Release notes for the Weights & Biases Python SDK, including features, fixes, and breaking changes.
This page includes release notes for the Weights & Biases Python SDK (wandb package). For the W&B Weave Python SDK (weave package), see Weave SDK releases. For W&B Server, see W&B Server release notes.
November 20, 2026
Added
The backup_sync_dir_fn setting takes a function that run.finish() calls with the run’s sync directory, for example to back up the data of an offline run.
The automations API supports sandbox events. OnSandboxState fires when a W&B sandbox in a project or team is created, completes, or fails, and you filter it with SandboxEvent.state.
W&B LEET TUI features:
wandb leet <project-url> browses the runs of a W&B project.
Online runs, opened with wandb leet <run-url> or from an online project, show their system metrics like local runs do.
Online runs that are still running update their history and system metrics every 15 seconds.
Press Esc on a run opened with wandb leet <run-url> to go to the run’s project, where you can browse its other runs, like you can for local runs. Previously, an online run was limited to the single-run view.
Press shift+↑ and shift+↓ to scroll the rows of a heatmap system chart that has more series than fit, so you can inspect every CPU core or GPU.
Google TPU:
System metrics and hardware metadata are collected on TPU 8t and TPU 8i hosts.
Run metadata on multi-host or multislice GKE TPU pods includes the host’s worker index and count within its slice, and the slice index and count (tpu.worker_id, tpu.worker_count, tpu.slice_id, tpu.num_slices).
Changed
Identity token logins stay signed in until the JWT issuer’s session timeout without presenting the identity token again, when the issuer is configured to allow it.
On SIGTERM, W&B’s background service takes up to 30 seconds to finish active runs before exiting instead of stopping immediately.
Artifact.link() and Run.link_artifact() warn when the target collection name contains :, for example my-coll:prod. Pass aliases with the aliases argument instead.
Fixed
Scripts that seed Python’s random module get the same sequence of random numbers from run to run again after wandb.init(). This was a regression in v0.30.0.
System metrics are collected when TMPDIR is longer than about 60 characters, missing, not writable, or on a filesystem without Unix socket support. Since v0.31.0, a long TMPDIR, which is common on HPC clusters, left runs without any system metrics.
wandb sync of an offline run uploads the run’s artifacts, such as tables and models, to the entity and project the run is synced to, including with --entity and --project. This was a regression in v0.28.1.
File.download() no longer fails with invalid URL escape when a file name contains a %.
wandb.watch histograms of parameters and gradients whose values are all identical show those values. Previously, the histogram was empty.
Logging an artifact with thousands of files sends its file list to the server in smaller batches, which makes request timeouts less likely.
Api.automations() no longer fails when an entity or project has a push notification automation, and skips automations it cannot parse instead of failing.
Failed requests to the W&B server report the HTTP status, such as 401 "Unauthorized" for an invalid API key. Previously, the error message was the literal text <no message>.
Launch agents in secure mode reject jobs with SageMaker or Vertex resource arguments.
Installing wandb from source in Termux on Android builds for Android. Previously, the build always failed.
W&B LEET TUI:
An online run opened before it logged any history now updates its state, summary, and system metrics while it is running.
wandb leet symon labels the y-axes of the GPU PCIe and NVLink throughput charts and the Trainium memory charts correctly, for example 7.67GB/s instead of 7.67e+09GB/s.
Security
The wandb/launch-agent Docker image is built on an updated base image with a patched Brotli (1.2.1) to address CVE-2026-111497, and with Go 1.27.4.
October 16, 2026
Notable changes
The default API endpoint is now CoreWeave Forge at https://api.forge.coreweave.com, and run links point to https://forge.coreweave.com/wandb. Existing API keys stored for api.wandb.ai keep working without logging in again.
Added
Python 3.15 is now supported.
wandb.init(resume="must") now works for offline runs. wandb sync fails if there is no run to resume.
wandb sweep-scheduler [entity/project/]sweep_id runs a sweep’s search locally with Optuna or Ax as the optimizer, configured in a scheduler block of the sweep config. wandb sweep prints the command for sweeps that set one.
wandb agent --max-consecutive-failed-runs N shuts the agent down once N runs in a row have failed.
wandb sync warns when a run was logged with a newer version of wandb than the one syncing it.
W&B LEET TUI features:
wandb leet inspect --summary prints a run’s state, latest metric values, config, and last console lines from its local .wandb file, which is a quick way for a script or a coding agent to check on a run. --json prints the records, or the summary, as JSON, and --follow keeps printing records as a running run writes them.
Press l to link the console logs to the charts in the single-run view. With the logs pane focused, the charts show a crosshair at the time the selected line was logged, and inspecting a chart with the mouse selects the console line logged nearest to that point. Press l again to unlink.
Press Ctrl+A to select every run that matches the runs filter, and press x to deselect all runs except the pinned one.
Selected and pinned runs are remembered for each wandb directory and selected again the next time you open it. The newest run is also selected if it started since the last session.
The CPU Core (%) chart opens as a heatmap with one row per core. Press y to switch it back to lines.
wandb leet symon, the live system monitor, has several new features:
It shows the utilization of every CPU core, the CPU temperature, and swap usage on Linux, and the hostname, uptime, and load average in its header.
It charts network and disk throughput in bytes per second instead of totals, and opens with the CPU, memory, network, and disk charts ahead of the rest.
It has a sidebar with the machine’s chip, cores, memory, and GPUs, live meters for CPU, memory, swap, disk, and GPU use, and the processes that use the most CPU. Press t to sort the processes by memory, f to filter them by name or PID, Tab to move between the process list and the charts, and [ to hide the sidebar. Drag the sidebar border to resize it, and press 0 to restore the default width.
New system metrics:
PCIe throughput (gpu.N.pcieTxBytes, gpu.N.pcieRxBytes) for NVIDIA GPUs older than Hopper.
Each GPU’s cumulative energy consumption in joules as gpu.N.energyJoules, for NVIDIA and AMD GPUs.
proc.cpu.throttledPercent, the percentage of CPU scheduler periods in which the container’s CPU limit throttled the run.
proc.memory.oomKills, the number of processes in the run’s container that the out-of-memory killer has killed since the run started.
network.tcpRetransmits, the number of TCP segments retransmitted since the run started.
The CPU and memory that W&B’s own background processes use, as wandb.cpu and wandb.memory.rssMB. These metrics are off by default. To turn them on, use the x_stats_self_usage setting.
Run metadata now records the CPU model and, for each NVIDIA GPU, its PCI bus ID and, on Linux, its NUMA node.
Changed
The default API endpoint is now CoreWeave Forge at https://api.forge.coreweave.com. Run, project, sweep, and login links point to https://forge.coreweave.com/wandb, including when base_url is set to https://api.wandb.ai or https://forge.coreweave.com/api/wandb. Existing API keys stored for api.wandb.ai keep working without logging in again.
The helper process that collects system metrics is now wandb-symon instead of wandb-xpu.
import wandb and the wandb CLI start about twice as fast.
The wandb package is smaller. wandb-core shrank by about a third, from 52 MB to 35 MB on Linux x86_64, and the parquet reader and the system metrics helper shrank by more than half.
Runs write their data to the local .wandb file at least every 15 seconds, so wandb leet shows recent data sooner for runs that log infrequently.
The cpu system metric reports the monitored process’s CPU utilization over the sampling interval instead of its average since the process started.
NVIDIA GPM system metrics such as gpu.N.smActive and gpu.N.dramActive are now averaged over the whole interval between samples instead of a 200 ms window.
AMD GPU system metrics are read through the ROCm SMI library (librocm_smi64.so) instead of by running the rocm-smi tool on every sample. Readings that a GPU doesn’t provide are left out instead of reported as 0.
OpenMetrics and DCGM exporter endpoints are scraped once per sampling interval for all runs on the machine instead of once per run, and all runs use the same openmetrics.<endpoint>.<metric>.<index> index for the same metric labels.
Removed
Removed the undocumented wandb.set_trace(). Use Python’s built-in breakpoint() instead.
Fixed
Resumed runs no longer report a huge or negative _runtime and run duration. This was a regression in v0.30.0.
Offline runs show warnings again, such as when run.log() data is ignored because its step is lower than the current step. This was a regression in v0.27.0.
Per-process GPU metrics (gpu.process.*) are logged again for NVIDIA GPUs that the process that called wandb.init() uses. Since v0.18.2, they were logged only when a subprocess used the GPU.
Runs that print thousands of console lines per second no longer spend minutes of CPU time uploading them and finishing.
Runs that finish right after they initialize, for example when wandb sync uploads a short run, no longer risk exiting before their final data and exit status are uploaded.
wandb sync no longer overwrites the earlier history of a resumed run when the server reports a stale step.
Calling wandb.init() after wandb.agent(sweep_id, function=...) returns creates a new run instead of reusing and overwriting the last sweep run.
wandb.Api(overrides={"base_url": ...}) no longer changes the W&B server URL that wandb.init() and other W&B calls use later in the same process.
Every wandb.Api() object no longer permanently retains a few MiB of memory after it is garbage collected.
