Spaces:
Running on CPU Upgrade
Running on CPU Upgrade
Merge pull request #3 from huggingface/job-tool-improvements
Browse files- agent/core/agent_loop.py +121 -3
- agent/core/session.py +1 -0
- agent/core/tools.py +9 -4
- agent/main.py +67 -1
- agent/prompts/system_prompt.yaml +1 -0
- agent/tools/jobs_tool.py +154 -27
agent/core/agent_loop.py
CHANGED
|
@@ -15,6 +15,16 @@ from agent.core.tools import ToolRouter
|
|
| 15 |
ToolCall = ChatCompletionMessageToolCall
|
| 16 |
|
| 17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
class Handlers:
|
| 19 |
"""Handler functions for each operation type"""
|
| 20 |
|
|
@@ -33,9 +43,10 @@ class Handlers:
|
|
| 33 |
|
| 34 |
Laminar.set_trace_session_id(session_id=session.session_id)
|
| 35 |
|
| 36 |
-
# Add user message to history
|
| 37 |
-
|
| 38 |
-
|
|
|
|
| 39 |
|
| 40 |
# Send event that we're processing
|
| 41 |
await session.send_event(
|
|
@@ -97,6 +108,28 @@ class Handlers:
|
|
| 97 |
tool_name = tc.function.name
|
| 98 |
tool_args = json.loads(tc.function.arguments)
|
| 99 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
await session.send_event(
|
| 101 |
Event(
|
| 102 |
event_type="tool_call",
|
|
@@ -191,6 +224,85 @@ class Handlers:
|
|
| 191 |
|
| 192 |
await session.send_event(Event(event_type="undo_complete"))
|
| 193 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 194 |
@staticmethod
|
| 195 |
async def shutdown(session: Session) -> bool:
|
| 196 |
"""Handle shutdown (like shutdown in codex.rs:1329)"""
|
|
@@ -226,6 +338,12 @@ async def process_submission(session: Session, submission) -> bool:
|
|
| 226 |
await Handlers.undo(session)
|
| 227 |
return True
|
| 228 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
if op.op_type == OpType.SHUTDOWN:
|
| 230 |
return not await Handlers.shutdown(session)
|
| 231 |
|
|
|
|
| 15 |
ToolCall = ChatCompletionMessageToolCall
|
| 16 |
|
| 17 |
|
| 18 |
+
def _needs_approval(tool_name: str, tool_args: dict) -> bool:
|
| 19 |
+
"""Check if a tool call requires user approval before execution"""
|
| 20 |
+
if tool_name != "hf_jobs":
|
| 21 |
+
return False
|
| 22 |
+
|
| 23 |
+
# Check if it's a run or uv operation
|
| 24 |
+
operation = tool_args.get("operation", "")
|
| 25 |
+
return operation in ["run", "uv"]
|
| 26 |
+
|
| 27 |
+
|
| 28 |
class Handlers:
|
| 29 |
"""Handler functions for each operation type"""
|
| 30 |
|
|
|
|
| 43 |
|
| 44 |
Laminar.set_trace_session_id(session_id=session.session_id)
|
| 45 |
|
| 46 |
+
# Add user message to history only if there's actual content
|
| 47 |
+
if text:
|
| 48 |
+
user_msg = Message(role="user", content=text)
|
| 49 |
+
session.context_manager.add_message(user_msg)
|
| 50 |
|
| 51 |
# Send event that we're processing
|
| 52 |
await session.send_event(
|
|
|
|
| 108 |
tool_name = tc.function.name
|
| 109 |
tool_args = json.loads(tc.function.arguments)
|
| 110 |
|
| 111 |
+
# Check if this tool requires user approval
|
| 112 |
+
if _needs_approval(tool_name, tool_args):
|
| 113 |
+
await session.send_event(
|
| 114 |
+
Event(
|
| 115 |
+
event_type="approval_required",
|
| 116 |
+
data={
|
| 117 |
+
"tool": tool_name,
|
| 118 |
+
"arguments": tool_args,
|
| 119 |
+
"tool_call_id": tc.id,
|
| 120 |
+
},
|
| 121 |
+
)
|
| 122 |
+
)
|
| 123 |
+
|
| 124 |
+
# Store pending approval and return early
|
| 125 |
+
session.pending_approval = {
|
| 126 |
+
"tool_call": tc,
|
| 127 |
+
"arguments": tool_args,
|
| 128 |
+
}
|
| 129 |
+
|
| 130 |
+
# Return early - wait for EXEC_APPROVAL operation
|
| 131 |
+
return None
|
| 132 |
+
|
| 133 |
await session.send_event(
|
| 134 |
Event(
|
| 135 |
event_type="tool_call",
|
|
|
|
| 224 |
|
| 225 |
await session.send_event(Event(event_type="undo_complete"))
|
| 226 |
|
| 227 |
+
@staticmethod
|
| 228 |
+
async def exec_approval(
|
| 229 |
+
session: Session, approved: bool, feedback: str | None = None
|
| 230 |
+
) -> None:
|
| 231 |
+
"""Handle job execution approval"""
|
| 232 |
+
if not session.pending_approval:
|
| 233 |
+
await session.send_event(
|
| 234 |
+
Event(
|
| 235 |
+
event_type="error",
|
| 236 |
+
data={"error": "No pending approval to process"},
|
| 237 |
+
)
|
| 238 |
+
)
|
| 239 |
+
return
|
| 240 |
+
|
| 241 |
+
tc = session.pending_approval["tool_call"]
