File size: 15,162 Bytes
a384097 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 | """ModalSpawnExecutor — production Modal-backed serverless executor.
This is the v0-finished sibling of `ModalExecutor` (which remains a
skeleton per Wave 18 contract). The skeleton class stays unchanged to
preserve `test_skeleton_executors.py`'s pinned NotImplementedError
contract; this class is the working alternative for users who want
real Modal execution.
Design choices vs the skeleton's docstring:
1. **User-provided `modal.Function` instead of internal app construction.**
The skeleton showed a pattern where ModalExecutor builds its own
`modal.App` and registers `run_replica` internally. That couples the
executor to image/GPU/Volume choices the user actually wants to own.
Instead, ModalSpawnExecutor takes a *pre-decorated* `modal.Function`
from the caller — the user defines:
@app.function(gpu="H100:4", image=my_image, volumes={"/vol": vol},
secrets=[modal.Secret.from_name("hf-token")],
timeout=4*3600)
def run_replica(rendezvous_uri: str, world_size: int,
rank: int, **entrypoint_args):
import os
os.environ["REPLICA_RANK"] = str(rank)
from composer_replication.diloco.serverless import (
MockManager, ObjectStoreAllReduce,
)
store = ObjectStoreAllReduce(rendezvous_uri, rank=rank,
world_size=world_size)
manager = MockManager(store)
# ... user's training loop with this manager ...
then constructs:
executor = ModalSpawnExecutor(modal_function=run_replica)
handles = executor.launch_replicas(
n_replicas=4,
entrypoint=run_replica, # ignored — function is bound
entrypoint_args={"rendezvous_uri": "/vol/diloco/run42",
"world_size": 4},
)
2. **Rank as explicit kwarg, not env-var indirection.** Modal Functions
start with a clean env, so the rank-via-env pattern that
LocalProcessExecutor uses is fragile here (Modal would need
container-level env injection per call, which `modal.Secret.from_dict`
does but adds a round-trip per spawn). We pass rank as a kwarg to
`.spawn(rank=i)` so it's plumbed through Modal's call args directly.
3. **Handle metadata = `call_id`, no in-process state.** Unlike
LocalProcessExecutor (which holds Process refs), this executor is
stateless after launch — handles are reconstructed via
`modal.FunctionCall.from_id(call_id)` for poll/cancel/collect.
Lets the executor survive process restart mid-run.
References:
- modal-client 1.4.x docs on FunctionCall: https://modal.com/docs/reference/modal.FunctionCall
- ADR-005 (executor protocol design)
"""
from __future__ import annotations
import time
from typing import Any, Callable, Mapping
from composer_replication.diloco.serverless.executor import (
ReplicaHandle,
ServerlessExecutor,
)
class ModalSpawnExecutor:
"""Run replicas as parallel Modal Function spawns.
Implements the `ServerlessExecutor` Protocol against Modal's
`Function.spawn()` API. The user must provide a pre-decorated
`modal.Function` (with `@app.function(...)` already applied) — see
module docstring for the expected signature.
Args:
modal_function: a `modal.Function` registered against a `modal.App`.
Must accept at minimum `rank: int` plus the kwargs in
`entrypoint_args`. Image / GPU / Volume / Secret / timeout
are pinned on the decorator and the executor won't override
them.
deploy: if True, calls `modal_function.app.deploy()` before
spawning. Required when running outside a `modal run` context
(e.g. from a regular Python script). Default False — assumes
the user is inside a `modal run` block where the app is
already live.
Raises:
RuntimeError: if `modal` client is not installed.
TypeError: if `modal_function` is not a `modal.Function`.