Run.scan_history(keys=...) no longer fails with 403 Forbidden on W&B deployments that store run history in Amazon S3.
Passing a run’s config to wandb.init(config=...) or run.config.update() copies its keys. Previously, each pass nested the config one level deeper under an _items key.
wandb verify no longer hangs on its final sweep check against W&B Server instances that deduplicate sweep run configs.
linked_at is accepted as a filter field in api.registries().collections().versions(filter={"linked_at": ...}), matching the existing linked_at order field.
The Docker image ID that W&B records for runs inside Kubernetes no longer loses the start of the image name and the end of the digest.
Integrations:
The CatBoost WandbCallback logs every evaluation, including the last one, at the iteration CatBoost makes it. With metric_period > 1, it used to log each value one evaluation late and drop the final ones.
wandb.tensorflow.WandbHook works again with TensorFlow 2.16 and newer.
wandb.integration.ultralytics imports again with ultralytics 8.3.200 and newer.
wandb.integration.huggingface.autolog() no longer requires the pytz package.
Calling a pipeline that wandb.integration.diffusers.autolog() doesn’t support no longer raises an error. The call is not logged.
Importing wandb.integration.lightning.fabric or wandb.integration.ultralytics without the libraries they need fails with an ImportError that names the missing package.
System metrics:
The x_stats_gpu_device_ids setting limits which GPUs are reported for every vendor, including AMD and Apple GPUs, as documented. It applied only to NVIDIA GPUs.
GPU, TPU, and Apple accelerator metrics resume when the helper process that collects them exits or fails to start. They used to stop for the rest of the run.
NVIDIA PCIe and NVLink throughput metrics from GPM (gpu.N.pcieTxBytes, gpu.N.pcieRxBytes, gpu.N.nvlinkTxBytes, gpu.N.nvlinkRxBytes) are reported in bytes per second, as their names say. They were in mebibytes per second.
network.sent and network.recv no longer count loopback traffic, or count traffic twice through bonded and bridged interfaces.
NVIDIA GPM metrics such as gpu.N.smActive are averaged over each run’s own sampling interval. When several runs shared one process, each run’s value covered only the time since any run’s previous sample.
Sampling system metrics no longer costs more as the number of runs in a process grows.
Run metadata on AWS Trainium records the Neuron device and NeuronCore counts. They were always 0.
W&B LEET TUI:
The system metrics filter applies to whichever run is highlighted, not only the run it was typed on. Space and Enter work in the runs list right after you filter runs with f from another pane.
Changing the system metrics grid rows or columns, including in wandb leet symon, no longer crashes and takes effect immediately.
wandb leet inspect exits with an error when the file is not a readable .wandb log, instead of printing “skipped corrupt data” forever when its output is piped.
Security
The wandb/launch-agent Docker image is built on an updated base image that fixes the CVEs reported against the 0.30.0 image.
September 8, 2026
Notable changes
wandb.InternalApi, the wandb.PublicApi and wandb.apis.PublicApi aliases, and the wandb.apis.internal and wandb.sdk.internal.internal_api modules are removed. Use wandb.Api() instead.
wandb.api and wandb.ensure_configured() are deprecated. Use wandb.login(prompt=False) to check whether a user is logged in, and wandb.Api() for everything else.
The legacy wandb sync options are removed. Run wandb sync with no arguments instead of wandb sync --sync-all, and use --job-type instead of --job_type. To list the current options, run wandb sync --help.
Added
wandb.login(prompt=False) uses the configured credentials if there are any and returns False otherwise, without asking for an API key or switching to offline mode. Use it to check whether a user is logged in without blocking.
The Automations API supports sending a prompt to ARIA as an automation action with SendPromptToAria. For more information, see Create an ARIA automation.
W&B LEET TUI features:
Charts support custom x-axes that you set with run.define_metric(), including glob definitions such as run.define_metric("train/*", step_metric="train/step"). The chart header shows the axis name as [x: train/step]. This applies to runs that you open from local .wandb files.
To filter console logs, focus the logs pane, press /, then type a regular expression, or press Tab and type a glob pattern. Only matching lines are shown, and new matches are followed as they arrive. Press Ctrl+/ to clear the filter.
The metrics, system metrics, and runs filters are saved for each wandb directory and applied again the next time you open it, in both the workspace and the single-run view. The filters are stored in .wandb-leet.json inside the directory. Clearing a filter removes it from the saved filters.
The border or separator line that you drag to resize a pane or sidebar is highlighted while you drag it.
Changed
wandb status --settings no longer prints the API key, and wandb projects no longer prints project descriptions.
System metrics on Apple Silicon Macs start about 1.5 seconds sooner after a run starts.
The W&B LEET TUI is faster on long runs. A run with 100 metrics and 400,000 steps opens 10 times faster, frames render about 30 percent faster, and live chart updates no longer redraw every point.
Deprecated
wandb.api and wandb.ensure_configured() are deprecated and will be removed in a later release. wandb.api now provides only api_key, default_entity, and viewer(). Use wandb.login(prompt=False) to check whether a user is logged in, and wandb.Api() for everything else.
wandb.sandbox and the wandb beta sandbox commands are deprecated and will be removed in a later release. Use the cwsandbox package directly.
Removed
Removed wandb.InternalApi, the wandb.PublicApi and wandb.apis.PublicApi aliases, and the wandb.apis.internal and wandb.sdk.internal.internal_api modules. Use wandb.Api() instead.
Removed the legacy wandb sync options. Run wandb sync with no arguments instead of wandb sync --sync-all, and use --job-type instead of --job_type.
Fixed
File uploads and downloads that take longer than 20 seconds no longer fail with CommError: Failed to execute API request: the service process is busy and did not respond in time. This was a regression in v0.29.0.
wandb.init() and wandb.Api() no longer stall for up to 10 seconds when the W&B server is slow to respond. This was a regression in v0.29.0.
Requests that are rejected with HTTP 402 Payment Required, which indicates a plan or quota limit such as an artifact usage limit, are no longer retried. The error is reported immediately.
Syncing a run whose artifacts were already uploaded no longer fails when the artifact files are no longer on disk. This was a regression in v0.29.0.
A failed artifact upload no longer deletes the files it was uploading, so wandb sync can retry it. After a successful upload, only files in the artifact’s staging directory are cleaned up.
When a .wandb file ends with an incomplete record, which happens while a run is still writing or after a run is stopped abruptly mid-write, wandb sync now explains what to do instead of reporting an internal error.
Passing aliases or tags to Run.log_artifact() in the public API no longer fails with a server error.
wandb.Image masks that you provide as float arrays with values outside the range [0, 255] now raise TypeError, as integer arrays already did, instead of being saved with incorrect class IDs.
System metrics are collected as soon as a run starts instead of one sampling interval later, which is 15 seconds by default. Short runs now record system metrics, and system charts no longer start with a gap.
wandb.init() no longer waits up to 5 seconds for the GPU and TPU metrics collector to start.
In Jupyter notebooks, system metrics are collected while a cell runs and paused between cells. Since v0.19.10 the two were swapped, so system metrics were collected only while the notebook was idle.
On machines with AMD GPUs, system metrics monitoring no longer runs rocm-smi up front, which took 1 to 3 seconds, and no longer crashes when rocm-smi reports no GPUs.
System metrics on Apple Silicon Macs:
CPU utilization, frequency, and power now cover the whole sampling interval instead of a snapshot of about 10 milliseconds. The snapshot made utilization and frequency noisy and overstated power by about 25 percent.
Macs with 256 GB or more of RAM no longer report a memory size of 0 GB.
CPU and GPU temperatures are no longer skewed by invalid sensor readings.
M5 Pro and M5 Max Macs no longer report 0 efficiency cores.
CPU frequencies on the MacBook Neo are no longer reported about 1,000 times too high.
CPU utilization and frequency are now collected for the efficiency cores of M5 Macs and the performance cores of M5 Pro and M5 Max Macs, which were missing.
W&B LEET TUI:
wandb leet prints an error when it cannot start, for example when no terminal is available, instead of exiting with status 1 and no output. Debug logs (WANDB_DEBUG=true) are written next to the LEET config file instead of to the current directory.
wandb leet no longer sends usage telemetry in offline or disabled mode (WANDB_MODE=offline or WANDB_MODE=disabled), like the rest of the SDK.
In narrow terminals, the workspace hides the run overview sidebar, and then the runs list, when they would leave the charts fewer than 24 columns wide. Previously, an 80-column terminal showed the charts as a one-column sliver between the two sidebars.
The top y-axis label of a chart is no longer cut off when it is wider than the other labels. For example, the label now shows 1.09e+03 instead of +03.
August 26, 2026
Notable changes
On macOS, wandb now requires macOS 13 (Ventura) or newer.
wandb.Image no longer normalizes pixel values. Make sure your data already falls within the range [0, 255].
wandb.Video() now requires the format argument when you initialize it from a NumPy array or an IO object.
The .length property on paginated objects, such as Api.runs(), Api.sweeps(), and run.files(), is removed. Use len(...) instead.