|
| 242 |
+
tool_args = session.pending_approval["arguments"]
|
| 243 |
+
tool_name = tc.function.name
|
| 244 |
+
|
| 245 |
+
if approved:
|
| 246 |
+
# Execute the pending tool
|
| 247 |
+
await session.send_event(
|
| 248 |
+
Event(
|
| 249 |
+
event_type="tool_call",
|
| 250 |
+
data={"tool": tool_name, "arguments": tool_args},
|
| 251 |
+
)
|
| 252 |
+
)
|
| 253 |
+
|
| 254 |
+
output, success = await session.tool_router.call_tool(tool_name, tool_args)
|
| 255 |
+
|
| 256 |
+
# Add tool result to context
|
| 257 |
+
tool_msg = Message(
|
| 258 |
+
role="tool",
|
| 259 |
+
content=output,
|
| 260 |
+
tool_call_id=tc.id,
|
| 261 |
+
name=tool_name,
|
| 262 |
+
)
|
| 263 |
+
session.context_manager.add_message(tool_msg)
|
| 264 |
+
|
| 265 |
+
await session.send_event(
|
| 266 |
+
Event(
|
| 267 |
+
event_type="tool_output",
|
| 268 |
+
data={
|
| 269 |
+
"tool": tool_name,
|
| 270 |
+
"output": output,
|
| 271 |
+
"success": success,
|
| 272 |
+
},
|
| 273 |
+
)
|
| 274 |
+
)
|
| 275 |
+
else:
|
| 276 |
+
# User rejected - add cancellation message to context
|
| 277 |
+
cancellation_msg = "Job execution cancelled by user"
|
| 278 |
+
if feedback:
|
| 279 |
+
cancellation_msg += f". User feedback: {feedback}"
|
| 280 |
+
|
| 281 |
+
tool_msg = Message(
|
| 282 |
+
role="tool",
|
| 283 |
+
content=cancellation_msg,
|
| 284 |
+
tool_call_id=tc.id,
|
| 285 |
+
name=tool_name,
|
| 286 |
+
)
|
| 287 |
+
session.context_manager.add_message(tool_msg)
|
| 288 |
+
|
| 289 |
+
await session.send_event(
|
| 290 |
+
Event(
|
| 291 |
+
event_type="tool_output",
|
| 292 |
+
data={
|
| 293 |
+
"tool": tool_name,
|
| 294 |
+
"output": cancellation_msg,
|
| 295 |
+
"success": False,
|
| 296 |
+
},
|
| 297 |
+
)
|
| 298 |
+
)
|
| 299 |
+
|
| 300 |
+
# Clear pending approval
|
| 301 |
+
session.pending_approval = None
|
| 302 |
+
|
| 303 |
+
# Continue agent loop with empty input to process the tool result
|
| 304 |
+
await Handlers.run_agent(session, "")
|
| 305 |
+
|
| 306 |
@staticmethod
|
| 307 |
async def shutdown(session: Session) -> bool:
|
| 308 |
"""Handle shutdown (like shutdown in codex.rs:1329)"""
|
|
|
|
| 338 |
await Handlers.undo(session)
|
| 339 |
return True
|
| 340 |
|
| 341 |
+
if op.op_type == OpType.EXEC_APPROVAL:
|
| 342 |
+
approved = op.data.get("approved", False) if op.data else False
|
| 343 |
+
feedback = op.data.get("feedback") if op.data else None
|
| 344 |
+
await Handlers.exec_approval(session, approved, feedback)
|
| 345 |
+
return True
|
| 346 |
+
|
| 347 |
if op.op_type == OpType.SHUTDOWN:
|
| 348 |
return not await Handlers.shutdown(session)
|
| 349 |
|
agent/core/session.py
CHANGED
|
@@ -53,6 +53,7 @@ class Session:
|
|
| 53 |
)
|
| 54 |
self.is_running = True
|
| 55 |
self.current_task: asyncio.Task | None = None
|
|
|
|
| 56 |
|
| 57 |
async def send_event(self, event: Event) -> None:
|
| 58 |
"""Send event back to client"""
|
|
|
|
| 53 |
)
|
| 54 |
self.is_running = True
|
| 55 |
self.current_task: asyncio.Task | None = None
|
| 56 |
+
self.pending_approval: Optional[dict[str, Any]] = None
|
| 57 |
|
| 58 |
async def send_event(self, event: Event) -> None:
|
| 59 |
"""Send event back to client"""
|
agent/core/tools.py
CHANGED
|
@@ -8,6 +8,7 @@ from dataclasses import dataclass
|
|
| 8 |
from typing import Any, Awaitable, Callable, Optional
|
| 9 |
|
| 10 |
from fastmcp import Client
|
|
|
|
| 11 |
from lmnr import observe
|
| 12 |
from mcp.types import EmbeddedResource, ImageContent, TextContent
|
| 13 |
|
|
@@ -166,10 +167,14 @@ class ToolRouter:
|
|
| 166 |
|
| 167 |
# Otherwise, use MCP client
|
| 168 |
if self._mcp_initialized:
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
|
| 174 |
return "MCP client not initialized", False
|
| 175 |
|
|
|
|
| 8 |
from typing import Any, Awaitable, Callable, Optional
|
| 9 |
|
| 10 |
from fastmcp import Client
|
| 11 |
+
from fastmcp.exceptions import ToolError
|
| 12 |
from lmnr import observe
|
| 13 |
from mcp.types import EmbeddedResource, ImageContent, TextContent
|
| 14 |
|
|
|
|
| 167 |
|
| 168 |
# Otherwise, use MCP client
|
| 169 |
if self._mcp_initialized:
|
| 170 |
+
try:
|
| 171 |
+
result = await self.mcp_client.call_tool(tool_name, arguments)
|
| 172 |
+
