mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-21 12:30:48 +08:00
modal: launch.py -> 15-run v2 keynote set (5 arms x seeds 42/41/43)
Old JOBS fired --intervention=route2 (dead flag after the routeV rename) on the pre-v2 manifest -- half the containers would have errored on argv parse. Replace with the n=3 keynote set generated from ARMS x SEEDS: vanilla, routeV real-V per-rollout, routeV per-token, random-V(157), placebo(vampire). Tag stems match the local pueue twins so Modal and local cross-replicate. id 1 = canary (seed-42 vanilla). Fix app.py::smoke route2->routeV and the subprocess modal binary (not on PATH; resolve next to sys.executable). v2 eval rides in via the runtime-mounted src/. Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
+2
-3
@@ -64,10 +64,9 @@ that costs ~5 min GPU per run; uploading the prebuilt ones is cheaper.)
|
||||
## Verify one run, then fan out
|
||||
|
||||
```bash
|
||||
modal run modal/app.py --action smoke # 4-step route2 sanity
|
||||
modal run modal/app.py --argv "fast --intervention=route2 --seed=43 --steps=60 --eval-ablate-every=10 --out-tag=_modal_verify"
|
||||
modal run modal/launch.py --only 1 # canary: seed-42 vanilla, confirm clean v2 FINAL EVAL
|
||||
# compare its per_mode_deploy.json to the local-box artifact for the same args
|
||||
modal run modal/launch.py # all 12; or --only 127,134
|
||||
modal run modal/launch.py # all 15 (5 arms x seeds 42/41/43)
|
||||
```
|
||||
|
||||
## Getting the outputs back
|
||||
|
||||
+8
-8
@@ -23,9 +23,9 @@ Design notes / deliberate choices (see modal/README.md for the runbook):
|
||||
|
||||
Usage:
|
||||
modal run modal/app.py::warm # download model + build svd_cache once
|
||||
modal run modal/app.py::smoke # 4-step route2 sanity on the real model
|
||||
modal run modal/app.py::train --argv "fast --intervention=route2 --seed=43 --steps=60 ..."
|
||||
modal run modal/launch.py # fan out jobs 124-135 (see launch.py)
|
||||
modal run modal/app.py::smoke # 4-step routeV sanity on the real model
|
||||
modal run modal/app.py::train --argv "fast --intervention=routeV --seed=43 --steps=60 ..."
|
||||
modal run modal/launch.py # fan out the 15-run v2 keynote set (see launch.py)
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
@@ -174,7 +174,7 @@ def _run_train(argv: list[str]) -> dict:
|
||||
@app.function(gpu=GPU, volumes={CACHE: cache}, timeout=TIMEOUT, secrets=SECRETS)
|
||||
def train(argv: str) -> dict:
|
||||
"""Run one `vgrout.train` invocation. `argv` is the CLI string after
|
||||
`python -m vgrout.train`, e.g. "fast --intervention=route2 --seed=43 --steps=60"."""
|
||||
`python -m vgrout.train`, e.g. "fast --intervention=routeV --seed=43 --steps=60"."""
|
||||
return _run_train(shlex.split(argv))
|
||||
|
||||
|
||||
@@ -190,11 +190,11 @@ def warm() -> dict:
|
||||
|
||||
@app.function(gpu=GPU, volumes={CACHE: cache}, timeout=TIMEOUT, secrets=SECRETS)
|
||||
def smoke() -> dict:
|
||||
"""4-step real-model route2 sanity (the user's smoke gate before fan-out).
|
||||
"""4-step real-model routeV sanity (the user's smoke gate before fan-out).
|
||||
Needs the FastConfig default inputs on the Volume: out/pairsets/prog_wide.json
|
||||
+ out/pools/substrate (upload via modal/upload_inputs.py first)."""
|
||||
+ the teacher pool (upload via modal/upload_inputs.py first)."""