The following deprecated APIs are removed: the start_method setting, the summary="best" option and goal argument of run.define_metric(), the quiet argument of run.finish() and wandb.finish(), and the run.project_name(), run.get_url(), run.get_project_url(), and run.get_sweep_url() methods.
Added
The new public API method run.console_logs() returns the console output that W&B captured for a run, for both finished and active runs. To return only the last N lines, pass last=N. Reading from the beginning of the log requires W&B Server v0.77 or newer.
The Automations API supports team and organization scopes.
The Automations API supports creating and editing automations that are scoped to a Registry object.
wandb.Artifact accepts a new digest_algorithm argument, either "MD5" (the default) or "XXH128". Pass digest_algorithm="XXH128" to opt an artifact into XXH128 hashing where possible. For artifacts created with digest_algorithm="XXH128", artifact.verify() fails on SDK versions older than v0.29.0.
The new property Artifact.linked_at returns the time a version was linked to the relevant collection. It applies only to linked versions and returns None for source artifacts.
Registry version queries (Api.registries().versions(), Api.registries().collections().versions(), Registry.versions(), and Registry.collections().versions()) accept an optional order string as a keyword argument for organizations with advanced search. You can order versions by created_at, artifact_size, or linked_at.
W&B LEET TUI features:
The new command wandb leet inspect [PATH] browses the raw records in a run’s .wandb transaction log, showing a filterable list of records alongside a text view of the selected record, and following live runs as they write. Press i in the single-run view to inspect the current run. Pipe or redirect the output to print the records instead, such as wandb leet inspect | less.
Press g to draw guides behind line charts, either a dotted background or horizontal lines aligned with the axis ticks. Your choice is saved, and you can also set it with chart_guides in wandb leet config.
Drag the separator lines between the run overview sections to resize them with the mouse. The proportions are saved for each view. Press 0 to reset them along with the other pane sizes.
Changed
The message format that the SDK uses to communicate between internal processes has changed slightly. If you use wandb beta core, restart the service after you upgrade wandb, because some operations can fail when the SDK and service versions differ.
wandb.sandbox now allows GPU resource requests for sandboxes instead of rejecting resources.gpu on the client.
Registry search methods Api.registries(), .collections(), and .versions() now validate filter field names and reject unsupported names.
Api.create_automation() and Api.update_automation() now raise UnsupportedError instead of CommError when the server does not support the given automation.
The macOS wheels now require macOS 13 (Ventura) or newer. macOS 12 reached end of life in 2024.
In the W&B LEET TUI, the run overview sections (Environment, Config, and Summary) now share the sidebar height proportionally and use all available space. Previously, they stopped growing at fixed sizes, which left the bottom of tall terminals empty.
Removed
Removed wandb.tensorboard.log() and wandb.tensorflow.log().
Removed the deprecated start_method setting, which already had no effect.
Removed the deprecated .length property from wandb.apis.paginator.SizedPaginator. Use len(...) instead.
Removed the deprecated summary="best" option and goal argument of run.define_metric(). Use summary="min" or summary="max" instead.
Removed the deprecated quiet argument of run.finish() and wandb.finish(). Use wandb.Settings(quiet=...) instead.
Removed the deprecated run.project_name(), run.get_url(), run.get_project_url(), and run.get_sweep_url() methods. Use the run.project, run.url, run.project_url, and run.sweep_url properties instead.
Removed the deprecated normalize argument from wandb.Image. Make sure pixel values in NumPy arrays and PyTorch tensors already fall within the range [0, 255].
Fixed
Artifact.new_file() now works for artifacts uploaded with wandb sync.
Run.scan_history(keys=...) no longer fails for runs with nested values.
At certain pane widths, such as while dragging a sidebar edge, the W&B LEET TUI drew metric charts wider than the available space, which pushed the right sidebar off screen.
W&B LEET TUI:
wandb leet config now works in short terminal windows. The field list scrolls with the selection, and Space toggles boolean settings.
Media panels no longer paint black bars around images in the ANSI view or flash black when you switch to the full-resolution view. The area around the image now shows the terminal background, and switching modes no longer flashes or moves the image.
Security
Downloading artifacts with artifact.files() or artifact.checkout() now rejects file and artifact names that contain relative paths. This fixes a path traversal vulnerability described in CVE-2026-91771.
August 12, 2026
Notable changes
The wandb sync --clean command now exits with code 1 and prints a hint to use the new wandb clean command instead.
wandb login now verifies credentials by default. To turn off verification, pass --no-verify on the command line or verify=False to wandb.login().
Added
Filestream requests support gzip compression, which reduces network traffic when you log metrics. Compression is opt-in and requires server support. To turn it on, set x_file_stream_no_gzip=False in wandb.Settings. Compression becomes the default in a future release.
The wandb agent command accepts a new --term-timeout flag.
The new property run.sync_dir returns the run’s sync directory.
Changed
The new wandb clean command replaces wandb sync --clean.
wandb.init() now reports the error it was retrying, such as a network error, when it times out, instead of a generic timeout message. The init_timeout setting now also bounds the backend’s retries during run initialization.
wandb.Api().create_run_queue(), wandb.Api().create_custom_chart(), and wandb.Api().upsert_run_queue() now raise WandbApiFailedError when the operation fails on the backend.
Paginated artifact and registry query methods now validate pagination arguments on the client before they fetch any results. This applies to Api.artifacts(), Api.artifact_collections(), Api.registries(), Registries.collections(), Registries.versions(), Collections.versions(), Registry.collections(), Registry.versions(), and Project.collections().
Ordered registry search now scopes per-registry collection and version queries to the current registry’s internal ID, which makes pagination and filtering more reliable.
Response parsing is faster for many wandb.Api operations, including artifact and registry queries.
Argument handling in wandb launch for the local-process resource is hardened so that job-supplied values are always shell-quoted.
The launch agent now restricts a job’s Git source URL to HTTPS and SSH remotes, and pins Git’s protocol allowlist when it fetches the source and updates submodules.
Deprecated
wandb sync --sync-tensorboard is deprecated and will be removed in a later release.
Removed
Releases no longer include 32-bit Windows (win32) wheels. Use 64-bit Python on Windows.
Fixed
Registry search with registries(order=...).collections(order=...).versions() now returns artifact versions in registry order, collection order, or both.
macOS x86_64 wheels now contain x86_64 builds of the wandb-xpu binary and the Rust parquet library. Previously, these shipped as arm64 builds, which could not run or be loaded on Intel Macs.
wandb.login(verify=True) and wandb login --verify now verify federated identity (identity token) credentials, which were previously not verified.
wandb login now validates API keys before it saves them to the .netrc file.
wandb login no longer removes or corrupts credentials that belong to other machines in your .netrc file.
wandb login and wandb verify no longer update the system host settings when a login attempt fails.
wandb verify now reports a failed check instead of crashing when an operation still fails after retries.
Fixed wandb.Api(overrides={"base_url": ...}) failing to authenticate with federated identity (identity token) credentials when the specified server was not the default one, such as a Dedicated Cloud deployment, unless WANDB_BASE_URL was also set.
When you use federated identity, the identity token file is now read again for each access token exchange instead of only at startup. Short-lived identity tokens that are re-minted to the same path now keep working for runs that outlive them.
When you use federated identity and the server rejects the identity token exchange, such as for an invalid or expired token, requests now fail immediately with the server’s error message. Previously, the rejected exchange was retried until requests timed out with a generic error.
Sweep agents now respect a custom WANDB_DIR when they write JSON output.
The global_step metric that is created when you sync TensorBoard files is no longer prefixed, such as train/global_step, which makes it easier to compare training and validation metrics.
The TensorBoard integration now produces fewer W&B steps by merging data for the same global_step into one W&B step where possible.
wandb sync no longer hangs on a run that set its name, tags, or notes after starting.
Log messages captured from Python loggers no longer add blank lines to a run’s logs.
del run.summary[key] now removes metrics that were logged as nested values, which it previously ignored.
Security
Debug logs no longer contain CoreWeave cluster credentials.
wandb docker and wandb docker-run no longer pass your API key on the command line, where other users of the same machine could read it.
July 16, 2026
Notable changes
This version requires protobuf v5+ and drops support for protobuf v4.
This version drops compatibility with W&B Server v0.69.x and older.
Added
Api().project().sweeps() accepts an optional filters argument, which uses the same filter format as Api().runs().
The new public API method Run.stop() (wandb.Api().run(...).stop()) requests that an active run stop gracefully, equivalent to the Stop run button in the W&B App.
The new method wandb.Api().organization() returns an Organization by name. If you omit the name, it returns your current default organization.
Changed
wandb sync now routes to wandb beta sync for supported sets of parameters. To use the previous implementation instead, pass --legacy.
wandb beta sync now skips online runs by default, matching wandb sync. To include online runs, pass --no-skip-online.
wandb.sandbox now defaults serverless sandboxes to a maximum lifetime of 12 hours (max_lifetime_seconds=43200). To override the default, pass max_lifetime_seconds for an individual sandbox or set it in SandboxDefaults.
wandb.Api().runs() and wandb.Api().run() now raise RunNotFoundError when a run’s data fails to load, such as when the run is deleted before its data finishes loading.
run.save() accepts a new glob argument. Pass glob=False to turn off glob expansion (*, ?, and [...]) and treat the path as a literal path. This is useful when a filename contains glob metacharacters.
