Move examples from cog-examples to cog by anish-sahoo · Pull Request #3055 · replicate/cog

anish-sahoo · 2026-06-11T21:39:17Z

Moves the example models from the separate replicate/cog-examples repo into this repo under examples/, migrates them to the BaseRunner/run API, and adds cog.Secret input support to coglet.

Add example models using run.py/BaseRunner: blur, canary, hello-concurrency, hello-context, hello-image, hello-replicate, hello-train, hello-world, notebook, z-image-turbo
Migrate the existing streaming-text example to run.py/BaseRunner
Replace examples/resnet with the classic torchvision ResNet50 classifier (BaseRunner, ImageNet weights bundled with torchvision -- no managed weights or import step), ported and modernized from replicate/cog-examples
Move the previous managed-weights ResNet example to examples/experimental/resnet-managed-weights (renamed model: resnet-managed-weights)
Update managed-weights cog.yaml to use the model: field (replacing image: <your-registry>/...) and refresh its weights.lock
hello-replicate demonstrates Secret usage by calling the Replicate API

coglet

Add cog.Secret input support: classify and coerce Secret, Optional[Secret], and Secret | None fields, wrapping string values in cog.types.Secret (list[Secret] is intentionally not coerced)
Ignore two pyo3 0.27 advisories in cargo-deny (RUSTSEC-2026-0176, RUSTSEC-2026-0177) with justification; pinned due to numpy/pyo3-async-runtimes lag, and neither code path is reachable in coglet

test harness

Add a base_dir field for repo: local models so local models resolve relative to the manifest (defaults to fixtures/models/)
Point all example models in manifest.yaml at the local examples/ directory instead of replicate/cog-examples
Migrate fixture models from predict.py/BasePredictor to run.py/BaseRunner
Only pass --setup-timeout to cog predict (cog train does not support it)

ask-bonk

I'm Bonk, and I've done a quick review of your PR.

This PR moves example models from the external cog-examples repo into the main cog repo and updates the test harness to support base_dir for local models. Good cleanup overall.

I found two actionable issues:

HIGH: The managed-weights weights.lock was regenerated from incomplete/stub weights. The parakeet model.safetensors is listed as 135 bytes instead of ~2.5GB.
MEDIUM: The hello-train model lost its train_tests entry in the manifest, removing training test coverage even though the example still supports training.

ask-bonk · 2026-06-11T21:46:16Z

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

HIGH — examples/managed-weights/weights.lock was regenerated from incomplete/stub weights. The parakeet model.safetensors is listed as 135 bytes instead of ~2.5GB, and the total size dropped from ~2.5GB to ~1.1MB. Please regenerate from the actual weights or revert this file.
MEDIUM — hello-train lost its train_tests entry in tools/test-harness/manifest.yaml even though the example still contains train.py and a train: field in cog.yaml. This removes training test coverage. I've posted a suggestion to restore it.

github run

ask-bonk

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

HIGH — examples/managed-weights/weights.lock was regenerated from incomplete/stub weights. The parakeet model.safetensors is listed as 135 bytes instead of ~2.5GB, and the total size dropped from ~2.5GB to ~1.1MB. Please regenerate from the actual weights or revert this file.
MEDIUM — hello-train lost its train_tests entry in tools/test-harness/manifest.yaml even though the example still contains train.py and a train: field in cog.yaml. This removes training test coverage. I've posted a suggestion to restore it.
MEDIUM — examples/z-image-turbo/run.py uses tempfile.mktemp() which creates a predictable temporary file path (flagged by CodeQL). Replace with tempfile.NamedTemporaryFile(delete=False) for safety.
MEDIUM — examples/hello-concurrency/run.py has an empty except FileNotFoundError: pass block. Add a logging statement to document the intent.

ask-bonk · 2026-06-11T21:48:55Z

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

HIGH — examples/managed-weights/weights.lock was regenerated from incomplete/stub weights. The parakeet model.safetensors is listed as 135 bytes instead of ~2.5GB, and the total size dropped from ~2.5GB to ~1.1MB. Please regenerate from the actual weights or revert this file.
MEDIUM — hello-train lost its train_tests entry in tools/test-harness/manifest.yaml even though the example still contains train.py and a train: field in cog.yaml. This removes training test coverage. I've posted a suggestion to restore it.
MEDIUM — examples/z-image-turbo/run.py uses tempfile.mktemp() which creates a predictable temporary file path (flagged by CodeQL). Replace with tempfile.NamedTemporaryFile(delete=False) for safety.
MEDIUM — examples/hello-concurrency/run.py has an empty except FileNotFoundError: pass block. Add a logging statement to document the intent.

