[model] Add support for GLM4.7 Flash

[zhuzilin]

Please change the "architectures" to "DeepseekV32ForCausalLM" and "model_type" to "deepseek_v3" before the sglang in the slime image supports Glm4MoeLiteForCausalLM, and add "moe_layer_freq": 1.

[sxthunder]

I modified the config.json of glm-4.7 by changing "architectures" to "DeepseekV3ForCausalLM" and "model_type" to "deepseek_v3", but it still failed with an AttributeError: 'DeepseekV3Config' object has no attribute 'rope_theta'

[ifififa]

@zhuzilin hi，can we convert glm-4.7-flash from huggingface to megatron format with --mtp-num-layers 1?
when I set --mtp-num-layers 1 in conversion script, I met this error:

[rank7]: Traceback (most recent call last):
[rank7]: File "/root/slime/tools/convert_hf_to_torch_dist.py", line 145, in
[rank7]: main()
[rank7]: File "/root/slime/tools/convert_hf_to_torch_dist.py", line 119, in main
[rank7]: bridge.load_weights(model, hf_model_path, memory_efficient=True)
[rank7]: File "/usr/local/lib/python3.12/dist-packages/mbridge/core/bridge.py", line 172, in load_weights
[rank7]: k: self._weight_name_mapping_mcore_to_hf(v)
[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank7]: File "/usr/local/lib/python3.12/dist-packages/mbridge/models/deepseek_v3.py", line 300, in _weight_name_mapping_mcore_to_hf
[rank7]: return self._convert_mtp_param(mcore_weights_name)
[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank7]: File "/usr/local/lib/python3.12/dist-packages/mbridge/models/deepseek_v3.py", line 361, in _convert_mtp_param
[rank7]: assert self.config.num_layers == 61, "only support 61 layers for now"
[rank7]: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank7]: AssertionError: only support 61 layers for now
[rank7]:[W209 12:07:19.756675305 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())