Temporary Unix socket files and directories are now removed at shutdown.
protobuf v4 is no longer supported.
Removed
Removed legacy fallback implementations for downloading artifact files from W&B Server releases older than v0.70.0. On these servers, Artifact.files() and Artifact.download() no longer work for any artifact, and Artifact.get_entry() and Artifact.get_path() file downloads no longer work for non-reference artifacts. To keep using these operations, upgrade W&B Server to v0.70.0 or newer.
Removed legacy fallback implementations for fetching an artifact by name from W&B Server releases older than v0.70.0. On these servers, wandb.Api().artifact() and other methods that fetch artifacts by path no longer work. To keep using these operations, upgrade W&B Server to v0.70.0 or newer.
Removed the GitPython dependency. Git metadata is now collected by invoking the git executable directly. The GIT_PYTHON_GIT_EXECUTABLE environment variable is still honored for locating the executable.
Fixed
Saving a linked registry artifact, such as when adding an alias, no longer fails when you lack write access to the source project.
np.float16 and np.float32NaN values logged with Run.log() are now recorded as NaN instead of being silently dropped, matching the behavior of np.float64 and native float.
Run.upload_file() (through wandb.Api().run(...)) now registers the uploaded file with the run on Self-Managed deployments. Previously, the file’s bytes were uploaded but never committed, so the file didn’t appear on the run in deployments without object-store notifications.
Sweep agents now let an in-progress run finish before exiting when the sweep is deleted or Api() returns a 404 error.
Fixed deadlocks introduced in SDK v0.25.1. These deadlocks were reported when using PyTorch with Python versions earlier than 3.14, which use the forkmultiprocessing start method by default. They could also occur during garbage collection when using some wandb.Api functionality.
June 22, 2026
Notable changes
This version drops compatibility with server versions older than 0.65.0.
Added
W&B LEET TUI features:
Media panels support high-resolution image rendering in terminals that support the Kitty protocol, with ANSI fallback.
Media panels support synced scrubbing: press l to link scrubbing, then navigate with arrow keys, Home, and End, which move a shared cursor over the union step timeline and every image tile follows it.
Basic remote-run support with wandb leet [run-url].
The following paginated artifacts and registry API methods now accept an optional order string as a keyword argument: Api.artifacts(), Api.artifact_collections(), Api.registries(), Api.registries().collections(), Registry.collections().
Changed
W&B LEET TUI changes:
The W&B LEET TUI is now generally available, and the command is updated from wandb beta leet to wandb leet. The prior command is still available as an alias.
Lists of images logged under a single key are now displayed in the W&B LEET media pane, one tile per image.
To improve query performance, wandb.Api().runs() no longer loads Sweeps for each run by default. Sweep data is loaded on first access of the sweep property.
Fixed
File.download() no longer fails after a hard-coded 5-second timeout; downloads respect the file transfer settings.
Fixed a regression introduced in 0.27.1 where wandb.Api().viewer, Api().user(), and Api().users() could fail with the error WandbApiFailedError: relogin required for some API keys.
When a wandb.Image is logged inside a wandb.Table with multiple box or mask layers with distinct class_labels, each layer’s labels are now preserved in new box_class_maps or mask_class_maps fields in the table.json. Previously, a single class_map would be incorrectly overwritten by each set of class labels.
Artifact file operations now consistently require normalized relative paths.
To fix silent failures when $TMPDIR is missing or unwritable, logging an artifact using a WandbLogger or run.log_artifact now writes the manifest file to the artifact’s staging directory instead of the OS temp dir.
Fixed a bug where logging artifacts in shared mode (wandb.init(mode="shared")) with code-saving enabled would raise an error.
To create a diff.patch file, the git_root setting is now used by default, falling back to the root_dir setting.
Apple system metrics like GPU, CPU, power, and temperature are now collected on Apple M5 Macs.
File download progress is now shown when exporting history using wandb.Api().run(...).download_history_exports.
Fixed a regression introduced in 0.27.1 where Run.scan_history() would return no rows for a run whose history exists but has not been exported to Parquet, like an active run.
June 6, 2026
Fixed
Run.scan_history() no longer returns duplicate or missing steps when fetching run history.
June 4, 2026
Added
The Automations API now supports artifact tag, collection tag, and artifact unlinking events.
Changed
wandb.Api GraphQL operations now route through wandb-core instead of the legacy Python GraphQL client.
Failures from wandb.Api GraphQL operations now raise WandbApiFailedError instead of requests HTTP exceptions.
Run.scan_history() now reads from exported parquet history when available, which can significantly improve throughput for runs with large history. This capability was introduced under beta_scan_history in SDK v0.23.1.
Removed
Removed the unsupported wandb.apis.importers API.
Removed stale OpenAI, Cohere, and LangChain LLM integrations, including legacy autologging and tracing APIs.
Removed the deprecated Keras WandbCallback and the legacy wandb.integration.yolov8 callback package.
Fixed
wandb.sandbox now rejects invalid argument values for placement, GPU, and egress.
May 13, 2026
Notable Changes
This version drops support for Python 3.9 and Pydantic v1. Pydantic v2.6 or newer is now required.
Added
The new package wandb.sandbox and the wandb beta sandbox CLI for using Serverless Sandboxes.
The new setting stop_on_fatal_error and environment variable WANDB_STOP_ON_FATAL_ERROR stop a run after a fatal error that prevents it from uploading metrics.
The new setting finish_timeout and corresponding environment variable WANDB_FINISH_TIMEOUT let you specify how many seconds to wait for data to upload at the end of a run. If a run times out, a warning message is printed. Use boolean setting finish_timeout_raises or corresponding environment variable WANDB_FINISH_TIMEOUT_RAISES to raise an error in addition to printing a warning message.
The new setting capture_loggers lets you specify Python loggers to capture into the run’s Logs tab, using a mapping of logger name to minimum log level. Log records emitted by those loggers are published as console output to the run, similar to stdout/stderr capture.
The new method run.write_logs() writes text directly to the run’s Logs tab instead of relying on automatic capture of stdout/stderr.
Changed
System metrics for runs in Linux containers now report CPU and memory usage relative to the container’s cgroup v2 limits instead of the host’s totals.
Python 3.9 is no longer supported.
Pydantic 2.6 or above is now required. Pydantic v1 is no longer supported.
April 23, 2026
Added
wandb.Api methods that return artifacts, registries, automations, and related paginators accept an optional start argument so you can resume iteration from a saved cursor.
The stop_fn setting on wandb.Settings customizes how the SDK asks the Python process to stop when a run stop is requested (the default sends SIGINT). Pass the setting with wandb.setup(settings=...) or wandb.init(settings=...). Because the value is a callable, it cannot be set with an environment variable.
Changed
In wandb beta leet, WASD keys and arrow keys are interchangeable for moving focus within each pane (chart grids, list navigation). Home, End, Page Up, and Page Down keys work consistently. The media pane still uses arrow keys to scrub the X-axis and WASD keys to move between image tiles.
Fixed
wandb.init(id=run_id, reinit="create_new") now raises an error if another run with the same run_id is still active in the same script.
wandb.Api no longer raises on some operations when offline mode is enabled with the WANDB_MODE environment variable or the mode setting.
April 13, 2026
Notable Changes
This version drops compatibility with Server versions older than 0.63.0 for Dedicated Cloud and Self-Managed W&B deployments.
Added
wandb beta core start|stop commands run a detached wandb-core service and reuse it across multiple processes via the WANDB_SERVICE environment variable.
The @wandb_log decorator adds support for Kubeflow Pipelines v2 (kfp>=2.0.0).
Set the property wandb.Settings(allow_media_symlink=true) to link to media files to the run directory rather than copying them. This improves performance and saves disk space. A symbolic link is used if the filesystem supports it. If the filesystem does not support symbolic links, such as in Windows, a hard link is used instead. The SDK does not check whether the media file was modified locally before uploading it.
Use method run.pin_config_keys(keys) to programmatically pin specific config keys for display in the References section on the Run Overview page.
Sweep.agent().Runs() and Sweep.agents().Runs() methods return details about the sweep’s runs running on a given agent or all agents.
On Google TPU infrastructure, the SDK attempts to collect new metrics such as tensorcore_util and latency distributions, in addition to duty_cycle_pct and hbm_capacity_total/hbm_capacity_usage. See TPU monitoring library: Supported metrics in the Google Cloud documentation.
On Nvidia Hopper+ GPUs (H100 and newer), NVML GPU Performance Monitoring profiling metrics are collected, providing visibility into SM utilization, tensor/FP pipeline activity, DRAM bandwidth, and PCIe/NVLink throughput without running the DCGM daemon.
wandb beta leet TUI updates:
Support for displaying tags and notes in the run overview.
Support for filtering runs by metadata in multi-run workspace mode using the f keyboard shortcut.