github run

ask-bonk

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

MEDIUM — examples/blur/run.py returns the built-in input function instead of the image parameter when blur == 0. This is a logic bug.
MEDIUM — hello-train is still missing its train_tests entry in tools/test-harness/manifest.yaml even though the example contains train.py and a train: field in cog.yaml. This removes training test coverage.

ask-bonk · 2026-06-12T16:53:26Z

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

MEDIUM — examples/blur/run.py returns the built-in input function instead of the image parameter when blur == 0. This is a logic bug.
MEDIUM — hello-train is still missing its train_tests entry in tools/test-harness/manifest.yaml even though the example contains train.py and a train: field in cog.yaml. This removes training test coverage.

Both issues have suggestions posted as review comments.

github run

ask-bonk

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

MEDIUM — hello-train is still missing its train_tests entry in tools/test-harness/manifest.yaml even though the example contains train.py and a train: field in cog.yaml. The test harness still fully supports train_tests (see manifest.go and runner.go), and there is no visible replacement mechanism or deprecation marker in the codebase. Removing the only training test without a documented alternative creates a coverage gap.

ask-bonk · 2026-06-12T17:07:41Z

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo and updates the test harness to support base_dir for local models.

Issues found:

MEDIUM — hello-train is still missing its train_tests entry in tools/test-harness/manifest.yaml even though the example contains train.py and a train: field in cog.yaml. The test harness still fully supports train_tests (see manifest.go and runner.go), and there is no visible replacement mechanism or deprecation marker in the codebase. Removing the only training test without a documented alternative creates a coverage gap.

Posted as review comment on PR #3055.

github run

ask-bonk

I'm Bonk, and I've done a quick review of your PR.

Summary: Moves example models from cog-examples into the main repo, updates the test harness for base_dir support, and adds Secret type coercion in coglet.

Issues found:

HIGH — The new Secret coercion in crates/coglet-python/src/input.rs calls Secret(value) in Python, but the Python Secret class in python/cog/types.py does not define an __init__ that accepts arguments. This will raise TypeError: Secret() takes no arguments at runtime when any predictor declares a Secret-typed input.
MEDIUM — hello-train is still missing its train_tests entry in tools/test-harness/manifest.yaml even though the example contains train.py and a train: field in cog.yaml. The test harness still processes train_tests (see manifest.go and runner.go), and there is no deprecation marker or replacement mechanism visible in the code. Removing the only training test creates a coverage gap.

ask-bonk · 2026-06-15T16:35:00Z

Done. I posted a review on PR #3055 with two actionable issues:

HIGH — The new Secret coercion in crates/coglet-python/src/input.rs:288 calls Secret(value), but Python's Secret class has no __init__ and will raise TypeError. I suggested creating the instance with call0() and setting the secret_value attribute directly.
MEDIUM — hello-train still lacks train_tests coverage despite having train.py and a train: field, with no deprecation marker in the codebase.

github run

- Document list[Secret] coercion exclusion in coglet input handling - Fix stale 'predict.py' comments in example cog.yaml files - Correct hello-train README to use 'cog predict' - Add trailing newlines to example cog.yaml files - Only pass --setup-timeout for 'cog predict' (not 'cog train') - Re-add train_tests for hello-train

ask-bonk · 2026-06-15T18:33:44Z

@anish-sahoo Bonk workflow was cancelled.

View workflow run · To retry, trigger Bonk again.

markphelps

Non-Secret review comments from a focused pass on the example migration and test harness. I left Secret/coglet scope comments out separately.

markphelps

I think the coglet Secret input support should be pulled out into a separate PR. This PR is already large and is primarily about moving/modernizing examples; adding new runtime input coercion behavior changes the review surface and deserves its own focused tests/review.

For this PR, can we either:

remove the coglet Secret changes and any example/test-harness dependency on them, or
keep the hello-replicate example only if it works with existing released behavior / reads REPLICATE_API_TOKEN from env rather than requiring new coglet Secret coercion?

Then a follow-up PR can add Secret, Optional[Secret], and any list/edge-case behavior with dedicated coglet tests.

