Serve gpt-oss LoRA from BF16 weights with the Triton MoE runner

[kevintli]

Summary

With the gpt-oss preset, the LoRA sampler failed at startup:

sglang/srt/lora/layers.py: NotImplementedError: LoRA MoE not supported for backend MoeRunnerBackend.TRITON_KERNELS

openai/gpt-oss-20b ships MXFP4 experts, and SGLang routes those through triton_kernel. SGLang's LoRA MoE layer only supports triton (which needs get_triton_quant_info, and Mxfp4MoEMethod doesn't have it) or marlin (CompressedTensors/NvFp4 only). Miles has the same constraint and says so: examples/lora/run-gpt-oss-20B-megatron-moe-lora.sh has "need to use bf16 ckpt when enable triton moe backend, eg, lmsys/gpt-oss-20b-bf16", and its gpt-oss LoRA CI uses gpt-oss-20b-bf16.

Changes:

• New recipe field model_weights: the HF repo for the base weights, defaulting to model. The served/client-facing base_model is still model; only the downloaded assets change.

  DeploymentConfig.weights_repo = recipe.model_weights or recipe.model
  asset_path = f"/assets/{weights_repo}"          # trainer hf_checkpoint + SGLang --model-path
  prepare_model_assets: snapshot_download(repo_id=weights_repo, ...)

  model_weights is now part of trainer_identity() and inference_identity(), so switching weights redeploys both.
• gpt-oss preset: model_weights = "lmsys/gpt-oss-20b-bf16" (the MXFP4 release upcast to BF16, same architecture and config apart from quantization_config), plus sglang_cfg.moe_runner_backend = "triton".

Stacked on #6.

Link to Devin session: https://modal.devinenterprise.com/sessions/f53cfabb210146de8f0338fe388d7973
Open in Devin Desktop: https://modal.devinenterprise.com/desktop/session/f53cfabb210146de8f0338fe388d7973?variant=devin
Requested by: @kevintli

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access that start with 'Devin'.

• Disable automatic comment, CI, and merge conflict monitoring