Support for displaying wandb.Image data as ANSI thumbnails, with grid layout, X-axis scrubbing, full-screen mode, and keyboard or pointer navigation.
Cycle through the following display modes in an individual chart using the y keyboard shortcut:
Press once to turn on log-scale display for the Y-axis.
For system metrics like GPU utilization expressed as percentages, press again to turn off log-scale display and turn on bucketed heatmap mode.
Press again to go back to the default display mode.
New standalone system monitor mode: wandb beta leet symon.
New colorblind-friendly color schemes:
dusk-shore: sienna/blue/gray gradient, suitable for per-plot single-run mode.
clear-signal: sienna/blue/gray theme with the colors interleaved for maximum differentiation between adjacent series.
Configure with wandb beta leet config.
Changed
JSON serialization and deserialization now use orjson for improved performance.
Improved observability of system metrics in wandb beta leet TUI with multi-series overlays, inspection, and live/history zoom.
Reduced the likelihood of run base color collisions in wandb beta leet. Now, a workspace-scoped run color allocator preserves hashed palette affinity, resolving collisions with nearby HSL variants.
Fixed
Fixed a bug where if no upstream tracking branch was set, the SDK would not create a git diff.patch file between the local branch and the nearest upstream branch. To opt in to this fix, pass property wandb.Settings(disable_git_fork_point=False). See Settings().
Fixed a bug where update_automation() could silently drop event filters, such as alias conditions on OnAddArtifactAlias, when a new event is provided.
Fixed a bug that could cause artifact client ID collisions in forked child processes by reseeding the fast ID generator after fork().
Fixed a bug where the environment variable WANDB__EXTRA_HTTP_HEADERS was not applied to presigned object-store upload and download requests.
Fixed a deadlock in artifact.download() for artifacts with many large files.
Fixed a bug where User.generate_api_key() could fail to fetch the list of existing API key names.
March 10, 2026
Added
New run console logs pane in wandb beta leet command. Toggle with l.
New system metrics pane in multi-run workspace mode in wandb beta leet command. Toggle with s.
wandb beta leet command now supports system metrics filtering. Toggle with \.
ArtifactType.collections() now supports filtering and ordering of collections.
A warning is now logged if run.log_artifact() does not create a new version because the artifact content is identical to an existing version.
The new method Project.collections() fetches filtered and ordered artifact collections in a project.
The new wandb purge-cache CLI command cleans up cached files.
Fixed
Fixed a rare deadlock caused when garbage collection triggers at an unlucky time and runs a __del__ method that prints.
Fixed a bug where loading multiple runs with api.Runs("..."), then attempting to get a single run’s user with code like the following could fail with an AttributeError:
for run in runs: print(run.user)
February 12, 2026
Notable changes
Python 3.8 is no longer supported.
Added
The wandb beta leet TUI now supports multi-run workspaces.
Configure wandb beta leet’s persisted configuration interactively with the wandb beta leet config command.
Access the project owner’s user information with the owner property of wandb.apis.public.Project.
Fixed
Fixed a bug where a sweep agent for a deleted sweep could run indefinitely with repeated 404 errors.
Fixed a bug where a wandb-core crash could produce extremely long and repetitive tracebacks in older Python versions.
Fixed a regression introduced in v0.24.0 where a TensorBoard sync could stop after syncing 1 MB of data.
February 4, 2026
Added
wandb.Api() now supports Federated Auth (JWT based authentication).
Fixed
The presigned download URL is now refreshed if it expires during artifact file downloads.
January 29, 2026
Notable Changes
Runs created with wandb==0.24.0 may fail to upload some data, which this release fixes. Missing data is stored in the run’s .wandb file and can be reuploaded with wandb sync.
Added
download_history_exports in api.Run class to download exported run history in parquet file format.
Changed
When a settings file (such as ./wandb/settings or ~/.config/wandb/settings) contains an invalid setting, all settings files are ignored and an error is printed.
Fixed
After wandb login --host <invalid-url>, using wandb login --host <valid-url> works as usual.
Fixed a bug where wandb beta sync can get stuck on Syncing... and fail to load credentials.
Fixed a bug that could occasionally lead to unuploaded data. Now, all data is stored in the run’s .wandb file and can be reuploaded with wandb sync.
These regressions were introduced in 0.24.0.
January 13, 2026
This release has been removed from PyPI. Do not use this version. If you are on this version, upgrade to v0.24.1 or a newer version immediately. Due to a regression fixed in v0.24.1, runs created with wandb==0.24.0 may fail to upload some data.
Notable Changes
This version removes the legacy, deprecated wandb.beta.workflows module, including its log_model()/use_model()/link_model() functions. This is formally a breaking change.
Added
wandb agent and wandb.agent() now accept a forward_signals flag (CLI: --forward-signals/-f) to relay SIGINT/SIGTERM and other catchable signals from the agent to its sweep child runs, enabling cleaner shutdowns when you interrupt an agent process.
wandb beta sync now supports a --live option for syncing a run while it’s being logged.
Removed
Removed the deprecated wandb.beta.workflows module, including its log_model(), use_model(), and link_model() functions, and whose modern successors are the Run.log_artifact, Run.use_artifact, and Run.link_artifact methods, respectively.
Fixed
Fixed Run.__exit__ type annotations to accept None values, which are passed when no exception is raised.
Fixed Invalid Client ID digest error when creating artifacts after calling random.seed(). Client IDs could collide when random state was seeded deterministically.
Fixed CLI error when listing empty artifacts.
Fixed regression for calling api.run() on a Sweeps run.
Fixed the “View run at” message printed at the end of a run which sometimes did not include a URL.
Runs queried from wandb.Api() now display a string representation in VSCode notebooks instead of a broken HTML window.
December 2, 2025
Added
Regex support in metrics and run overview filters in wandb beta leet command.
Chart inspection in wandb beta leet command: right-click and drag to show (x, y) at the nearest data point; hold Alt for synchronized inspection across all visible charts.
The Automations API now supports creating and editing automations that trigger on run states.
The Automations API now supports basic zscore automation events.
beta_history_scan method to Run objects for faster history scanning performance with wandb.Api.
Changed
wandb.Api() now raises a UsageError if WANDB_IDENTITY_TOKEN_FILE is set and an explicit API key is not provided.
Deprecated
Anonymous mode is deprecated and will emit warnings. This deprecation includes the anonymous setting, the WANDB_ANONYMOUS environment variable, wandb.init(anonymous=...), and commands wandb login --anonymously and wandb.login(anonymous=...).
Fixed
wandb.Image() no longer prints a deprecation warning.
Registry.description and ArtifactCollection.description no longer reject empty strings.
Instantiating Artifact objects is now significantly faster.
wandb.Run.save() now falls back to hardlinks and, if needed, copying (downgrading the ‘live’ file policy to ‘now’, if applicable) when symlinks are disabled or unavailable (e.g., cross‑volume or no Developer Mode on Windows).
Artifact collection aliases are now fetched lazily on accessing ArtifactCollection.aliases instead of on instantiating ArtifactCollection, improving performance of Api.artifact_collections(), Api.registries().collections(), etc.
Use explicitly provided API key in wandb.init(settings=wandb.Settings(api_key="...")) for login. Use the key from run when logging artifacts via run.log_artifact().
W&B LEET TUI correctly displays negative Y axis tick values and base/display units of certain system metrics.
Fixed a rare infinite loop in console_capture.py.
File upload/download now respects WANDB_X_EXTRA_HTTP_HEADERS except for reference artifacts.
November 11, 2025
Added
Experimental wandb beta leet command - Lightweight Experiment Exploration Tool - a terminal UI for viewing W&B runs locally with real-time metrics visualization and system monitoring.
The registry API now supports programmatic management of user and team members of individual registries.
Registry.id has been added as a (read-only) property of Registry objects.
Fixed
Artifact.files() now has a correct len() when filtering by the names parameter.
The numerator for file upload progress no longer occasionally exceeds the total file size.
Artifact.link() now logs unsaved artifacts instead of raising an error, consistent with the behavior of Run.link_artifact().
Automatic code saving now works when running ipython notebooks in VSCode’s Jupyter notebook extension.
Logging an artifact with infinite floats in Artifact.metadata now raises a ValueError early, instead of waiting on request retries to time out.
October 28, 2025
Added
Settings console_chunk_max_seconds and console_chunk_max_bytes for size- and time-based multipart console logs file chunking.
Registry API query methods (Api.registries(), Registry.{collections,versions}(), Api.registries().{collections,versions}()) now accept a per_page keyword arg to override the default batch size for paginated results.
Changed
API keys longer than 40 characters are now supported.
Fixed
run.config now properly returns a dict when calling artifact.logged_by() in v0.22.1.
wandb.Api(api_key=...) now prioritizes the explicitly provided API key over thread-local cached credentials.
Fixed a rare deadlock in console_capture.py.If you dump thread tracebacks during the deadlock and see the wandb-AsyncioManager-main thread stuck on a line in console_capture.py: this is now fixed.
Fixed an issue where TensorBoard sync would sometimes stop working if the tfevents files were being written live.
