chatlearn/models/vllm/hooks/vllm_0_6_3/loader.py [36:74]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    convert_llama_state_dict_from_megatron_to_vllm,
    convert_llama_state_dict_from_mcore_to_vllm,
    convert_qwen_state_dict_from_megatron_to_vllm,
    load_checkpoint
)

def load_weights(self, model_args):
    torch.distributed.barrier()
    self.model_args = model_args
    load_checkpoint(self, None, None, model_args=model_args)
    torch.distributed.barrier()

def load_state_dict(self, state_dict, strict=True, assign=False):
    qwen_version = None
    if isinstance(self, LlamaForCausalLM):
        use_legacy_models = get_use_legacy_models(self.model_args)
        if use_legacy_models:
            convert_state_dict_internal = convert_llama_state_dict_from_megatron_to_vllm
        else:
            convert_state_dict_internal = convert_llama_state_dict_from_mcore_to_vllm
    elif isinstance(self, QWenLMHeadModel):
        qwen_version = 1.0
        convert_state_dict_internal = convert_qwen_state_dict_from_megatron_to_vllm
    elif isinstance(self, Qwen2ForCausalLM) or (Qwen2MoeForCausalLM is not None and isinstance(self, Qwen2MoeForCausalLM)):
        qwen_version = 2.0
        convert_state_dict_internal = convert_qwen_state_dict_from_megatron_to_vllm
    else:
        raise RuntimeError(f"Unsupported model for vllm backend. \
            support [LlamaForCausalLM, QWenLMHeadModel, Qwen2ForCausalLM, Qwen2MoeForCausalLM] only, while {self}")

    state_dict = convert_state_dict_internal(self.model_args, self.config, qwen_version=qwen_version)
    super(type(self), self).load_state_dict(state_dict, strict=strict)


def init(self, load_config):
    # remove 'Model loader extra config' assert.
    self.load_config = load_config

loader.DummyModelLoader.__init__ = init
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



chatlearn/models/vllm/hooks/vllm_0_6_6/loader.py [37:75]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    convert_llama_state_dict_from_megatron_to_vllm,
    convert_llama_state_dict_from_mcore_to_vllm,
    convert_qwen_state_dict_from_megatron_to_vllm,
    load_checkpoint
)

def load_weights(self, model_args):
    torch.distributed.barrier()
    self.model_args = model_args
    load_checkpoint(self, None, None, model_args=model_args)
    torch.distributed.barrier()

def load_state_dict(self, state_dict, strict=True, assign=False):
    qwen_version = None
    if isinstance(self, LlamaForCausalLM):
        use_legacy_models = get_use_legacy_models(self.model_args)
        if use_legacy_models:
            convert_state_dict_internal = convert_llama_state_dict_from_megatron_to_vllm
        else:
            convert_state_dict_internal = convert_llama_state_dict_from_mcore_to_vllm
    elif isinstance(self, QWenLMHeadModel):
        qwen_version = 1.0
        convert_state_dict_internal = convert_qwen_state_dict_from_megatron_to_vllm
    elif isinstance(self, Qwen2ForCausalLM) or (Qwen2MoeForCausalLM is not None and isinstance(self, Qwen2MoeForCausalLM)):
        qwen_version = 2.0
        convert_state_dict_internal = convert_qwen_state_dict_from_megatron_to_vllm
    else:
        raise RuntimeError(f"Unsupported model for vllm backend. \
            support [LlamaForCausalLM, QWenLMHeadModel, Qwen2ForCausalLM, Qwen2MoeForCausalLM] only, while {self}")

    state_dict = convert_state_dict_internal(self.model_args, self.config, qwen_version=qwen_version)
    super(type(self), self).load_state_dict(state_dict, strict=strict)


def init(self, load_config):
    # remove 'Model loader extra config' assert.
    self.load_config = load_config

loader.DummyModelLoader.__init__ = init
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -