Skip to content

feat: add vLLM SDK backend for evaluation and self-evolution scripts #46

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Yan0613 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

feat: add vLLM SDK backend for evaluation and self-evolution scripts #46

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions scripts/run_eval.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,17 @@ class EvalSettings(BaseSettings):
default=Path("~/officeqa/officeqa.csv").expanduser(),
description="Path to OfficeQA dataset CSV",
)
sdk: Literal["claude", "opencode", "codex", "goose", "openhands"] = Field(
sdk: Literal[
"claude", "opencode", "codex", "goose", "openhands"
] = Field(
default="claude",
description="SDK to use: 'claude', 'opencode', 'codex', or 'goose'",
description="SDK to use: 'claude', 'opencode', 'codex', 'goose', or 'openhands'",
)


async def main(settings: EvalSettings):
set_sdk(settings.sdk)
print(f"[SDK] {settings.sdk} backend, model: {settings.model}")

# Load dataset
data = pd.read_csv(settings.dataset_path)
Expand Down
50 changes: 47 additions & 3 deletions scripts/run_eval_dabstep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@

from src.harness import Agent
from src.agent_profiles import dabstep_agent_options, make_dabstep_agent_options
from src.agent_profiles import set_sdk, set_vllm_config
from src.evaluation.eval_full import evaluate_full, load_results
from src.evaluation.dabstep_scorer import question_scorer
from src.schemas import AgentResponse
Expand Down Expand Up @@ -80,8 +81,46 @@ async def main():
default="claude-opus-4-5-20251101",
help="Model for agent (default: claude-opus-4-5-20251101)",
)
parser.add_argument(
"--sdk",
type=str,
default="claude",
choices=["claude", "opencode", "vllm"],
help="SDK backend to use: claude (default), opencode, or vllm",
)
parser.add_argument(
"--vllm-base-url",
type=str,
default="http://localhost:8000/v1",
help="vLLM server base URL (used when --sdk vllm, default: http://localhost:8000/v1)",
)
parser.add_argument(
"--vllm-max-tokens",
type=int,
default=8192,
help="Max tokens for vLLM generation (default: 8192)",
)
parser.add_argument(
"--vllm-context-length",
type=int,
default=131072,
help="vLLM model context window size (default: 131072 for 128K models like Qwen2.5-72B)",
)
args = parser.parse_args()

# Configure SDK backend
set_sdk(args.sdk)
if args.sdk == "vllm":
set_vllm_config(
base_url=args.vllm_base_url,
model_name=args.model,
max_tokens=args.vllm_max_tokens,
context_length=args.vllm_context_length,
)
print(f"[SDK] vLLM backend: {args.vllm_base_url}, model: {args.model}, context: {args.vllm_context_length}, max_tokens: {args.vllm_max_tokens}")
else:
print(f"[SDK] {args.sdk} backend, model: {args.model}")

# Load dataset
data = pd.read_csv(args.dataset)

Expand All @@ -97,9 +136,14 @@ async def main():

# Auto-discover context files from data-dir
data_dir = Path(args.data_dir).resolve()
context_file_names = sorted(f.name for f in data_dir.iterdir() if f.is_file())
context_files_text = "\n".join(f"- {data_dir / name}" for name in context_file_names)
print(f"Context files ({len(context_file_names)}): {', '.join(context_file_names)}")
if data_dir.exists():
context_file_names = sorted(f.name for f in data_dir.iterdir() if f.is_file())
context_files_text = "\n".join(f"- {data_dir / name}" for name in context_file_names)
print(f"Context files ({len(context_file_names)}): {', '.join(context_file_names)}")
else:
context_file_names = []
context_files_text = "(no context files available)"
print(f"[WARNING] data-dir not found: {data_dir}. Context files will be unavailable.")

# Prepare items: (task_id, formatted_prompt, answer)
items = [
Expand Down
34 changes: 31 additions & 3 deletions scripts/run_eval_livecodebench.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,7 @@
import pandas as pd

from src.harness import Agent, set_sdk
from src.agent_profiles import make_livecodebench_agent_options
from src.agent_profiles import make_livecodebench_agent_options, set_vllm_config
from src.evaluation.eval_full import evaluate_full, load_results
from src.evaluation.livecodebench import (
score_livecodebench,
Expand Down Expand Up @@ -77,14 +77,42 @@ async def main():
parser.add_argument(
"--sdk",
type=str,
choices=["claude", "opencode"],
choices=["claude", "opencode", "vllm"],
default="claude",
help="SDK to use: 'claude' or 'opencode' (default: claude)",
help="SDK to use: 'claude', 'opencode', or 'vllm' (default: claude)",
)
parser.add_argument(
"--vllm-base-url",
type=str,
default="http://localhost:8000/v1",
help="vLLM server base URL (used when --sdk vllm, default: http://localhost:8000/v1)",
)
parser.add_argument(
"--vllm-max-tokens",
type=int,
default=8192,
help="Max tokens for vLLM generation (default: 8192)",
)
parser.add_argument(
"--vllm-context-length",
type=int,
default=131072,
help="vLLM model context window size (default: 131072 for 128K models like Qwen2.5-72B)",
)
args = parser.parse_args()