"""
backend_name = "modal_spawn"
supports_inter_replica_network = False # Modal containers are isolated by default
def __init__(
self,
modal_function: Any,
*,
deploy: bool = False,
) -> None:
try:
import modal # noqa: F401
except ImportError as e:
raise RuntimeError(
"ModalSpawnExecutor requires the modal client. Install with "
"`pip install modal` and configure with `modal token new`. "
f"Got: {e!r}"
)
# Duck-type check — modal.Function objects expose .spawn / .remote /
# ._app, which the user-supplied function will have if they used the
# @app.function(...) decorator. We avoid `isinstance(_, modal.Function)`
# to stay tolerant of modal-client minor-version changes that may
# restructure the class.
if not (hasattr(modal_function, "spawn") and hasattr(modal_function, "remote")):
raise TypeError(
f"modal_function must be a modal.Function (decorated via "
f"`@app.function(...)`). Got {type(modal_function)!r} which "
f"has no `.spawn()` method. "
f"See ModalSpawnExecutor docstring for expected signature."
)
self.modal_function = modal_function
self._deploy_requested = deploy
self._deployed = False
self._handles: dict[int, dict[str, Any]] = {}
# -----------------------------------------------------------------
# Lifecycle
# -----------------------------------------------------------------
def _maybe_deploy(self) -> None:
if self._deploy_requested and not self._deployed:
# `modal_function.app` exposes the underlying App. Calling
# `.deploy()` registers it with Modal so spawn() works from
# outside `modal run`.
app = getattr(self.modal_function, "app", None)
if app is None:
raise RuntimeError(
"modal_function.app is None — can't deploy. The function "
"must have been decorated against a real modal.App."
)
app.deploy()
self._deployed = True
# -----------------------------------------------------------------
# ServerlessExecutor Protocol
# -----------------------------------------------------------------
def launch_replicas(
self,
n_replicas: int,
entrypoint: str | Callable[..., Any],
entrypoint_args: Mapping[str, Any],
*,
gpu: str | None = None,
timeout: int = 3600,
) -> list[ReplicaHandle]:
"""Spawn N parallel Modal Function calls.
Note: `entrypoint` is **ignored** — the actual entrypoint is the
`modal_function` passed to `__init__`. This keeps the executor
Protocol-compatible while preserving the user's image/GPU
decoration. `gpu` and `timeout` are similarly ignored (pinned
on the function decorator).
"""
del entrypoint, gpu, timeout # pinned on the decorated function
if n_replicas < 1:
raise ValueError(f"n_replicas must be >= 1, got {n_replicas}")
self._maybe_deploy()
# Strip rank_env if present — we use explicit `rank` kwarg.
spawn_kwargs = {k: v for k, v in entrypoint_args.items()
if k != "rank_env"}
handles: list[ReplicaHandle] = []
for rank in range(n_replicas):
try:
fcall = self.modal_function.spawn(rank=rank, **spawn_kwargs)
except Exception as e:
# Best-effort cancel any already-launched siblings
for prior in handles:
try:
self.cancel(prior)
except Exception:
pass
raise RuntimeError(
f"ModalSpawnExecutor.launch_replicas failed at rank={rank} "
f"of {n_replicas} (already-launched siblings cancelled). "
f"Underlying error: {e!r}"
) from e
handle = ReplicaHandle(
rank=rank,
backend_name=self.backend_name,
metadata={
"call_id": fcall.object_id,
"spawn_ts": time.time(),
},
)
self._handles[rank] = {
"fcall": fcall,
"result": None,
}
handles.append(handle)
return handles
def poll(self, handle: ReplicaHandle) -> str:
"""Poll a Modal call's status.
Modal's FunctionCall doesn't expose a non-blocking status getter
directly (the API is `.get(timeout=...)`), so we poll by trying
`.get(timeout=0)` and treating Timeout/Pending as "running".
Returns one of: "pending" | "running" | "succeeded" | "failed" |
"cancelled".
