Lumbridge Bench
CI / verify (push) Successful in 24s
CI / deploy (push) Failing after 1m14s

This commit is contained in:
Karti Tripathi
2026-08-04 00:44:07 -07:00
commit 006feee0f7
65 changed files with 13516 additions and 0 deletions
+167
View File
@@ -0,0 +1,167 @@
"""Model registry.
A *target* is not a model. It is a specific way of serving a specific model:
(model × host × quantization × serving config × checkpoint)
This is the central design decision of the bench and it cannot be retrofitted
later without invalidating every result already recorded. On unified-memory
boxes like the DGX Spark, serving flags move throughput more than a model swap
does -- `--enforce-eager` alone can cost 30%+, and `--gpu-memory-utilization`
determines how much KV cache exists, which sets the concurrency ceiling. A
registry keyed on model name alone produces numbers nobody can reproduce.
`checkpoint` is first-class for the same reason: once we fine-tune our own
models, we will compare `karti-7b@step2000` against `karti-7b@step4000` far
more often than we compare Qwen against Llama.
"""
from __future__ import annotations
import os
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
import yaml
REGISTRY_PATH = Path(__file__).parent / "models.yaml"
@dataclass(frozen=True)
class Host:
"""A machine that can serve models."""
id: str
ssh: str | None = None
hardware: str | None = None
memory_gb: int | None = None
memory_bandwidth_gbs: int | None = None
notes: str | None = None
@dataclass(frozen=True)
class Serving:
"""How a target is served. Part of the identity of a result."""
engine: str # vllm | sglang | llama.cpp | anthropic | openai | ...
model_name: str # value sent as "model" in the API request
base_url: str | None = None
service: str | None = None # inspect openai-api service prefix
max_model_len: int | None = None
flags: dict[str, Any] = field(default_factory=dict)
# How to sample this model during quality runs. Part of the identity of a result: the
# same model at temperature 1 and at temperature 0.6 is two different measurements.
#
# Not a global constant, because the right value is a property of the model. Greedy
# decoding looks like the obvious choice for reproducibility and is actively wrong for
# reasoning models -- Qwen3 in thinking mode degenerates into repetition at
# temperature 0, and a 2-sample probe that took two minutes ran past twelve without
# terminating. Reproducibility comes from pinning the seed, not from killing entropy.
sampling: dict[str, Any] = field(default_factory=dict)
@property
def is_local(self) -> bool:
return self.engine in {"vllm", "sglang", "llama.cpp", "ollama"}
@dataclass(frozen=True)
class Target:
"""One benchmarkable configuration."""
id: str
display_name: str
serving: Serving
host: str
family: str | None = None
params_b: float | None = None
active_params_b: float | None = None # MoE: params active per token
quant: str | None = None
checkpoint: str | None = None
tier: str = "candidate" # production | candidate | reference | baseline
aliases: list[str] = field(default_factory=list)
notes: str | None = None
@property
def inspect_model(self) -> str:
"""The model string Inspect uses to address this target."""
s = self.serving
if s.engine in {"anthropic", "openai", "google", "openrouter", "grok"}:
return f"{s.engine}/{s.model_name}"
if s.is_local:
if not s.service:
raise ValueError(f"target {self.id}: local serving requires a 'service'")
# Inspect convention: openai-api/<service>/<model>, with base_url and
# api key read from <SERVICE>_BASE_URL / <SERVICE>_API_KEY.
return f"openai-api/{s.service}/{s.model_name}"
raise ValueError(f"target {self.id}: unsupported engine {s.engine!r}")
@property
def env(self) -> dict[str, str]:
"""Environment variables Inspect needs to reach this target."""
s = self.serving
if not s.is_local or not s.service:
return {}
prefix = s.service.upper().replace("-", "_")
env = {f"{prefix}_API_KEY": os.environ.get(f"{prefix}_API_KEY", "no-key-required")}
if s.base_url:
env[f"{prefix}_BASE_URL"] = s.base_url
return env
def apply_env(self) -> None:
"""Export this target's env vars into the current process."""
os.environ.update(self.env)
@property
def slug(self) -> str:
"""Filesystem-safe identifier used in results filenames."""
return self.id.replace("/", "_").replace("@", "__")
@dataclass(frozen=True)
class Registry:
hosts: dict[str, Host]
targets: dict[str, Target]
def target(self, ident: str) -> Target:
"""Look up a target by id or alias."""
if ident in self.targets:
return self.targets[ident]
for t in self.targets.values():
if ident in t.aliases:
return t
known = ", ".join(sorted(self.targets))
raise KeyError(f"unknown target {ident!r}. known targets: {known}")
def host(self, ident: str) -> Host:
if ident not in self.hosts:
raise KeyError(f"unknown host {ident!r}")
return self.hosts[ident]
def by_tier(self, tier: str) -> list[Target]:
return [t for t in self.targets.values() if t.tier == tier]
def load_registry(path: Path | None = None) -> Registry:
raw = yaml.safe_load((path or REGISTRY_PATH).read_text())
hosts = {
hid: Host(id=hid, **(spec or {}))
for hid, spec in (raw.get("hosts") or {}).items()
}
targets: dict[str, Target] = {}
for spec in raw.get("targets") or []:
spec = dict(spec)
serving = Serving(**spec.pop("serving"))
target = Target(serving=serving, **spec)
if target.host not in hosts:
raise ValueError(f"target {target.id}: unknown host {target.host!r}")
if target.id in targets:
raise ValueError(f"duplicate target id {target.id!r}")
targets[target.id] = target
return Registry(hosts=hosts, targets=targets)
__all__ = ["Host", "Registry", "Serving", "Target", "load_registry"]