cog

New features

• GPU compatibility warnings in cog doctor. cog doctor now warns when the resolved PyTorch and CUDA wheel ships no kernels for the local GPU, catching images that build successfully but fail when they run CUDA operations. (#3104)
• OpenTelemetry tracing. Models can opt into distributed traces spanning Cog's HTTP server, parent and worker processes, and Python model code. Cog continues incoming W3C trace context, exports through OTLP, and leaves tracing disabled by default. (#3157)
• JSON output for cog push. cog push <target> --json emits one versioned JSON document with digest-pinned image, model, and managed-weight references for image and bundle projects. Progress and diagnostics remain on stderr. (#3178)

Improvements

• .dockerignore guidance. The getting-started documentation now explains Cog's Docker build context and how to exclude development-only files to reduce build time and image size. (#3165)

Bug fixes

• CUDA 13 builds work with uv-managed Python. Cog now passes --break-system-packages and creates the expected python alias for both prebuilt and direct CUDA 13 base images. (#3160)

Breaking changes

• Local servers now bind to 127.0.0.1 instead of 0.0.0.0. cog serve, cog run -p, cog predict, and cog train no longer expose the prediction endpoint to your whole network by default. To get the old behavior back, use cog serve --host 0.0.0.0 or cog run -p 0.0.0.0:8888 — -p now accepts host:port. (#2812)

New features

• Browser playground. cog playground starts a local, schema-driven web UI for talking to a running model — a Postman-like tool that reads the model's OpenAPI schema and builds the input form for you, with a request inspector and webhook relay. Point it at anything speaking the Cog HTTP API with --target. (#3123)

• cog serve --playground. Serve the model and the playground in one command. The playground gets its own port (--playground-port, default 9000, 0 picks a free one) and both URLs are printed on startup. (#3154)

• @cog.concurrent decorator. Configure async runner concurrency on the method that does the work instead of in cog.yaml:

  class Runner(cog.BaseRunner):
      @cog.concurrent(max=4)
      async def run(self, prompt: str = cog.Input()) -> str:
          return await generate(prompt)
  

  The value is extracted statically at build time. concurrency.max in cog.yaml still works and still takes precedence, but it now warns on build — remove it when you migrate. (#3089)

• Input validation before build and run. cog predict, cog run, and cog train now validate -i inputs against the model's OpenAPI schema before building the image or starting the container, so a typo fails in a second rather than after a full build. Covers unknown input names, missing required inputs, type mismatches, enum values, numeric bounds, string length, and regex patterns. The schema is generated statically from cog.yaml; for existing images it's read from the openapi_schema label. (#3097)

• accept on Path and File inputs. Declare the MIME types or extensions an input takes — Input(accept="image/*"), Input(accept="audio/wav,audio/mp3"), Input(accept=".safetensors,.bin") — and Cog emits it as x-cog-accept in the OpenAPI schema so UIs and clients know what to offer. Using accept on a non-file type is a build error. (#2863)

• Local wheels and source archives in requirements. build.python_requirements now accepts bare local paths like ./dist/mylib-0.1.0-py3-none-any.whl or ./packages/localpkg.tar.gz. Paths resolve relative to the requirements file, and only the referenced artifacts get staged into the build. .whl, .zip, .tar.gz, .tgz, .tar.bz2, and .tar.xz are supported. (#3095)

Improvements

• Readable logs from local commands. cog serve, cog run, cog predict, and cog train now run the model with console-formatted logs instead of production JSON. Deployed models are unaffected. (#2806)
• Download progress for cog weights. cog weights import and cog weights pull show live Docker-style progress while they populate the local weight store. (#3087)
• Prebuilt base images are documented. New base images page covers what's on Cog's prebuilt base images, how --use-cog-base-image works, and how it interacts with --use-cuda-base-image. (#3088)

Bug fixes

• GPU builds on CUDA base images work again. Builds on nvidia/cuda images failed with externally-managed-environment because uv-managed Python is marked PEP 668 externally managed and uv 0.9.x enforces it. Cog now passes --break-system-packages wherever it installs into a uv-managed Python, not just in cog base-image. (#3073)
• Explicit cudnn with a CUDA minor version now validates. A config pairing cuda: "11.6" with cudnn: "8" was rejected up front even though 11.6 resolves fine to a patch-level base image like nvidia/cuda:11.6.2-cudnn8-devel-ubuntu20.04. Validation now matches versions the same way image selection do

New features

• Server-Sent Event prediction streams. HTTP prediction requests can now ask for Accept: text/event-stream to receive start, output, log, metric, and terminal completed events for predictors that explicitly opt in with @streaming / @cog.streaming. Reconnecting clients can replay retained in-flight prediction events with PUT /predictions/{id}. (#3019)
• JSON-native union inputs. Cog now supports union-typed predictor inputs in generated schemas and HTTP request handling, allowing compatible JSON-native values to be resolved correctly. (#3048)

Improvements

• Example models now live in the Cog repository. The example models from replicate/cog-examples have moved into this repository under examples/, with updated run.py / BaseRunner examples for common workflows including images, training, notebooks, streaming, concurrency, context, and Replicate API usage. (#3055)
• Experimental warning for cog weights. Every cog weights subcommand now prints a warning that the weights workflow is experimental and should not be relied on in production workflows yet. (#3025)
• Static schema parser target resolution. The static Python schema parser now uses run() as the primary prediction entry point while preserving legacy predict() fallback behavior, and resolves inherited and imported targets more consistently. (#3027)

Bug fixes

• cog.Secret inputs now work with coglet-backed predictions. Predictors that annotate inputs as Secret, Optional[Secret], or Secret | None now receive cog.types.Secret values at runtime again, restoring behavior that regressed in the Rust/coglet runtime rewrite. (#3057)
• cog doctor --fix now shows available remediation text. Findings without an auto-fix now display their remediation message instead of incorrectly saying no auto-fix is available. (#3031)
• Cog-managed weight uploads now send the correct layer media type. Weight layer uploads now propagate the Cog weight media type during registry finalization while preserving regular image-layer behavior. (#3033)
• Predictor validation now handles PEP 563 string annotations. Predictor files using from __future__ import annotations no longer reject valid setup() -> None methods or accept invalid run() -> None methods because None annotations were stored as strings. (#3034)
• cog serve now mounts weights like cog run. Models served locally now receive configured weights consistently with local prediction runs. (#3044)
• Omitted optional inputs now resolve to None at the HTTP edge. Optional predictor inputs omitted from HTTP requests are now passed as None instead of being treated inconsistently by request handling. (#3051)

Changelog

• d6c2b7064006b93cf4e608fd7157950bd5296417 Bump version to 0.21.0-rc.3 (#3050)
• 63c65dfcbe0f2507ded4f1518d41be9480f323fb feat: support JSON-native union inputs (#3048)