Artifact.manifest delays downloading and generating the download URL for the artifact manifest until it’s first used. If the manifest has not been locally modified, Artifact.size and Artifact.digest can return without waiting to fetch the full manifest.
Fixed uploading GCS folder references via artifact.add_reference.
The SDK now correctly infers notebooks paths in Jupyter sessions, using the server’s root directory, so code saving works in subdirectories (e.g. code/nested/<notebook>.ipynb).
Optimize artifacts downloads re-verification with checksum caching.
Lazy loading support for Api().runs() to improve performance when listing runs. The new lazy=True parameter (default) loads only essential metadata initially, with automatic on-demand loading of heavy fields like config and summary when accessed.
Add storage_region option when creating artifacts. Users can use CoreWeave AI Object Storage by specifying wandb.Artifact(storage_region="coreweave-us") when using wandb.ai for faster artifact upload/download on CoreWeave’s infrastructure.
Fixed
Api.artifact_exists() and Api.artifact_collection_exists() now raise on encountering timeout errors, rather than (potentially erroneously) returning False.
Remove a build targeting MacOS 10.x due to multiple security and supply chain considerations.
Fixed
Resuming a run with a different active run will now raise an error unless you call run.finish() first, or call wandb.init() with the parameter reinit='create_new'.
Fix Api().runs() for wandb server < 0.51.0 (when project.internalId was added to gql API).
Sweeps: command run scripts that import readline whether directly or indirectly (e.g. import torch on Python 3.13) should no longer deadlock.
Add DSPy integration: track evaluation metrics over time, log predictions and program signature evolution to W&B Tables, and save DSPy programs as W&B Artifacts (complete program or state as JSON/PKL).
This version raises errors that would previously have been suppressed during calls to Artifact.link() or Run.link_artifact(). While this prevents undetected failures in those methods, it is also a breaking change.
Added
New settings for max_end_of_run_history_metrics and max_end_of_run_summary_metrics.
New wandb.integration.weave module for automatically initializing Weave when a W&B run is active and weave is imported.
Changed
Errors encountered while linking an artifact are no longer suppressed/silenced, and Artifact.link() and Run.link_artifact() no longer return None.
The “Run history” and “Run summary” printed at the end of a run are now limited to 10 metrics each.
Fixed
Dataclasses in a run’s config no longer raise Object of type ... is not JSON serializable when containing real classes as fields to the dataclass.
Artifact.link() and Run.link_artifact() should be faster on server versions 0.74.0+, requiring 4-5 fewer unnecessary blocking GraphQL requests.
The default ordering for Api().runs(...) and Api().sweeps(...) is now ascending order based on the runs created_at time.
Added
Support first summary option in define_metric.
Add support for paginated sweeps.
pattern parameter to Api().run().files to only get files matching a given pattern from the W&B backend.
Add optional format key to Launch input JSONSchema to specify a string with a secret format.
Changed
Sweep.name property will now return user-edited display name if available (falling back to original name from sweep config, then sweep ID as before).
Api().runs(...) and Api().sweeps(...) now returns runs in ascending order according to the runs created_at time.
Artifact with large file (>2GB) uploads faster by using parallel hashing on system with more cores.
Remove the implementation of __bool__ for the registry iterators to align with python lazy iterators.
Deprecated
The wandb.beta.workflows module and its contents (including log_model(), use_model(), and link_model()) are deprecated and will be removed in a future release.
Fixed
Correct the artifact url for organization registry artifacts to be independent of the artifact type.
Suffixes on sanitized InternalArtifact names have been shortened to 6 alphanumeric characters.
wandb.Video will not print a progress spinner while encoding video when WANDB_SILENT/WANDB_QUIET environment variables are set.
Fixed registries fetched using api.registries() from having an extra wandb-registry- prefix in the name and full_name fields.
Fixed a crash that could happen when using sync_tensorboard.
Api().run(...).upload_file no longer throws an error when uploading a file in a different path relative to the provided root directory.
Calling load() function on a public API run object no longer throws TypeError.
When a Sweeps run function called by wandb.agent() API throws an exception, it will now appear on the logs page for the run. (This previously only happened for runs called by the wandb agent CLI command.).
This version removes the legacy implementation of the service process. This is a breaking change.
Added
Setting x_stats_track_process_tree to track process-specific metrics such as the RSS, CPU%, and thread count in use for the entire process tree, starting from x_stats_pid. This can be expensive and is disabled by default.
Notes are now returned to the client when resuming a run.
Added support for creating custom Vega chart presets through the API. Users can now define and upload custom chart specifications that can be then reused across runs with wandb.plot_table().
Changed
Calling Artifact.link() no longer instantiates a throwaway placeholder run.
wandb now attempts to use Unix sockets for IPC instead of listening on localhost, making it work in environments with more restrictive permissions (such as Databricks).
Api.artifact() will now display a warning while fetching artifacts from migrated model registry collections.
The .length for objects queried from wandb.Api has been deprecated. Use len(...) instead.
Removed
Removed the legacy python implementation of the service process. The legacy-service option of wandb.require() as well as the x_require_legacy_service and x_disable_setproctitle settings with the corresponding environment variables have been removed and will now raise an error if used.
Removed the private wandb.Run._metadata attribute. To override the auto-detected CPU and GPU counts as well as the GPU type, please use the new settings x_stats_{cpu_count,cpu_logical_count,gpu_count,gpu_type}.
Fixed
Allow s3 style CoreWeave URIs for reference artifacts.
Fixed rare bug that made Ctrl+C ineffective after logging large amounts of data.
Respect silent, quiet, and show_warnings settings passed to a Run instance for warnings emitted by the service process.
api.Runs no longer makes an API call for each run loaded from W&B.
Correctly parse the x_extra_http_headers setting from the env variable.
.length calls the W&B backend to load the length of objects when no data has been loaded rather than returning None.
wandb.Image() was broken in 0.20.0 when given NumPy arrays with values in the range [0, 1], now fixed.
wandb.Table: Added new constructor param, log_mode, with options "IMMUTABLE" and "MUTABLE". IMMUTABLE log mode (default) is existing behavior that only allows a table to be logged once. MUTABLE log mode allows the table to be logged again if it has been mutated. (@domphan-wandb in https://github.com/wandb/wandb/pull/9758)
This version removes the ability to disable the service process. This is a breaking change.
Added
Added merge parameter to Artifact.add_dir to allow overwrite of previously-added artifact files.
Support for pytorch.tensor for masks and boxes parameters when creating a wandb.Image object.
sync_tensorboard now supports syncing tfevents files stored in S3, GCS and Azure.
GCS paths use the format gs://bucket/path/to/log/dir and rely on application-default credentials, which can be configured using gcloud auth application-default login
S3 paths use the format s3://bucket/path/to/log/dir and rely on the default credentials set through aws configure
Azure paths use the format az://account/container/path/to/log/dir and the az login credentials, but also require the AZURE_STORAGE_ACCOUNT and AZURE_STORAGE_KEY environment variables to be set. Some other environment variables are supported as well, see here.
Added support for initializing some Media objects with pathlib.Path.
New setting x_skip_transaction_log that allows to skip the transaction log. Note: Should be used with caution, as it removes the guarantees about recoverability.
normalize parameter to wandb.Image initialization to normalize pixel values for Images initialized with a numpy array or pytorch tensor.
Changed
Various APIs now raise TypeError instead of ValueError or other generic errors when given an argument of the wrong type.
Various Artifacts and Automations APIs now raise CommError instead of ValueError upon encountering server errors, so as to surface the server error message.
wandb.sdk.wandb_run.Run::save method now requires the glob_str argument.
Removed
Removed support for disabling the service process. The x_disable_service/_disable_service setting and the WANDB_DISABLE_SERVICE/WANDB_X_DISABLE_SERVICE environment variable have been deprecated and will now raise an error if used.
Removed ability to use wandb.docker after only importing wandb.
wandb.docker is not part of wandb’s public interface and is subject to breaking changes. Please do not use it.
Removed no-op sync argument from wandb.Run::log function.
The start_method setting is deprecated and has no effect; it is safely ignored.
The property Artifact.use_as and parameter use_as for run.use_artifact() are deprecated since these have not been in use for W&B Launch.
Fixed
Calling wandb.teardown() in a child of a process that called wandb.setup() no longer raises WandbServiceNotOwnedError.
This error could have manifested when using W&B Sweeps
Offline runs with requested branching (fork or rewind) sync correctly.
Log exception as string when raising exception in Job wait_until_running method.
wandb.Image initialized with tensorflow data would be normalized differently than when initialized with a numpy array.
Using wandb login no longer prints a warning about wandb.require("legacy-service").
Logging a Table (or other objects that create internal artifacts) no longer raises ValueError when logged from a run whose ID contains special characters.
wandb.Api initialized with the base_url now respects the provided url, rather than the last login url.
Added creation, deletion, and updating of registries in the SDK.
artifact.is_link property to artifacts to determine if an artifact is a link artifact (such as in the Registry) or source artifact.
artifact.linked_artifacts to fetch all the linked artifacts to a source artifact and artifact.source_artifact to fetch the source artifact of a linked artifact.
run.link_artifact(), artifact.link(), and run.link_model() all return the linked artifact upon linking.