# Set SDK
set_sdk(args.sdk)
if args.sdk == "vllm":
set_vllm_config(
base_url=args.vllm_base_url,
model_name=args.model,
max_tokens=args.vllm_max_tokens,
context_length=args.vllm_context_length,
)
print(f"[SDK] vLLM backend: {args.vllm_base_url}, model: {args.model}, context: {args.vllm_context_length}, max_tokens: {args.vllm_max_tokens}")
else:
print(f"[SDK] {args.sdk} backend, model: {args.model}")

# Ensure dataset is downloaded
if args.dataset is None:
Expand Down
39 changes: 39 additions & 0 deletions scripts/run_eval_sealqa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@

from src.harness import Agent
from src.agent_profiles import sealqa_agent_options, make_sealqa_agent_options
from src.agent_profiles import set_sdk, set_vllm_config
from src.evaluation.eval_full import evaluate_full, load_results
from src.evaluation.sealqa_scorer import score_sealqa
from src.schemas import AgentResponse
Expand Down Expand Up @@ -62,8 +63,46 @@ async def main():
default="claude-opus-4-5-20251101",
help="Model for agent (default: claude-opus-4-5-20251101)",
)
parser.add_argument(
"--sdk",
type=str,
default="claude",
choices=["claude", "opencode", "vllm"],
help="SDK backend to use: claude (default), opencode, or vllm",
)
parser.add_argument(
"--vllm-base-url",
type=str,
default="http://localhost:8000/v1",
help="vLLM server base URL (used when --sdk vllm, default: http://localhost:8000/v1)",
)
parser.add_argument(
"--vllm-max-tokens",
type=int,
default=8192,
help="Max tokens for vLLM generation (default: 8192)",
)
parser.add_argument(
"--vllm-context-length",
type=int,
default=131072,
help="vLLM model context window size (default: 131072 for 128K models like Qwen2.5-72B)",
)
args = parser.parse_args()

# Configure SDK backend
set_sdk(args.sdk)
if args.sdk == "vllm":
set_vllm_config(
base_url=args.vllm_base_url,
model_name=args.model,
max_tokens=args.vllm_max_tokens,
context_length=args.vllm_context_length,
)
print(f"[SDK] vLLM backend: {args.vllm_base_url}, model: {args.model}, context: {args.vllm_context_length}, max_tokens: {args.vllm_max_tokens}")
else:
print(f"[SDK] {args.sdk} backend, model: {args.model}")

# Load dataset
data = pd.read_csv(args.dataset)

Expand Down
26 changes: 24 additions & 2 deletions scripts/run_loop.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,6 +21,7 @@
skill_generator_options,
prompt_generator_options,
set_sdk,
set_vllm_config,
)
from src.agent_profiles.skill_generator import get_project_root
from src.registry import ProgramManager
Expand Down Expand Up @@ -84,9 +85,23 @@ class LoopSettings(BaseSettings):
model: Optional[str] = Field(
default=None, description="Model for base agent (opus, sonnet, haiku)"
)
sdk: Literal["claude", "opencode", "codex", "goose", "openhands"] = Field(
sdk: Literal[
"claude", "opencode", "codex", "goose", "openhands", "vllm"
] = Field(
default="claude",
description="SDK to use: 'claude', 'opencode', 'codex', or 'goose'",
description="SDK to use: 'claude', 'opencode', 'codex', 'goose', 'openhands', or 'vllm'",
)
vllm_base_url: str = Field(
default="http://localhost:8000/v1",
description="vLLM server base URL (only used when sdk='vllm')",
)
vllm_max_tokens: int = Field(
default=8192,
description="Max tokens for vLLM generation (only used when sdk='vllm')",
)
vllm_context_length: int = Field(
default=131072,
description="vLLM model context window size (default: 131072 for 128K models like Qwen2.5-72B)",
)


Expand Down Expand Up @@ -138,6 +153,13 @@ def stratified_split(
async def main(settings: LoopSettings):
# Set SDK based on CLI argument
set_sdk(settings.sdk)
if settings.sdk == "vllm":
set_vllm_config(
base_url=settings.vllm_base_url,
model_name=settings.model,
max_tokens=settings.vllm_max_tokens,
context_length=settings.vllm_context_length,
)

data = pd.read_csv(settings.dataset)

Expand Down