output = convert_mcp_content_to_string(result.content)
|
| 173 |
+
return output, not result.is_error
|
| 174 |
+
except ToolError as e:
|
| 175 |
+
# Catch MCP tool errors and return them to the agent
|
| 176 |
+
error_msg = f"Tool error: {str(e)}"
|
| 177 |
+
return error_msg, False
|
| 178 |
|
| 179 |
return "MCP client not initialized", False
|
| 180 |
|
agent/main.py
CHANGED
|
@@ -47,10 +47,13 @@ class Submission:
|
|
| 47 |
|
| 48 |
async def event_listener(
|
| 49 |
event_queue: asyncio.Queue,
|
|
|
|
| 50 |
turn_complete_event: asyncio.Event,
|
| 51 |
ready_event: asyncio.Event,
|
| 52 |
) -> None:
|
| 53 |
"""Background task that listens for events and displays them"""
|
|
|
|
|
|
|
| 54 |
while True:
|
| 55 |
try:
|
| 56 |
event = await event_queue.get()
|
|
@@ -96,6 +99,69 @@ async def event_listener(
|
|
| 96 |
old_tokens = event.data.get("old_tokens", 0) if event.data else 0
|
| 97 |
new_tokens = event.data.get("new_tokens", 0) if event.data else 0
|
| 98 |
print(f"📦 Compacted context: {old_tokens} → {new_tokens} tokens")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
# Silently ignore other events
|
| 100 |
|
| 101 |
except asyncio.CancelledError:
|
|
@@ -145,7 +211,7 @@ async def main():
|
|
| 145 |
|
| 146 |
# Start event listener in background
|
| 147 |
listener_task = asyncio.create_task(
|
| 148 |
-
event_listener(event_queue, turn_complete_event, ready_event)
|
| 149 |
)
|
| 150 |
|
| 151 |
# Wait for agent to initialize
|
|
|
|
| 47 |
|
| 48 |
async def event_listener(
|
| 49 |
event_queue: asyncio.Queue,
|
| 50 |
+
submission_queue: asyncio.Queue,
|
| 51 |
turn_complete_event: asyncio.Event,
|
| 52 |
ready_event: asyncio.Event,
|
| 53 |
) -> None:
|
| 54 |
"""Background task that listens for events and displays them"""
|
| 55 |
+
submission_id = [1000] # Use list to make it mutable in closure
|
| 56 |
+
|
| 57 |
while True:
|
| 58 |
try:
|
| 59 |
event = await event_queue.get()
|
|
|
|
| 99 |
old_tokens = event.data.get("old_tokens", 0) if event.data else 0
|
| 100 |
new_tokens = event.data.get("new_tokens", 0) if event.data else 0
|
| 101 |
print(f"📦 Compacted context: {old_tokens} → {new_tokens} tokens")
|
| 102 |
+
elif event.event_type == "approval_required":
|
| 103 |
+
# Display job details and prompt for approval
|
| 104 |
+
tool_name = event.data.get("tool", "") if event.data else ""
|
| 105 |
+
arguments = event.data.get("arguments", {}) if event.data else {}
|
| 106 |
+
|
| 107 |
+
print("\n" + "=" * 60)
|
| 108 |
+
print("⚠️ JOB EXECUTION APPROVAL REQUIRED")
|
| 109 |
+
print("=" * 60)
|
| 110 |
+
|
| 111 |
+
operation = arguments.get("operation", "")
|
| 112 |
+
args = arguments.get("args", {})
|
| 113 |
+
|
| 114 |
+
print(f"Operation: {operation}")
|
| 115 |
+
|
| 116 |
+
if operation == "uv":
|
| 117 |
+
script = args.get("script", "")
|
| 118 |
+
dependencies = args.get("dependencies", [])
|
| 119 |
+
print(f"Script to run:\n{script}")
|
| 120 |
+
if dependencies:
|
| 121 |
+
print(f"Dependencies: {', '.join(dependencies)}")
|
| 122 |
+
elif operation == "run":
|
| 123 |
+
image = args.get("image", "")
|
| 124 |
+
command = args.get("command", "")
|
| 125 |
+
print(f"Docker image: {image}")
|
| 126 |
+
print(f"Command: {command}")
|
| 127 |
+
|
| 128 |
+
# Common parameters
|
| 129 |
+
flavor = args.get("flavor", "cpu-basic")
|
| 130 |
+
detached = args.get("detached", False)
|
| 131 |
+
print(f"Hardware: {flavor}")
|
| 132 |
+
print(f"Detached mode: {detached}")
|
| 133 |
+
|
| 134 |
+
secrets = args.get("secrets", [])
|
| 135 |
+
if secrets:
|
| 136 |
+
print(f"Secrets: {', '.join(secrets)}")
|
| 137 |
+
|
| 138 |
+
print("=" * 60)
|
| 139 |
+
|
| 140 |
+
# Get user decision
|
| 141 |
+
loop = asyncio.get_event_loop()
|
| 142 |
+
response = await loop.run_in_executor(
|
| 143 |
+
None,
|
| 144 |
+
input,
|
| 145 |
+
"Approve? (y=yes, n=no, or provide feedback to reject): ",
|
| 146 |
+
)
|
| 147 |
+
|
| 148 |
+
response = response.strip()
|
| 149 |
+
approved = response.lower() in ["y", "yes"]
|
| 150 |
+
feedback = (
|
| 151 |
+
None if approved or response.lower() in ["n", "no"] else response
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
# Submit approval
|
| 155 |
+
submission_id[0] += 1
|
| 156 |
+
