vLLM Gaudi Plugin System Architecture

Plugin architecture: platform plugin (vllm.platform_plugins, name=hpu) registers HpuPlatform; general plugin (vllm.general_plugins, name=hpu_custom_ops) registers custom ops. set_torch_compile() configures: PT_HPU_WEIGHT_SHARING=0 (disable weight sharing for HPU); if lazy mode: torch._dynamo.config.disable=True, PT_HPU_ENABLE_LAZY_COLLECTIVES=true; if eager: RUNTIME_SCALE_PATCHING=1. Custom ops registered via register_ops(): hpu_rejection_sampler, hpu_nixl_connector, hpu_fused_moe, hpu_grouped_topk_router, hpu_layernorm, hpu_lora, hpu_mamba_mixer2, hpu_rotary_embedding, hpu_compressed_tensors, hpu_fp8, hpu_gptq, hpu_awq, hpu_mm_encoder_attention.