|
||||
return _run_train(shlex.split(
|
||||
"fast --intervention=route2 --seed=43 --steps=4 --eval-ablate-every=2 "
|
||||
"fast --intervention=routeV --seed=43 --steps=4 --eval-ablate-every=2 "
|
||||
"--eval-n-prompts=2 --out-tag=_modal_smoke"
|
||||
))
|
||||
|
||||
@@ -203,7 +203,7 @@ def smoke() -> dict:
|
||||
def main(argv: str = "", action: str = "train"):
|
||||
"""`modal run modal/app.py --action warm`
|
||||
`modal run modal/app.py --action smoke`
|
||||
`modal run modal/app.py --argv "fast --intervention=route2 --seed=43 --steps=60 ..."`"""
|
||||
`modal run modal/app.py --argv "fast --intervention=routeV --seed=43 --steps=60 ..."`"""
|
||||
if action == "warm":
|
||||
print(warm.remote())
|
||||
elif action == "smoke":
|
||||
|
||||
+32
-23
@@ -1,14 +1,17 @@
|
||||
"""Fan out the 2026-06-06 manifest's Running/Queued jobs (124-135) as parallel
|
||||
Modal containers. argv copied verbatim from docs/spec/20260606_job_manifest.md
|
||||
(the `vgrout.train` entries), so each Modal run == the pueue run it replaces.
|
||||
"""Fan out the n=3 v2-eval keynote set as parallel Modal containers.
|
||||
|
||||
Run only AFTER `modal run modal/app.py --action smoke` is verified clean.
|
||||
15 runs = 5 arms x seeds {42,41,43}, argv mirrors the local pueue runs (same
|
||||
preset/steps/eval cadence), so each Modal run == its local twin and the two
|
||||
environments cross-replicate. All on the v2 train/test token-gap eval (the
|
||||
mounted src/ carries the current committed code).
|
||||
|
||||
modal run modal/launch.py # all 12
|
||||
modal run modal/launch.py --only 127,134 # a subset (verify-one path)
|
||||
id 1 = the canary: seed-42 vanilla (needs only the shared teacher pool, no
|
||||
pairset/direction inputs). Fire it alone first, then fan out once it's green:
|
||||
.venv/bin/modal run modal/launch.py --only 1 # canary -> clean v2 FINAL EVAL
|
||||
.venv/bin/modal run modal/launch.py # all 15
|
||||
|
||||
Each container writes out/runs/<ts>_<tag>/per_mode_deploy.json to the Volume;
|
||||
this entrypoint also copies them into modal/results/<tag>.json locally.
|
||||
this entrypoint pulls the full run dir + log back into out/runs/ locally.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
@@ -21,33 +24,39 @@ sys.path.insert(0, str(Path(__file__).parent))
|
||||
from app import app, train # noqa: E402 (same dir; registers the functions)
|
||||
|
||||
VOL = "vgrout-cache"
|
||||
# `modal` isn't on PATH; the CLI binary sits next to the interpreter running
|
||||
# this entrypoint (.venv/bin/python -> .venv/bin/modal).
|
||||
MODAL_BIN = str(Path(sys.executable).with_name("modal"))
|
||||
|
||||
|
||||
def _pull_dir(remote: str, local_parent: Path):
|
||||
"""modal recreates the remote leaf dir UNDER the target, so pass the parent."""
|
||||
local_parent.mkdir(parents=True, exist_ok=True)
|
||||
subprocess.run(["modal", "volume", "get", "--force", VOL, remote, str(local_parent)], check=False)
|
||||
subprocess.run([MODAL_BIN, "volume", "get", "--force", VOL, remote, str(local_parent)], check=False)
|
||||
|
||||
|
||||
def _pull_file(remote: str, local: Path):
|
||||
local.parent.mkdir(parents=True, exist_ok=True)
|
||||
subprocess.run(["modal", "volume", "get", "--force", VOL, remote, str(local)], check=False)
|
||||
subprocess.run([MODAL_BIN, "volume", "get", "--force", VOL, remote, str(local)], check=False)