approval_submission = Submission(
|
| 157 |
+
id=f"approval_{submission_id[0]}",
|
| 158 |
+
operation=Operation(
|
| 159 |
+
op_type=OpType.EXEC_APPROVAL,
|
| 160 |
+
data={"approved": approved, "feedback": feedback},
|
| 161 |
+
),
|
| 162 |
+
)
|
| 163 |
+
await submission_queue.put(approval_submission)
|
| 164 |
+
print("=" * 60 + "\n")
|
| 165 |
# Silently ignore other events
|
| 166 |
|
| 167 |
except asyncio.CancelledError:
|
|
|
|
| 211 |
|
| 212 |
# Start event listener in background
|
| 213 |
listener_task = asyncio.create_task(
|
| 214 |
+
event_listener(event_queue, submission_queue, turn_complete_event, ready_event)
|
| 215 |
)
|
| 216 |
|
| 217 |
# Wait for agent to initialize
|
agent/prompts/system_prompt.yaml
CHANGED
|
@@ -85,6 +85,7 @@ system_prompt: |
|
|
| 85 |
- Always search Hugging Face Hub for existing resources before suggesting custom implementations
|
| 86 |
- When referencing models, datasets, or papers, include direct links from search results
|
| 87 |
- Never assume a library is available - check documentation first
|
|
|
|
| 88 |
- Follow ML best practices: proper train/val/test splits, reproducibility, evaluation metrics
|
| 89 |
- For training tasks, consider compute requirements and suggest appropriate hardware
|
| 90 |
- Never expose or log API keys, tokens, or secrets
|
|
|
|
| 85 |
- Always search Hugging Face Hub for existing resources before suggesting custom implementations
|
| 86 |
- When referencing models, datasets, or papers, include direct links from search results
|
| 87 |
- Never assume a library is available - check documentation first
|
| 88 |
+
- Before processing any dataset: inspect its actual structure first using the mcp__hf-mcp-server__hub_repo_details tool. Never assume column names: verify them beforehand.
|
| 89 |
- Follow ML best practices: proper train/val/test splits, reproducibility, evaluation metrics
|
| 90 |
- For training tasks, consider compute requirements and suggest appropriate hardware
|
| 91 |
- Never expose or log API keys, tokens, or secrets
|
agent/tools/jobs_tool.py
CHANGED
|
@@ -6,6 +6,7 @@ Refactored to use official huggingface-hub library instead of custom HTTP client
|
|
| 6 |
|
| 7 |
import asyncio
|
| 8 |
import base64
|
|
|
|
| 9 |
from typing import Any, Dict, Literal, Optional
|
| 10 |
|
| 11 |
from huggingface_hub import HfApi
|
|
@@ -60,10 +61,32 @@ OperationType = Literal[
|
|
| 60 |
]
|
| 61 |
|
| 62 |
# Constants
|
| 63 |
-
DEFAULT_LOG_WAIT_SECONDS = 10
|
| 64 |
UV_DEFAULT_IMAGE = "ghcr.io/astral-sh/uv:python3.12-bookworm"
|
| 65 |
|
| 66 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
def _build_uv_command(
|
| 68 |
script: str,
|
| 69 |
with_deps: list[str] | None = None,
|
|
@@ -103,6 +126,20 @@ def _wrap_inline_script(
|
|
| 103 |
return f'echo "{encoded}" | base64 -d | {uv_command_str}'
|
| 104 |
|
| 105 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
def _resolve_uv_command(
|
| 107 |
script: str,
|
| 108 |
with_deps: list[str] | None = None,
|
|
@@ -316,7 +353,8 @@ Call this tool with:
|
|
| 316 |
{{
|
| 317 |
"operation": "uv",
|
| 318 |
"args": {{
|
| 319 |
-
"script": "import random\\nprint(42 + random.randint(1, 5))"
|
|
|
|
| 320 |
}}
|
| 321 |
}}
|
| 322 |
```
|
|
@@ -335,7 +373,6 @@ Call this tool with:
|
|
| 335 |
**String format (simple cases only):**
|
| 336 |
- Still accepted for backwards compatibility, parsed with POSIX shell semantics
|
| 337 |
- Rejects shell operators and can mis-handle characters such as `&`; switch to arrays when things turn complex
|
| 338 |
-
- `$HF_TOKEN` stays literal—forward it via `secrets: {{ "HF_TOKEN": "$HF_TOKEN" }}`
|
| 339 |
|
| 340 |
### Show command-specific help
|
| 341 |
Call this tool with:
|
|
@@ -345,9 +382,10 @@ Call this tool with:
|
|
| 345 |
|
| 346 |
## Tips
|
| 347 |
|
| 348 |
-
- Jobs default to non-detached mode (
|
| 349 |
- Prefer array commands to avoid shell parsing surprises
|
| 350 |
-
- To access private Hub assets,
|
|
|
|
| 351 |
"""
|
| 352 |
return {"formatted": usage_text, "totalResults": 1, "resultsShared": 1}
|
| 353 |
|
|
@@ -356,6 +394,33 @@ Call this tool with:
|
|
| 356 |
help_text = f"Help for operation: {operation}\n\nCall with appropriate arguments. Use the main help for examples."
