call check_quantized_moe_compatibility after initialize

[chunyuan-w]

Motivation

Fixes the error when running DeepSeek-V3.1-Terminus-FP8 with TP=6 on CPU.

  File "/sglang/srt/model_executor/model_runner.py", line 306, in __init__
    self.check_quantized_moe_compatibility()
  File "/sglang/srt/model_executor/model_runner.py", line 597, in check_quantized_moe_compatibility
    raise ValueError(
ValueError: moe_intermediate_size 2048 must be divisible by moe_tp_size (6) which is tp_size (6) divided by moe_ep_size (1).

Modifications

Move check_quantized_moe_compatibility() to be after self.initialize(min_per_gpu_memory).

On CPU, we will do padding for moe_intermediate_size if it can't be divided by the tp_size inside self.initialize(min_per_gpu_memory):

sglang/python/sglang/srt/model_executor/model_runner.py

Lines 775 to 778 in de430b6

	if self.device == "cpu":
	self.model_config = adjust_config_with_unaligned_cpu_tp(
	self.model_config, self.load_config, self.tp_size
	)

We need to call check_quantized_moe_compatibility() after this padding otherwise we will run into the above error.
We're not able to move the adjust_config_with_unaligned_cpu_tp() call to an earlier place because it requires self.load_config to be set first and this is done inside the self.initialize(min_per_gpu_memory) here:

sglang/python/sglang/srt/model_executor/model_runner.py

Lines 765 to 774 in de430b6

	self.load_config = LoadConfig(
	load_format=self.server_args.load_format,
	download_dir=self.server_args.download_dir,
	model_loader_extra_config=self.server_args.model_loader_extra_config,
	tp_rank=self.tp_rank,
	remote_instance_weight_loader_seed_instance_ip=self.server_args.remote_instance_weight_loader_seed_instance_ip,
	remote_instance_weight_loader_seed_instance_service_port=self.server_args.remote_instance_weight_loader_seed_instance_service_port,
	remote_instance_weight_loader_send_weights_group_ports=self.server_args.remote_instance_weight_loader_send_weights_group_ports,
	modelopt_config=modelopt_config,
	)

[chunyuan-w]

/rerun-failed-ci

[chunyuan-w]

@zhyncs @Alcanderian could you please help review this PR? The CI failures are unrelated.

[Fridge003]

maybe cc @JustinTong0323

[chunyuan-w]

Hi @JustinTong0323 could you please take a look at this PR?

[chunyuan-w]

Hi @Alcanderian I checked that the CI failures are unrelated to this PR. Could you please help land this PR?