
[rank5]: assert param.size() == loaded_weight.size(), (
[rank5]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank5]: AssertionError: Attempted to load weight (torch.Size([4096, 2560])) into parameter (torch.Size([6144, 2560]))

How to Reproduce / 如何复现

Reproduce:
export NVTE_CUDA_INCLUDE_DIR=/usr/local/cuda/include
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True,garbage_collection_threshold:0.8'
OMP_NUM_THREADS=14
NPROC_PER_NODE=4
CUDA_VISIBLE_DEVICES=4,5,6,7
IMAGE_MAX_TOKEN_NUM=1024
VIDEO_MAX_TOKEN_NUM=128
FPS_MAX_FRAMES=16
MAX_PIXELS=602112
MASTER_PORT=29600
megatron rlhf
--rlhf_type grpo
--model Qwen/Qwen3-VL-4B-Instruct-FP8
--ref_load Qwen/Qwen3-VL-4B-Instruct-FP8
--load_safetensors true
--save_safetensors true
--fp8_recipe delayed
--fp8_format e4m3
--fp8_param_gather true
--context_parallel_size 1
--tensor_model_parallel_size 2
--pipeline_model_parallel_size 2
--dataset /home/team_cv/tdkien/kms-agentic-vtv/test_local/data_train_ocr.jsonl
--max_epochs 1
--global_batch_size 64
--micro_batch_size 1
--steps_per_generation 1
--num_generations 4
--external_plugins /home/team_cv/tdkien/ocr-training-model-vdu/rewards/plugin_glm.py
--reward_funcs external_ocr_classification_tag_reward
--use_vllm true
--vllm_mode colocate
--vllm_gpu_memory_utilization 0.3
--vllm_max_model_len 4096
--max_length 2048
--max_completion_length 2048
--lr 1e-6
--beta 0.001
--importance_sampling_level token
--epsilon 0.2
--epsilon_high 0.2
--dynamic_sample false
--overlong_filter true
--loss_type grpo
--sleep_level 2
--offload_model true
--offload_bridge false
--offload_optimizer true
--log_interval 1
--recompute_granularity selective
--finetune
--num_workers 8
--dataset_num_proc 8
--no_save_optim true
--no_save_rng true
--attention_backend flash
--temperature 1.0
--padding_free true
--log_completions true
--report_to wandb
--train_iters 100
--eval_interval 1000
--save_interval 1000
--use_hf true

Additional Information / 补充信息

No response

Start by reproducing the command with Qwen/Qwen3-VL-4B-Instruct-FP8, then inspect model_executor/models/qwen3_vl.py, qwen3.py, qwen2.py, and model_loader/weight_utils.py around load_weights and the reported assertion. Determine why the checkpoint shape differs from the parameter shape; done means the model loads successfully without this mismatch.