|
| 357 |
return {"formatted": help_text, "totalResults": 1, "resultsShared": 1}
|
| 358 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
async def _run_job(self, args: Dict[str, Any]) -> ToolResult:
|
| 360 |
"""Run a job using HfApi.run_job()"""
|
| 361 |
try:
|
|
@@ -363,8 +428,8 @@ Call this tool with:
|
|
| 363 |
self.api.run_job,
|
| 364 |
image=args.get("image", "python:3.12"),
|
| 365 |
command=args.get("command"),
|
| 366 |
-
env=args.get("env"),
|
| 367 |
-
secrets=args.get("secrets"),
|
| 368 |
flavor=args.get("flavor", "cpu-basic"),
|
| 369 |
timeout=args.get("timeout", "30m"),
|
| 370 |
namespace=args.get("namespace") or self.namespace,
|
|
@@ -382,14 +447,28 @@ To check logs, call this tool with `{{"operation": "logs", "args": {{"job_id": "
|
|
| 382 |
To inspect, call this tool with `{{"operation": "inspect", "args": {{"job_id": "{job.id}"}}}}`"""
|
| 383 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 384 |
|
| 385 |
-
# Not detached -
|
| 386 |
-
|
|
|
|
| 387 |
|
| 388 |
-
|
| 389 |
-
|
|
|
|
|
|
|
| 390 |
|
| 391 |
-
|
| 392 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 393 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 394 |
|
| 395 |
except Exception as e:
|
|
@@ -402,10 +481,18 @@ Note: Logs are being collected. Check the job page for real-time logs.
|
|
| 402 |
if not script:
|
| 403 |
raise ValueError("script is required")
|
| 404 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 405 |
# Resolve the command based on script type (URL, inline, or file)
|
| 406 |
command = _resolve_uv_command(
|
| 407 |
script=script,
|
| 408 |
-
with_deps=
|
| 409 |
python=args.get("python"),
|
| 410 |
script_args=args.get("script_args"),
|
| 411 |
)
|
|
@@ -415,20 +502,46 @@ Note: Logs are being collected. Check the job page for real-time logs.
|
|
| 415 |
self.api.run_job,
|
| 416 |
image=UV_DEFAULT_IMAGE,
|
| 417 |
command=command,
|
| 418 |
-
env=args.get("env"),
|
| 419 |
-
secrets=args.get("secrets"),
|
| 420 |
-
flavor=args.get("flavor"
|
| 421 |
timeout=args.get("timeout", "30m"),
|
| 422 |
namespace=args.get("namespace") or self.namespace,
|
| 423 |
)
|
| 424 |
|
| 425 |
-
|
|
|
|
|
|
|
| 426 |
|
|
|
|
| 427 |
**Status:** {job.status.stage}
|
| 428 |
**View at:** {job.url}
|
| 429 |
|
| 430 |
-
To check logs, call this tool with `{{"operation": "logs", "args": {{"job_id": "{job.id}"}}}}`
|
| 431 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 432 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 433 |
|
| 434 |
except Exception as e:
|
|
@@ -578,8 +691,8 @@ To verify, call this tool with `{{"operation": "inspect", "args": {{"job_id": "{
|
|
| 578 |
image=args.get("image", "python:3.12"),
|
| 579 |
command=args.get("command"),
|
| 580 |
schedule=args.get("schedule"),
|
| 581 |
-
env=args.get("env"),
|
| 582 |
-
secrets=args.get("secrets"),
|
| 583 |
flavor=args.get("flavor", "cpu-basic"),
|
| 584 |
timeout=args.get("timeout", "30m"),
|
| 585 |
namespace=args.get("namespace") or self.namespace,
|
|
@@ -613,10 +726,18 @@ To list all, call this tool with `{{"operation": "scheduled ps"}}`"""
|
|
| 613 |
if not schedule:
|
| 614 |
raise ValueError("schedule is required")
|
| 615 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 616 |
# Resolve the command based on script type
|
| 617 |
command = _resolve_uv_command(
|
| 618 |
script=script,
|
| 619 |
-
with_deps=
|
| 620 |
python=args.get("python"),
|
| 621 |
script_args=args.get("script_args"),
|
| 622 |
)
|
|
@@ -627,9 +748,9 @@ To list all, call this tool with `{{"operation": "scheduled ps"}}`"""
|
|
| 627 |
image=UV_DEFAULT_IMAGE,
|
| 628 |
command=command,
|
| 629 |
schedule=schedule,
|
| 630 |
-
env=args.get("env"),
|
| 631 |
-
secrets=args.get("secrets"),
|
| 632 |
-
flavor=args.get("flavor"
|
| 633 |
timeout=args.get("timeout", "30m"),
|
| 634 |
namespace=args.get("namespace") or self.namespace,
|
| 635 |
)
|
|
@@ -788,6 +909,7 @@ HF_JOBS_TOOL_SPEC = {
|
|
| 788 |
"description": (
|
| 789 |
"Manage Hugging Face CPU/GPU compute jobs. Run commands in Docker containers, "
|
| 790 |
"execute Python scripts with UV. List, schedule and monitor jobs/logs. "
|
|
|
|
| 791 |
"Call this tool with no operation for full usage instructions and examples."
