MojiDiff

← experiments

owned-gpu-smoke-2c3b248-7c03a644

failed —

Hypothesis

With the derived image assigned to the correct owned worker, the complete tiny smoke will pass through the verified sibling-container and scoped named-volume topology.

Why run it

Confirms or falsifies the tokenizer, renderer, CUDA forward/backward, checkpoint read-back, and durable named-volume artifact recovery path.

Written result

Status: failed at the adapter-output contract after successful GPU work.

The tokenizer and renderer passed, the RTX 4080 completed one forward/backward step, checkpoint read-back passed, and the 16,179-byte checkpoint was durably recovered with SHA-256 77e93414e04707efb0469718426bb2364d5a7b22e575e519124d5c628442348f. The run is still classified as failed because the inherited NGC entrypoint printed its banner before the JSON result and the adapter correctly rejected mixed stdout. A new run will bypass the image entrypoint explicitly.

State transitions

  1. planned2026-09-20T09:16:31Z
  2. staged2026-09-20T09:17:49Z
  3. failed2026-09-20T09:18:15Zngc_entrypoint_banner_broke_json_contract

Run record

Verbatim from runs/owned-gpu-smoke-2c3b248-7c03a644/run.yaml, the record committed before launch.

schema_version
1
run_id
owned-gpu-smoke-2c3b248-7c03a644
state
failed
parent_run
owned-gpu-smoke-2c3b248-c9e0f1f2
hypothesis
With the derived image assigned to the correct owned worker, the complete tiny smoke will pass through the verified sibling-container and scoped named-volume topology.
expected_information_gain
Confirms or falsifies the tokenizer, renderer, CUDA forward/backward, checkpoint read-back, and durable named-volume artifact recovery path.
code
git_commit
2c3b2488e32c8dc18aa04c0ac43e8933672453be
source_patch_sha256
e3b0c44298fc1c14…
snapshot_mode
git-archive
untracked_experiment_outputs_included
false
config
path
state/hosts.local.yaml
sha256
7c03a6440fe64d2a…
dataset
fixture
embedded-opacity-smoke-fixture
source
scripts/remote/vast_tiny_smoke.py
worker
alias
owned-gpu
gpu
NVIDIA GeForce RTX 4080
driver
595.71.05
image
mojidiff/owned-gpu-smoke:2c3b248
image_digest
sha256:fd065ec98130193b324cf9462fa97e025041bd6df72313249edd902608a2ac00
torch
2.8.0a0+5228986c39.nv25.06
cuda
12.9
cudnn
91002
cairosvg
2.9.0
libcairo2
1.18.0-3build1
workspace_root
/home/dev/workspace
workspace_volume
code-server-gpu_gpubox-workspace
artifact_root
/home/dev/.cache
artifact_volume
code-server-gpu_gpubox-home
artifact_volume_subpath
.cache
execution
seed_set
20250823
deterministic_algorithms
true
resource_cap
max_steps
20000
max_storage_gb
50
smoke_steps
1
network
none
command
owned Docker adapter stage then smoke; encoded remote configuration redacted
checkpoint_policy
retain smoke checkpoint and result in the persistent artifact volume
outputs
local_metadata
runs/owned-gpu-smoke-2c3b248-7c03a644
remote_workspace
/home/dev/workspace/owned-gpu-smoke-2c3b248-7c03a644
durable_artifacts
/home/dev/.cache/owned-gpu-smoke-2c3b248-7c03a644
planned_at
2026-09-20 09:16:31+00:00
failed_at
2026-09-20 09:18:15+00:00
failure_reason
The complete GPU smoke and durable artifact recovery succeeded, but the inherited NGC entrypoint printed a banner before the JSON result. The adapter rejected stdout as invalid JSON, so the run remains failed at the orchestration-contract boundary.