Multipart download for artifact file larger than 2GB, user can control it directly using artifact.download(multipart=True).
Project.id property to get the project ID on a wandb.public.Project.
New public API for W&B Automations.
New submodules and classes in wandb.automations.* to support programmatically managing W&B Automations.
Api.integrations(), Api.slack_integrations(), Api.webhook_integrations() to fetch a team’s existing Slack or webhook integrations.
Api.create_automation(), Api.automation()/Api.automations(), Api.update_automation(), Api.delete_automation() to create, fetch, edit, and delete Automations.
Create and edit automations triggered on RUN_METRIC_CHANGE events, i.e. on changes in run metric values (absolute or relative deltas).
Ability to collect profiling metrics for Nvidia GPUs using DCGM. To enable, set the WANDB_ENABLE_DCGM_PROFILING environment variable to true. Requires the nvidia-dcgm service to be running on the machine. Enabling this feature can lead to increased resource usage.
Fixed
run.log_code correctly sets the run configs code_path value.
Correctly use WANDB_CONFIG_DIR for determining system settings file path.
Prevent invalid Artifact and ArtifactCollection names (which would make them unloggable), explicitly raising a ValueError when attempting to assign an invalid name.
Prevent pydantic ConfigError in Pydantic v1 environments from not calling .model_rebuild()/.update_forward_refs() on generated types with ForwardRef fields.
wandb.init() no longer raises Permission denied error when the wandb directory is not writable or readable.
Calling file.delete() on files queried via api.Runs(...) no longer raises CommError.
The new reinit="create_new" setting causes wandb.init() to create a new run even if other runs are active, without finishing the other runs (in contrast to reinit="finish_previous"). This will eventually become the default.
Added Artifact.history_step to return the nearest run step at which history metrics were logged for the artifact’s source run.
Added data_is_not_path flag to skip file checks when initializing wandb.Html with a sting that points to a file.
Changed
Artifact.download() no longer raises an error when using WANDB_MODE=offline or when an offline run exists.
Removed
Dropped the -q / --quiet argument to the wandb magic in IPython / Jupyter; use the quiet run setting instead.
Deprecated
The following wandb.Run methods are deprecated in favor of properties and will be removed in a future release (@kptkin in https://github.com/wandb/wandb/pull/8925):.
run.project_name() is deprecated in favor of run.project
run.get_url() method is deprecated in favor of run.url
run.get_project_url() method is deprecated in favor of run.project_url
run.get_sweep_url() method is deprecated in favor of run.sweep_url
Fixed
Fixed ValueError on Windows when running a W&B script from a different drive.
Fix base_url setting was not provided to wandb.login().
wandb.Html() no longer raises IsADirectoryError with a value that matched a directory on the users system.
Added support for building artifact file download urls using the new url scheme, with artifact collection membership context.
Changed
Boolean values for the reinit setting are deprecated; use “return_previous” and “finish_previous” instead.
The “wandb” logger is configured with propagate=False at import time, whereas it previously happened when starting a run. This may change the messages observed by the root logger in some workflows.
Metaflow now requires plum-dispatch package.
Relaxed the pydantic version requirement to support both v1 and v2.
Existing pydantic types have been adapted to be compatible with Pydantic v1.
wandb.init(dir=...) now creates any nonexistent directories in dir if it has a parent directory that is writeable.
The server now supports fetching artifact files by providing additional collection information; updated the artifacts api to use the new endpoints instead.
Paginated methods (and underlying paginators) that accept a per_page argument now only accept int values. Default per_page values are set directly in method signatures, and explicitly passing None is no longer supported.
Fixed
Calling wandb.init() in a notebook finishes previous runs as previously documented.
Bug introduced in 0.19.0
Fixed an error being thrown when logging jpg/jpeg images containing transparency data.
wandb.init(resume_from=...) now works without explicitly specifying the run’s id.
Deleting files with the Public API works again.
Bug introduced in 0.19.1
Fixed media files not displaying in the UI when logging to a run with a custom storage bucket.
Prometheus API support for Nvidia DCGM GPU metrics collection.
Changed
Changed Nvidia GPU ECC counters from aggregated to volatile.
Fixed
Fixed a performance issue causing slow instantiation of wandb.Artifact, which in turn slowed down fetching artifacts in various API methods.
Some errors from wandb.Api have better string representations.
Artifact.add_reference, when used with file URIs for a directory and the name parameter, was incorrectly adding the value of name to the path of the file references.
Fixed a bug causing Artifact.add_reference() with checksum=False to log new versions of local reference artifacts without changes to the reference URI.
Added wandb login --base-url {host_url} to login as an alias of wandb login --host {host_url}.
Changed
Temporarily disabled collecting per-core CPU utilization stats.
Fixed
Fixed a bug causing offline mode to make network requests when logging media artifacts. If you are using an older version of W&B Server that does not support offline artifact uploads, use the setting allow_offline_artifacts=False to revert to older compatible behavior.
Expand sanitization rules for logged table artifact name to allow for hyphens and dots. This update brings the rules up-to-date with the current rules for artifact names. (Allowing letters, numbers, underscores, hyphens, and dots).
Correctly handle run rewind settings fork_from and resume_from.
Fix incorrectly reported device counts and duty cycle measurements for TPUs with single devices per chip / multiple devices on the host and make TPU metrics sampling more robust.
Handle non-consecutive TPU device IDs in system monitor.
Calling wandb.init(mode="disabled") no longer disables all later runs by default. Use wandb.setup(settings=wandb.Settings(mode="disabled")) for this instead, or set mode="disabled" explicitly in each call to wandb.init().
Fixed
The stop button correctly interrupts runs whose main Python thread is running C code, sleeping, etc.
Remove unintentional print that occurs when inspecting wandb.Api().runs().
Fix uploading large artifacts when using Azure Blob Storage.
The wandb offline command no longer adds an unsupported setting to wandb.Settings, resolving ValidationError.
Fix error when reinitializing a run, caused by accessing a removed attribute.
Fixed occasional deadlock when using multiprocessing to update a single run from multiple processes.
Prevent errors from bugs in older versions of botocore < 1.5.76.
Fixed various checks against invalid anonymous settings value.
Removed
The wandb.wandb_sdk.wandb_setup._setup() function’s reset parameter has been removed. Note that this is a private function, even though there appear to be usages outside of the repo. Please wandb.teardown() instead of _setup(reset=True).
In the private wandb.wandb_sdk.wandb_setup module, the logger and _set_logger symbols have been removed.
Security
Bump github.com/go-git/go-git version to 5.13.0 to address CVE-2025-21613.
Bump golang.org/x/net version to 0.33.0 to address CVE-2024-45338.
Remove wandb.Run.plot_table method. The functionality is still available and should be accessed using wandb.plot_table, which is now the recommended way to use this feature.
Drop support for Python 3.7.
Fixed
Fix ultralytics reporting if there are no positive examples in a validation batch.
Debug printing for hyperband stopping algorithm printed one char per line.
Include the missing log_params argument when calling lightgbm’s wandb_callback function.
Added create_and_run_agent to __all__ in wandb/sdk/launch/__init__.py to expose it as a public API.
Changed
Tables logged in offline mode now have updated keys to artifact paths when syncing. To revert to old behavior, use setting allow_offline_artifacts = False.
Deprecated
The quiet argument to wandb.run.finish() is deprecated, use wandb.Settings(quiet=...) to set this instead.
Fixed
Fix api.artifact() to correctly pass the enable_tracking argument to the Artifact._from_name() method.
Track detailed metrics for Apple ARM systems including GPU, eCPU, and pCPU utilization, power consumption, and temperature, and memory/swap utilization.
Allow users to link Registry artifacts without inputting the organization entity name.
Added a warning message indicating that the fps argument will be ignored when creating a wandb.Video object from a file path string or a bytes object.
Update docstrings for logged_artifacts and used_artifacts methods in Run class.
The _show_operation_stats setting enables a preview of a better run.finish() UX.
Fixed
Log power on AMD MI300X series GPUs.
Fixed typing issue of wandb.Api.
Ensure artifact objects are fully updated on Artifact.save().
Changed
Updated minimum version of sentry-sdk to 2.0.0 to address deprecation warnings.
Add the ability to monitor the utilization metrics of Google’s Cloud TPU devices.
Fixed
Capture Nvidia GPU stats on Windows.
Fixed a regression introduced in v0.18.2 that affected capturing the names of Nvidia GPU devices.
run.log_artifact() no longer blocks other data uploads until the artifact upload finishes.
Fixed media dependency for rdkit updated from rdkit-pypi to rdkit.
Saving an artifact with many large files no longer exhausts OS threads.
Changed
After artifact = run.log_artifact(), you must use artifact.wait() before operations that rely on the artifact having been uploaded. Previously, this wasn’t necessary in some cases because run.log_artifact() blocked other operations on the run.