|
||||
|
||||
# job id -> argv after `python -m vgrout.train`. Verbatim from the manifest.
|
||||
JOBS: dict[int, str] = {
|
||||
124: "fast --intervention=route2 --seed=41 --teacher-off-step=40 --steps=200 --eval-ablate-every=20 --out-tag=_route2_toff40_s41",
|
||||
125: "fast --intervention=route --seed=41 --v-hack-path=out/vhack/v_hack_pairset_prog_wide_randomV.safetensors --vhack-refresh-every=0 --eval-ablate-every=5 --steps=60 --out-tag=_route_randomV_s41",
|
||||
126: "fast --intervention=route2 --seed=41 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --v-hack-path=out/vhack/v_hack_a5_runtests.safetensors --steps=200 --eval-ablate-every=10 --eval-n-prompts=24 --gate-anchor-teacher-only --out-tag=_a5_route2_teacheronly_s41",
|
||||
127: "fast --intervention=erase --seed=41 --eval-ablate-every=5 --out-tag=_erase_realv_s41",
|
||||
128: "fast --intervention=erase --seed=41 --vhack-pairs-path=out/pairsets/null_city.json --eval-ablate-every=5 --out-tag=_erase_placebo_nullcity_s41",
|
||||
129: "fast --intervention=none --seed=41 --beta=1e-5 --adam-beta1=0.9 --adam-beta2=0.99 --steps=200 --eval-ablate-every=20 --out-tag=_none200_kl5_s41",
|
||||
130: "fast --intervention=route2 --seed=41 --beta=1e-5 --adam-beta1=0.9 --adam-beta2=0.99 --steps=200 --eval-ablate-every=20 --out-tag=_route2200_kl5_s41",
|
||||
131: "fast --intervention=none --seed=42 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --steps=200 --eval-ablate-every=10 --eval-n-prompts=24 --out-tag=_a5_vanilla_tmrt_s42",
|
||||
132: "fast --intervention=none --seed=43 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --steps=200 --eval-ablate-every=10 --eval-n-prompts=24 --out-tag=_a5_vanilla_tmrt_s43",
|
||||
133: "fast --intervention=route2 --seed=42 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --v-hack-path=out/vhack/v_hack_a5_runtests.safetensors --steps=200 --eval-ablate-every=10 --eval-n-prompts=24 --gate-anchor-teacher-only --out-tag=_a5_route2_teacheronly_s42",
|
||||
134: "fast --intervention=route2 --seed=43 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --v-hack-path=out/vhack/v_hack_a5_runtests.safetensors --steps=200 --eval-ablate-every=10 --eval-n-prompts=24 --gate-anchor-teacher-only --out-tag=_a5_route2_teacheronly_s43",
|
||||
135: "fast --intervention=route2 --seed=41 --teacher-pool-dir=out/pools/substrate --teacher-modes run_tests --route2-random-v-seed=0 --steps=200 --eval-ablate-every=20 --eval-n-prompts=24 --gate-anchor-teacher-only --out-tag=_a5_haar_d0_teacheronly_s41",
|
||||
# The keynote arms (A2 method+baseline, A3 ablation). Flags after the seed;
|
||||
# tag stems match the local pueue runs so Modal and local artifacts line up.
|
||||
COMMON = "fast --eval-ablate-every=10 --steps=60"
|
||||
ARMS: dict[str, str] = {
|
||||
"vanilla": "--intervention=none",
|
||||
"routeV_real": "--intervention=routeV --vhack-pairs-path=out/pairsets/prog_wide.json --vhack-refresh-every=5",
|
||||
"routeV_pertok_real": "--intervention=routeV --vhack-pairs-path=out/pairsets/prog_wide.json --vhack-refresh-every=5 --routeV-per-token",
|
||||
"routeV_randomV157": "--intervention=routeV --vhack-pairs-path=out/pairsets/prog_wide.json --vhack-refresh-every=5 --routeV-random-v-seed=157",
|
||||
"routeV_placebo_vampire": "--intervention=routeV --vhack-pairs-path=out/pairsets/null_vampire.json --vhack-refresh-every=5",
|
||||
}
|
||||
# seed 42 first so id 1 == the canary (vanilla s42, the seed missing locally).
|
||||
SEEDS = [42, 41, 43]
|
||||
JOBS: dict[int, str] = {}
|
||||
_jid = 1
|
||||
for _s in SEEDS:
|
||||
for _arm, _flags in ARMS.items():
|
||||
JOBS[_jid] = f"{COMMON} --seed={_s} {_flags} --out-tag=_{_arm}_s{_s}"
|
||||
_jid += 1
|
||||
|
||||
|
||||
@app.local_entrypoint()
|
||||
|
||||
Reference in New Issue
Block a user