|
| 792 |
),
|
| 793 |
"parameters": {
|
|
@@ -818,7 +940,12 @@ HF_JOBS_TOOL_SPEC = {
|
|
| 818 |
},
|
| 819 |
"args": {
|
| 820 |
"type": "object",
|
| 821 |
-
"description":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 822 |
"additionalProperties": True,
|
| 823 |
},
|
| 824 |
},
|
|
|
|
| 6 |
|
| 7 |
import asyncio
|
| 8 |
import base64
|
| 9 |
+
import os
|
| 10 |
from typing import Any, Dict, Literal, Optional
|
| 11 |
|
| 12 |
from huggingface_hub import HfApi
|
|
|
|
| 61 |
]
|
| 62 |
|
| 63 |
# Constants
|
|
|
|
| 64 |
UV_DEFAULT_IMAGE = "ghcr.io/astral-sh/uv:python3.12-bookworm"
|
| 65 |
|
| 66 |
|
| 67 |
+
def _substitute_hf_token(params: Dict[str, Any] | None) -> Dict[str, Any] | None:
|
| 68 |
+
"""
|
| 69 |
+
Substitute $HF_TOKEN with actual token value from environment.
|
| 70 |
+
|
| 71 |
+
Args:
|
| 72 |
+
params: Dictionary that may contain "$HF_TOKEN" in values
|
| 73 |
+
|
| 74 |
+
Returns:
|
| 75 |
+
Dictionary with $HF_TOKEN substituted
|
| 76 |
+
"""
|
| 77 |
+
if params is None:
|
| 78 |
+
return None
|
| 79 |
+
|
| 80 |
+
result = {}
|
| 81 |
+
for key, value in params.items():
|
| 82 |
+
if value == "$HF_TOKEN":
|
| 83 |
+
result[key] = os.environ.get("HF_TOKEN", "")
|
| 84 |
+
else:
|
| 85 |
+
result[key] = value
|
| 86 |
+
|
| 87 |
+
return result
|
| 88 |
+
|
| 89 |
+
|
| 90 |
def _build_uv_command(
|
| 91 |
script: str,
|
| 92 |
with_deps: list[str] | None = None,
|
|
|
|
| 126 |
return f'echo "{encoded}" | base64 -d | {uv_command_str}'
|
| 127 |
|
| 128 |
|
| 129 |
+
def _ensure_hf_transfer_dependency(deps: list[str] | None) -> list[str]:
|
| 130 |
+
"""Ensure hf-transfer is included in the dependencies list"""
|
| 131 |
+
if deps is None:
|
| 132 |
+
return ["hf-transfer"]
|
| 133 |
+
|
| 134 |
+
if isinstance(deps, list):
|
| 135 |
+
deps_copy = deps.copy() # Don't modify the original
|
| 136 |
+
if "hf-transfer" not in deps_copy:
|
| 137 |
+
deps_copy.append("hf-transfer")
|
| 138 |
+
return deps_copy
|
| 139 |
+
|
| 140 |
+
return ["hf-transfer"]
|
| 141 |
+
|
| 142 |
+
|
| 143 |
def _resolve_uv_command(
|
| 144 |
script: str,
|
| 145 |
with_deps: list[str] | None = None,
|
|
|
|
| 353 |
{{
|
| 354 |
"operation": "uv",
|
| 355 |
"args": {{
|
| 356 |
+
"script": "import random\\nprint(42 + random.randint(1, 5))",
|
| 357 |
+
"dependencies" : ["torch", "huggingface_hub"]
|
| 358 |
}}
|
| 359 |
}}
|
| 360 |
```
|
|
|
|
| 373 |
**String format (simple cases only):**
|
| 374 |
- Still accepted for backwards compatibility, parsed with POSIX shell semantics
|
| 375 |
- Rejects shell operators and can mis-handle characters such as `&`; switch to arrays when things turn complex
|
|
|
|
| 376 |
|
| 377 |
### Show command-specific help
|
| 378 |
Call this tool with:
|
|
|
|
| 382 |
|
| 383 |
## Tips
|
| 384 |
|
| 385 |
+
- Jobs default to non-detached mode (stream logs until completion). Set `detach: true` to return immediately.
