feat: add PINNED_MODELS and PRELOAD_API_KEY for preload on serverless

[hansent]

Summary

• PRELOAD_API_KEY: Dedicated API key for model preloading. On user-facing deployments, setting API_KEY globally causes unintended side effects (fallback auth on unauthenticated requests, billing attribution, model-access changes). PRELOAD_API_KEY provides the credential needed for model download during startup without affecting per-request behaviour. Falls back to API_KEY when not set.
• PINNED_MODELS: Comma-separated list of model IDs that are always preloaded at startup (bypassing the LAMBDA/GCP_SERVERLESS gate) and pinned in the LRU cache so they are never evicted under size or memory pressure. This replaces the need for a separate FORCE_PRELOAD flag.
• Improved preload logging: Load timing and resolved model IDs logged for each preloaded model.
• Direct model_manager.add_model() call: The model_add route handler is defined inside if not (LAMBDA or GCP_SERVERLESS), so it doesn't exist on serverless deployments. Preload now calls the underlying method directly.

New env vars

Variable	Values	Default
PRELOAD_API_KEY	Roboflow API key for preloading	Falls back to API_KEY
PINNED_MODELS	Comma-separated model IDs	unset

Example usage

# User-facing deployment on GCP with GCP_SERVERLESS=True:
PINNED_MODELS=sam2/hiera_large,sam3/sam3_final,sam3/sam3_interactive
PRELOAD_API_KEY=rf_your_api_key

Files changed

• inference/core/env.py — new PRELOAD_API_KEY and PINNED_MODELS env vars
• inference/core/interfaces/http/http_api.py — preload gate uses PRELOAD_API_KEY and PINNED_MODELS, calls model_manager.add_model() directly, improved logging
• inference/core/managers/decorators/fixed_size_cache.py — pin_model() method + eviction skip logic

Test plan

• Verify server starts normally without new env vars set (backward compat)
• Set PINNED_MODELS + PRELOAD_API_KEY with GCP_SERVERLESS=True, verify preloading works
• Verify logs show model loading with timing
• Verify /model/registry shows preloaded model after startup
• Load enough models to exceed MAX_ACTIVE_MODELS and verify pinned models are NOT evicted
• Verify PRELOAD_MODELS still works with API_KEY on non-serverless (backward compat)
• Verify per-request auth is unaffected when only PRELOAD_API_KEY is set

🤖 Generated with Claude Code