"""
meta = self._handles.get(handle.rank)
if meta is None:
return "cancelled"
# If we already collected this one, return cached result
if meta["result"] is not None:
return meta["result"]["status"]
import modal
from modal.exception import OutputExpiredError
fcall = meta["fcall"]
# Re-hydrate to get fresh state
try:
# `.get(timeout=0)` returns immediately if done; raises TimeoutError otherwise.
result_value = fcall.get(timeout=0)
meta["result"] = {
"rank": handle.rank,
"status": "succeeded",
"exit_code": 0,
"error": None,
"result": result_value,
"call_id": handle.metadata.get("call_id"),
}
return "succeeded"
except TimeoutError:
return "running"
except OutputExpiredError as e:
meta["result"] = {
"rank": handle.rank,
"status": "failed",
"exit_code": 1,
"error": f"OutputExpiredError: {e!r}",
"result": None,
"call_id": handle.metadata.get("call_id"),
}
return "failed"
except Exception as e:
# User-code exception bubbles up here as the original exception class
meta["result"] = {
"rank": handle.rank,
"status": "failed",
"exit_code": 1,
"error": f"{type(e).__name__}: {e!r}",
"result": None,
"call_id": handle.metadata.get("call_id"),
}
return "failed"
def stream_logs(self, handle: ReplicaHandle, *, n_lines: int = 200) -> str:
"""Read recent Modal logs for this call.
Modal exposes per-FunctionCall logs via the dashboard URL. The
client API doesn't expose log-streaming directly in 1.4.x, so we
return a pointer to the dashboard URL plus any captured error
from poll().
"""
meta = self._handles.get(handle.rank)
if meta is None:
return f"<replica {handle.rank}: no metadata>"
call_id = handle.metadata.get("call_id", "<unknown>")
try:
dashboard_url = meta["fcall"].get_dashboard_url()
except Exception:
dashboard_url = (
f"https://modal.com/apps/<workspace>/<env>/calls/{call_id}"
)
if meta.get("result"):
err = meta["result"].get("error") or "<no error>"
return (
f"[rank {handle.rank}] call_id={call_id}\n"
f" Dashboard: {dashboard_url}\n"
f" Result: {meta['result']['status']}\n"
f" Error: {err[-2000:] if err else '<none>'}"
)
return (
f"[rank {handle.rank}] call_id={call_id} (still running)\n"
f" Dashboard: {dashboard_url}\n"
f" Logs not streamable via client API in modal-client 1.4.x; "
f"use the dashboard URL or `modal app logs <app-id>`."
)
def cancel(self, handle: ReplicaHandle) -> None:
"""Best-effort cancel of a Modal call."""
meta = self._handles.get(handle.rank)
if meta is None:
return
try:
meta["fcall"].cancel()
except Exception:
# Already terminated, network blip, etc. — best-effort.
pass
def collect(
self,
handles: list[ReplicaHandle],
*,
timeout: int | None = None,
) -> list[dict[str, Any]]:
"""Block until all replicas finish; return per-replica result dicts.
Modal's `.get(timeout=...)` blocks until the call completes or
the timeout elapses. We iterate handles, calling `.get()` with
the remaining time budget, so the cumulative wall-clock is
bounded by `timeout`.
"""
deadline = time.time() + (timeout if timeout is not None else 86400)
results: list[dict[str, Any]] = []
for h in handles:
meta = self._handles.get(h.rank)
if meta is None:
results.append({
"rank": h.rank,
"status": "cancelled",
"exit_code": None,
"error": "handle has no metadata (cancelled or unknown)",
"result": None,
"call_id": h.metadata.get("call_id"),
})
continue
# Already collected by an earlier poll()
if meta["result"] is not None:
results.append(meta["result"])
continue
remaining = max(0.0, deadline - time.time())
try:
result_value = meta["fcall"].get(timeout=remaining)
result_dict = {
"rank": h.rank,
"status": "succeeded",
"exit_code": 0,
"error": None,
"result": result_value,
"call_id": h.metadata.get("call_id"),
}
except TimeoutError as e:
result_dict = {
"rank": h.rank,
"status": "running",
"exit_code": None,
"error": f"TimeoutError after deadline: {e!r}",
"result": None,
"call_id": h.metadata.get("call_id"),
}
except Exception as e:
result_dict = {
"rank": h.rank,
"status": "failed",
"exit_code": 1,
"error": f"{type(e).__name__}: {e!r}",
"result": None,
"call_id": h.metadata.get("call_id"),
}
meta["result"] = result_dict
results.append(result_dict)
return results
__all__ = ["ModalSpawnExecutor"]
|