|
| 386 |
- Prefer array commands to avoid shell parsing surprises
|
| 387 |
+
- To access private Hub assets (spaces, private models, datasets, collections), pass `secrets: {{ "HF_TOKEN": "$HF_TOKEN" }}`
|
| 388 |
+
- Before calling a job, think about dependencies (they must be specified), which hardware flavor to run on (choose simplest for task), and whether to include secrets.
|
| 389 |
"""
|
| 390 |
return {"formatted": usage_text, "totalResults": 1, "resultsShared": 1}
|
| 391 |
|
|
|
|
| 394 |
help_text = f"Help for operation: {operation}\n\nCall with appropriate arguments. Use the main help for examples."
|
| 395 |
return {"formatted": help_text, "totalResults": 1, "resultsShared": 1}
|
| 396 |
|
| 397 |
+
async def _wait_for_job_completion(
|
| 398 |
+
self, job_id: str, namespace: Optional[str] = None
|
| 399 |
+
) -> tuple[str, list[str]]:
|
| 400 |
+
"""
|
| 401 |
+
Stream job logs until completion, printing them in real-time.
|
| 402 |
+
|
| 403 |
+
Returns:
|
| 404 |
+
tuple: (final_status, all_logs)
|
| 405 |
+
"""
|
| 406 |
+
all_logs = []
|
| 407 |
+
|
| 408 |
+
# Fetch logs - generator streams logs as they arrive and ends when job completes
|
| 409 |
+
logs_gen = self.api.fetch_job_logs(job_id=job_id, namespace=namespace)
|
| 410 |
+
|
| 411 |
+
# Stream logs in real-time
|
| 412 |
+
for log_line in logs_gen:
|
| 413 |
+
print("\t" + log_line)
|
| 414 |
+
all_logs.append(log_line)
|
| 415 |
+
|
| 416 |
+
# After logs complete, fetch final job status
|
| 417 |
+
job_info = await _async_call(
|
| 418 |
+
self.api.inspect_job, job_id=job_id, namespace=namespace
|
| 419 |
+
)
|
| 420 |
+
final_status = job_info.status.stage
|
| 421 |
+
|
| 422 |
+
return final_status, all_logs
|
| 423 |
+
|
| 424 |
async def _run_job(self, args: Dict[str, Any]) -> ToolResult:
|
| 425 |
"""Run a job using HfApi.run_job()"""
|
| 426 |
try:
|
|
|
|
| 428 |
self.api.run_job,
|
| 429 |
image=args.get("image", "python:3.12"),
|
| 430 |
command=args.get("command"),
|
| 431 |
+
env=_substitute_hf_token(args.get("env")),
|
| 432 |
+
secrets=_substitute_hf_token(args.get("secrets")),
|
| 433 |
flavor=args.get("flavor", "cpu-basic"),
|
| 434 |
timeout=args.get("timeout", "30m"),
|
| 435 |
namespace=args.get("namespace") or self.namespace,
|
|
|
|
| 447 |
To inspect, call this tool with `{{"operation": "inspect", "args": {{"job_id": "{job.id}"}}}}`"""
|
| 448 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 449 |
|
| 450 |
+
# Not detached - wait for completion and stream logs
|
| 451 |
+
print(f"Job started: {job.id}")
|
| 452 |
+
print("Streaming logs...\n---\n")
|
| 453 |
|
| 454 |
+
final_status, all_logs = await self._wait_for_job_completion(
|
| 455 |
+
job_id=job.id,
|
| 456 |
+
namespace=args.get("namespace") or self.namespace,
|
| 457 |
+
)
|
| 458 |
|
| 459 |
+
# Format all logs for the agent
|
| 460 |
+
log_text = "\n".join(all_logs) if all_logs else "(no logs)"
|
| 461 |
+
|
| 462 |
+
response = f"""Job completed!
|
| 463 |
+
|
| 464 |
+
**Job ID:** {job.id}
|
| 465 |
+
**Final Status:** {final_status}
|
| 466 |
+
**View at:** {job.url}
|
| 467 |
+
|
| 468 |
+
**Logs:**
|
| 469 |
+
```
|
| 470 |
+
{log_text}
|
| 471 |
+
```"""
|
| 472 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 473 |
|
| 474 |
except Exception as e:
|
|
|
|
| 481 |
if not script:
|
| 482 |
raise ValueError("script is required")
|
| 483 |
|
| 484 |
+
# Get dependencies and ensure hf-transfer is included
|
| 485 |
+
deps = (
|
| 486 |
+
args.get("with_deps")
|
| 487 |
+
or args.get("dependencies")
|
| 488 |
+
or args.get("packages")
|
| 489 |
+
)
|
| 490 |
+
deps = _ensure_hf_transfer_dependency(deps)
|
| 491 |
+
|
| 492 |
# Resolve the command based on script type (URL, inline, or file)
|
| 493 |
command = _resolve_uv_command(
|
| 494 |
script=script,
|
| 495 |
+
with_deps=deps,
|
| 496 |
python=args.get("python"),
|
| 497 |
script_args=args.get("script_args"),
|
| 498 |
)
|
|
|
|
| 502 |
self.api.run_job,
|
| 503 |
image=UV_DEFAULT_IMAGE,
|
| 504 |
command=command,
|
| 505 |
+
env=_substitute_hf_token(args.get("env")),
|
| 506 |
+
secrets=_substitute_hf_token(args.get("secrets")),
|
| 507 |
+
flavor=args.get("flavor") or args.get("hardware") or "cpu-basic",
|
| 508 |
timeout=args.get("timeout", "30m"),
|
| 509 |
namespace=args.get("namespace") or self.namespace,
|
| 510 |
)
|
| 511 |
|
| 512 |
+
# If detached, return immediately
|
| 513 |
+
if args.get("detach", False):
|
| 514 |
+
response = f"""UV Job started successfully!