This version switches wandb to a new backend by enabling wandb.require("core") by default. This should not be a breaking change, but the new backend may have unexpected differences in behavior for legacy functionality and rare edge cases.
Added
Add support for artifact tags, via Artifact.tags and Run.log_artifact().
Fixed
Detect the notebook name in VS Code’s built-in jupyter server.
Changed
The new “core” backend, previously activated using wandb.require(“core”), is now used by default. To revert to the legacy behavior, add wandb.require("legacy-service") at the beginning of your script. Note: In a future minor release, the option to disable this new behavior will be removed.
Changed the default system metrics sampling interval to 10 seconds without averaging, while allowing custom intervals via wandb.init(settings=wandb.Settings(_stats_sampling_interval=...)).
Deprecated
define_metric(summary='best', goal=...) is deprecated and soon will be removed, use define_metric(summary='min') or define_metric(summary='min') instead.
Capture SM (Streaming Multiprocessor), memory, and graphics clock speed (MHz), (un)corrected error counts, fan speed (%), and encoder utilization for Nvidia GPU devices when using core.
Allow iterating over wandb.Config like a dictionary.
PR curves, images and histograms are supported when using TensorBoard with core enabled.
Added wandb.require("legacy-service") as the opt-out analog of wandb.require("core").
Fixed
Avoid leaving behind wandb-core process if user hits Ctrl+C twice.
Fix deprecation warnings arising from NumPy >= 2.1 by removing newshape argument from numpy.reshape.
Skip uploading/downloading GCS reference artifact manifest entries corresponding to folders.
Deprecated
Ability to disable the service process (WANDB__DISABLE_SERVICE) is deprecated and will be removed in the next minor release.
Added api.runs().histories() to fetch history metrics for runs that meet specified conditions.
Display warning when Kubernetes pod fails to schedule.
Added ArtifactCollection.save() to allow persisting changes.
Added the ability to overwrite history of previous runs at an arbitrary step and continue logging from that step.
Added new Workspace API for programmatically editing W&B Workspaces.
Added Artifact.unlink() to allow programmatic unlinking of artifacts.
Added basic TensorBoard support when running with wandb.require("core").
The TensorBoard tab in W&B will work.
Charts show up in W&B, possibly better than when running without core.
Not all types of data are supported yet. Unsupported data is not shown in charts.
Fixed
Fix define_metric behavior for multiple metrics in shared mode.
Correctly pass in project name to internal api from run while calling run.use_artifact().
Correctly upload console output log files when resuming runs enabled with console_multipart setting by @kptkin in PR #7694 and @dmitryduev in PR #7697.
Renamed wandb.plots to wandb.plot, renamed all integrations from wandb.<name> to wandb.integration.<name>, and removed the [async] extra.This version packages the wandb-core binary, formerly installed by the wandb-core Python package on PyPI. The wandb-core package is now unused and can be uninstalled.
Added
The wandb package now includes the wandb-core binary.
wandb-core is a new and improved backend for the W&B SDK that focuses on performance, versatility, and robustness.
Currently, it is opt-in. To start using the new backend, add wandb.require("core") to your script after importing wandb.
Added artifact_exists() and artifact_collection_exists() methods to Api to check if an artifact or collection exists.
wandb launch -u <git-uri | local-path> creates and launches a job from the given source code.
Fixed
Prevent crash on run.summary for finished runs.
Correctly report file upload errors when using wandb-core.
Implemented a stricter check for AMD GPU availability.
Fixed run.save() on Windows.
Show a warning instead of failing when using registries other than ECR and GAR with the Kaniko builder.
Fixed wandb.init() type signature including None.
Changed
When using wandb-core need to specify a required flag (wandb.require("core")) to enable it, before it was picked up automatically.
Use ETags instead of MD5 hashes for GCS reference artifacts.
Removed
Removed the deprecated wandb.plots.* functions and top-level third-party integrations wandb.[catboost,fastai,keras,lightgbm,sacred,xgboost]. Please use wandb.plot instead of wandb.plots and wandb.integration.[catboost,fastai,keras,lightgbm,sacred,xgboost] instead of wandb.[catboost,fastai,keras,lightgbm,sacred,xgboost].
Removed the [async] extra and the _async_upload_concurrency_limit setting.
Removed undocumented settings: _except_exit and problem.
Added support for overriding kaniko builder settings in the agent config.
Added link to the project workspace of a run in the footer.
Added support for overriding stopped run grace period in the agent config.
Added setting (_disable_update_check) to disable version checks during init.
WandbLogger.sync in the OpenAI Fine-Tuning integration gets a new log_datasets boolean argument to turn off automatic logging of datasets to Artifacts.
Reduced default status print frequency of launch agent. Added verbosity controls to allow for increased status print frequency and printing debug information to stdout.
Changed
Limit policy option on artifact cli’s put() to choices, [“mutable”, “immutable”].
Updated artifact public api methods to handle nullable Project field on the ArtifactSequence/ArtifactCollection type, based on gorilla server changes.
Fixed
Fixed run.save() not working with files inside run.dir, introduced in previous release.
Fixed rare panic during large artifact uploads.
Fixed wandb.login() causing runs not to be associated with launch queue.
Fixed job artifact download failing silently and causing run crash when using W&B Launch by @KyleGoyette PR #7285.
Fix handling of saving training files to Artifacts in the OpenAI Fine-Tuning integration.
Fixed a race where the wandb process could become unresponsive at the end of a run.
December 22, 2018
Bug Fix
Fix resuming in Jupyter on kernel restart.
Wandb.save ensures files are pushed regardless of growth.
Enhancement
Added replace=True keyword to init for auto-resuming.
New run.resumed property that can be used to detect if we’re resuming.
New run.step property to use for setting an initial epoch on resuming.
Made Keras callback save the best model as it improves.
December 20, 2018
Bug Fix
Really don’t require numpy.
Better error message if wandb.log() is called before wandb.init().
Prevent calling wandb.watch() multiple times.
Handle datetime attributes in logs / plotly.
Enhancement
Add environment to sweeps.
Enable tagging in the public API and in wandb.init().
New media type wandb.Html for logging arbitrary html.
Add Public api.create_run method for custom integrations.
Added glob support to wandb.save(), files save as they’re written to.
Added wandb.restore() for pulling files on resume.
December 06, 2018
Bug Fix
Added a timeout for generating diffs on large repos.
Fixed edge case where file syncing could hang.
Ensure all file changes are captured before exit.
Handle cases of sys.exit where code isn’t passed.
Don’t require numpy.
Enhancement
New wandb sync command that pushes a local directory to the cloud.
Support for syncing tfevents file during training.
Detect when running as TFJob and auto group.
New Kubeflow module with initial helpers for pipelines.
November 26, 2018
Bug Fix
Fixed history / summary bug.
November 24, 2018
Enhancement
Initial support for AWS SageMaker.
hook_torch renamed to watch with a deprecation warning.
Projects are automatically created if they don’t exist.
Additional GPU memory_allocated metric added.
Keras Graph stores edges.
Bug Fix
PyTorch graph parsing is more robust.
Fixed PyTorch 0.3 support.
File download API supports WANDB_API_KEY authentication.
November 13, 2018
Enhancement
Sweeps work with new backend (early release).
Summary tracks all history metrics unless they’re overridden by directly writing.
to summary.
Files support in data API.
Bug Fix
Show ongoing media file uploads in final upload progress.
November 09, 2018
Enhancement
Wandb.Audio supports duration.
Bug Fix
Pass username header in filestream API.
November 08, 2018
Enhancement
New wandb.Audio data type.
New step keyword argument when logging metrics.
Ability to specify run group and job type when calling wandb.init() or via.
environment variables. This enables automatic grouping of distributed training runs
in the UI
Ability to override username when using a service account API key.
Bug Fix
Handle non-tty environments in Python2.
Handle non-existing git binary.
Fix issue where sometimes the same image was logged twice during a Keras step.
October 19, 2018
Enhancement
PyTorch.
Added a new wandb.hook_torch method which records the graph and logs gradients & parameters of pytorch models
wandb.Image detects pytorch tensors and uses torchvision.utils.make_grid to render the image.
Bug Fix
wandb restore handles the case of not being run from within a git repo.
October 18, 2018
Bug Fix
We now open stdout and stderr in raw mode in Python 2 ensuring tools like bpdb work.
October 12, 2018
Enhancement
Catastrophic errors are now reported to Sentry unless WANDB_ERROR_REPORTING is set to false.
Improved error handling and messaging on startup.
October 05, 2018
Bug Fix
The first image when calling wandb.log() was not being written, now it is.
wandb.log() and run.summary now remove whitespace from keys.
October 05, 2018
Bug Fix
Vendored prompt_toolkit < 1.0.15 because the latest ipython is pinned > 2.0.
Lazy load wandb.h5 only if summary is accessed to improve Data API performance.
Enhancement
Jupyter.
Deprecated wandb.monitor in favor of automatically starting system metrics after the first wandb.log() call
Added new %%wandb jupyter magic method to display live results
Removed jupyter description iframe
The Data API now supports per_page and order options to the api.runs method.