- Remove coglet Secret coercion (moved to dedicated PR #3057) - Remove hello-replicate example and its test-harness entry (moved to #3057) - Drop Secret usage from the scalar-types fixture - Gate managed-weights harness run behind COG_MANAGED_WEIGHTS_READY - Fix stale predict.py/run.py and max:32/max:4 references in hello-concurrency README - Fix hello-context README to match the text input - Fix resnet README to describe GPU + torchvision setup-time weight fetch - Update resnet-managed-weights README to reference model: instead of image:

anish-sahoo · 2026-06-16T15:48:36Z

Re: pulling the coglet Secret support into its own PR — done. The coglet Secret coercion (Secret, Optional[Secret], Secret | None, with list[Secret] excluded) plus dedicated classification tests and the hello-replicate example now live in #3057. This PR drops the coglet Secret changes, the hello-replicate example + its harness entry, and the Secret usage in the scalar-types fixture, so it's back to being just the example migration + test-harness work.

ask-bonk · 2026-06-16T15:56:50Z

LGTM

github run

markphelps

one nit

- Resolve local models by base_dir (relative to repo root) instead of requiring 'repo: local'; precedence is base_dir, then repo, then the default fixtures/models dir. - Reject 'repo: local' in manifest.Load with a migration message. - Drop redundant 'repo: local' from all manifest entries; examples use 'base_dir: examples', fixtures use 'base_dir: tools/test-harness/fixtures/models'. - Add Model.Source() for clearer 'list' output when repo is empty. - Add tests for resolution, repo:local rejection, and Source().

markphelps

looks great! thank you for taking this on

feat: add examples

9d15254

github-advanced-security AI found potential problems Jun 11, 2026

View reviewed changes

Comment thread examples/z-image-turbo/run.py Fixed

github-code-quality Bot found potential problems Jun 11, 2026

View reviewed changes

Comment thread examples/hello-concurrency/run.py Fixed

format: go and python

c846fec

ask-bonk Bot reviewed Jun 11, 2026

View reviewed changes

Comment thread examples/managed-weights/weights.lock Outdated

Comment thread tools/test-harness/manifest.yaml

ask-bonk Bot reviewed Jun 11, 2026

View reviewed changes

Comment thread examples/managed-weights/weights.lock

Comment thread tools/test-harness/manifest.yaml

Comment thread examples/z-image-turbo/run.py Outdated

Comment thread examples/hello-concurrency/run.py

fix: lint and format refactoring

ffe15ec

ask-bonk Bot reviewed Jun 12, 2026

View reviewed changes

Comment thread examples/blur/run.py Outdated

Comment thread tools/test-harness/manifest.yaml

fix: code review and update cargo deny

f5515f6

ask-bonk Bot reviewed Jun 12, 2026

View reviewed changes

Comment thread tools/test-harness/manifest.yaml

feat: add secret support to coglet

a190ffa

ask-bonk Bot reviewed Jun 15, 2026

View reviewed changes

Comment thread crates/coglet-python/src/input.rs Outdated

Comment thread tools/test-harness/manifest.yaml

anish-sahoo marked this pull request as ready for review June 15, 2026 18:13

anish-sahoo requested a review from a team as a code owner June 15, 2026 18:13

feat: move resnet to experimental

2112b99

markphelps reviewed Jun 15, 2026

View reviewed changes

markphelps requested changes Jun 15, 2026

View reviewed changes

anish-sahoo mentioned this pull request Jun 16, 2026

feat: add cog.Secret input support to coglet #3057

Merged

markphelps reviewed Jun 16, 2026

View reviewed changes

Comment thread tools/test-harness/manifest.yaml Outdated

markphelps approved these changes Jun 16, 2026

View reviewed changes

anish-sahoo added this pull request to the merge queue Jun 16, 2026

Merged via the queue into main with commit caf9ea7 Jun 16, 2026
44 of 45 checks passed

anish-sahoo deleted the update-examples branch June 16, 2026 18:16

Uh oh!

Conversation

anish-sahoo commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

coglet

test harness

Uh oh!

Uh oh!

Uh oh!

ask-bonk Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ask-bonk Bot commented Jun 11, 2026

Uh oh!

ask-bonk Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

ask-bonk Bot commented Jun 11, 2026

Uh oh!

ask-bonk Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ask-bonk Bot commented Jun 12, 2026

Uh oh!

ask-bonk Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ask-bonk Bot commented Jun 12, 2026

Uh oh!

ask-bonk Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

ask-bonk Bot commented Jun 15, 2026

Uh oh!

ask-bonk Bot commented Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

markphelps left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

markphelps left a comment

Choose a reason for hiding this comment

Uh oh!

anish-sahoo commented Jun 16, 2026

Uh oh!

ask-bonk Bot commented Jun 16, 2026

Uh oh!

markphelps left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

markphelps left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

anish-sahoo commented Jun 11, 2026 •

edited

Loading

ask-bonk Bot commented Jun 15, 2026 •

edited

Loading