|
| 515 |
|
| 516 |
+
**Job ID:** {job.id}
|
| 517 |
**Status:** {job.status.stage}
|
| 518 |
**View at:** {job.url}
|
| 519 |
|
| 520 |
+
To check logs, call this tool with `{{"operation": "logs", "args": {{"job_id": "{job.id}"}}}}`"""
|
| 521 |
+
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 522 |
+
|
| 523 |
+
# Not detached - wait for completion and stream logs
|
| 524 |
+
print(f"UV Job started: {job.id}")
|
| 525 |
+
print("Streaming logs...\n---\n")
|
| 526 |
+
|
| 527 |
+
final_status, all_logs = await self._wait_for_job_completion(
|
| 528 |
+
job_id=job.id,
|
| 529 |
+
namespace=args.get("namespace") or self.namespace,
|
| 530 |
+
)
|
| 531 |
+
|
| 532 |
+
# Format all logs for the agent
|
| 533 |
+
log_text = "\n".join(all_logs) if all_logs else "(no logs)"
|
| 534 |
+
|
| 535 |
+
response = f"""UV Job completed!
|
| 536 |
+
|
| 537 |
+
**Job ID:** {job.id}
|
| 538 |
+
**Final Status:** {final_status}
|
| 539 |
+
**View at:** {job.url}
|
| 540 |
+
|
| 541 |
+
**Logs:**
|
| 542 |
+
```
|
| 543 |
+
{log_text}
|
| 544 |
+
```"""
|
| 545 |
return {"formatted": response, "totalResults": 1, "resultsShared": 1}
|
| 546 |
|
| 547 |
except Exception as e:
|
|
|
|
| 691 |
image=args.get("image", "python:3.12"),
|
| 692 |
command=args.get("command"),
|
| 693 |
schedule=args.get("schedule"),
|
| 694 |
+
env=_substitute_hf_token(args.get("env")),
|
| 695 |
+
secrets=_substitute_hf_token(args.get("secrets")),
|
| 696 |
flavor=args.get("flavor", "cpu-basic"),
|
| 697 |
timeout=args.get("timeout", "30m"),
|
| 698 |
namespace=args.get("namespace") or self.namespace,
|
|
|
|
| 726 |
if not schedule:
|
| 727 |
raise ValueError("schedule is required")
|
| 728 |
|
| 729 |
+
# Get dependencies and ensure hf-transfer is included
|
| 730 |
+
deps = (
|
| 731 |
+
args.get("with_deps")
|
| 732 |
+
or args.get("dependencies")
|
| 733 |
+
or args.get("packages")
|
| 734 |
+
)
|
| 735 |
+
deps = _ensure_hf_transfer_dependency(deps)
|
| 736 |
+
|
| 737 |
# Resolve the command based on script type
|
| 738 |
command = _resolve_uv_command(
|
| 739 |
script=script,
|
| 740 |
+
with_deps=deps,
|
| 741 |
python=args.get("python"),
|
| 742 |
script_args=args.get("script_args"),
|
| 743 |
)
|
|
|
|
| 748 |
image=UV_DEFAULT_IMAGE,
|
| 749 |
command=command,
|
| 750 |
schedule=schedule,
|
| 751 |
+
env=_substitute_hf_token(args.get("env")),
|
| 752 |
+
secrets=_substitute_hf_token(args.get("secrets")),
|
| 753 |
+
flavor=args.get("flavor") or args.get("hardware") or "cpu-basic",
|
| 754 |
timeout=args.get("timeout", "30m"),
|
| 755 |
namespace=args.get("namespace") or self.namespace,
|
| 756 |
)
|
|
|
|
| 909 |
"description": (
|
| 910 |
"Manage Hugging Face CPU/GPU compute jobs. Run commands in Docker containers, "
|
| 911 |
"execute Python scripts with UV. List, schedule and monitor jobs/logs. "
|
| 912 |
+
"Example hardware/flavor: cpu-basic, cpu-performance, t4-medium. "
|
| 913 |
"Call this tool with no operation for full usage instructions and examples."
|
| 914 |
),
|
| 915 |
"parameters": {
|
|
|
|
| 940 |
},
|
| 941 |
"args": {
|
| 942 |
"type": "object",
|
| 943 |
+
"description": (
|
| 944 |
+
"Operation-specific arguments as a JSON object. "
|
| 945 |
+
"Common args: script (for uv), packages/dependencies (array), "
|
| 946 |
+
"flavor/hardware (e.g., a10g-large, cpu-basic), command (array), "
|
| 947 |
+
"image (string), env (object), secrets (object)."
|
| 948 |
+
),
|
| 949 |
"additionalProperties": True,
|
| 950 |
},
|
| 951 |
},
|