Release 0.3.0 by piotrekno1 · Pull Request #98 · wild-edge/wildedge-python

piotrekno1 · 2026-10-21T17:49:47Z

Release 0.3.0 to main. Tag v0.3.0 on the merge commit.

Highlights (full list in CHANGELOG):

New integrations: mlx_whisper and faster_whisper, OpenAI responses.create() and embeddings.create(), Anthropic messages.stream().
The llama.cpp integration is renamed llama_cpp (gguf remains an alias) and now tracks chat and completion calls.
Instrumentation ownership: init(dsn=...) with the runner's DSN reuses its client. A conflicting second client warns, and --strict-integrations raises (Release 0.3.0 #96).
Fixes: duplicate events from flush() (Release 0.3.0 #96), transformers double counting, timm and ONNX model naming, Hugging Face hub patch importing unused transformers modules (Stop the Hugging Face hub patch from importing unused transformers modules #97).

This PR also moves the Hugging Face hub entry from Unreleased into the 0.3.0 section, since #97 merged before tagging.

Verification on this branch

910 passed, 28 skipped on Python 3.10, 3.13 and 3.14. Ruff is clean. The wheel and sdist pass twine check. check_tag_version.py, the release notes script and llms.txt all run.
All examples ran against local ingest through a logging proxy. Every batch was accepted, with no rejected events and no duplicate event IDs. This covers local models, both Whisper backends, attachments, OpenAI, Responses, embeddings, Anthropic, agentic, llm_api, the cli wrapper, FastAPI with --reload, Django with waitress, and the README snippets.
Checks for the fixes: the demo app's init() pattern under wildedge run, the same-DSN, other-DSN and strict cases, the flush race script (0 duplicates), a real fork(), and the Hugging Face hub patch (no extra imports, download functions still tracked).

🤖 Generated with Claude Code

Merge release/0.2.2 into devel

The gguf integration patched only Llama.__call__, but create_chat_completion() runs llama.create_completion() through the chat handler, so chat apps produced no inference events. Patch create_completion instead: __call__, create_completion and create_chat_completion all run it, so each call is tracked once. patch_instance_call_once takes a method name (default __call__). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…N comes from Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

README: PyPI badge, links to the app, docs and blog, and where the DS…

mlx 0.22.0 is gone from PyPI, which broke the daily run. Current pins move to the latest releases; numpy 2.2.6 is the newest that covers 3.10 to 3.13, with 3.10 overrides for onnxruntime and keras. Anthropic had pins and a test but was missing from the CI matrix. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

create_chat_completion now exposes its messages to the create_completion it runs, so chat events get prompt_type="chat" and turn_index while tracking stays in one place. Non-streamed calls report finish_reason, every call reports its sampling arguments with the method's defaults filled in, and with attachments enabled the prompt or messages and the reply are attached. Closes #64

Every compat test used importorskip, so a broken install skipped and the job passed (ultralytics min on 3.12/3.13 did), and half of them only called instrument(). Each test now loads a real model with the pinned library and checks the load and inference events, without network access: canned HTTP for openai and anthropic, small committed ONNX and GGUF files, and models built from config for the rest. A conftest hook fails any skip. Bugs found are marked xfail(strict=True) with their issue: #64, #70, #71, #72, #73. tests/compat is left out of the regular pytest run.

…el PRs The workflow's matrix was hand-written and drifted from the script (anthropic was missing), and pull requests into devel never ran compat. compat_matrix.py now builds the matrix: a pr profile with each integration's min and current pins, and the full nightly profile. A verify step checks the installed versions match the pins. tests/test_compat_matrix.py keeps the script, the compat tests and docs/compatibility.md in step.

ultralytics 8.0.0 imports pkg_resources without declaring setuptools, and Python 3.12+ environments no longer include it. The old compat test skipped here; with skips now failing, the full run caught it.

Every devel pull request running the matrix is more than needed; devel gets the nightly run. Pull requests into main or a release branch run the full matrix, so the separate pr profile is gone.

The gguf compat test fetches the 1.2 MB TinyStories model from a pinned Hugging Face revision on first use and checks its sha256. A broken link or changed file fails the test with the URL to update. Models are cached in tests/compat/.cache, which the gguf jobs keep between runs.

Fix the compat matrix and make it test each integration

…e-completion

Replaces the #64 xfail with compat tests for create_completion (stop reason and generation config with defaults), create_chat_completion (prompt_type, turn_index) and streamed chat, each tracked once. They fail on devel and pass with this branch on llama-cpp-python 0.2.90 and 0.3.36.

The example only called llm(...). It now runs three streamed create_chat_completion turns and one create_completion, so it exercises the chat tracking from this PR.

Lint was only run by hand, so misplaced imports and similar slips reached review. CI gets a lint job running ruff check and ruff format --check, and the rule set adds bugbear (B), comprehensions (C4), PIE, import placement (PLC0415), pytest style (PT), RET, RUF and SIM. Test doubles may keep mutable class attributes (RUF012).

Mostly mechanical: sorted __all__, removed unused noqa directives, dropped else after return, dict literals, itertools.pairwise, ClassVar on the client's class-level tables, contextlib.suppress in the hub cache walk, and an explicit zip strict flag (strict where lengths must match, not where a list can fall back to empty). llm_api imports get_client at module level: the import cycle its noqa mentioned no longer exists.

Imports move to module level, except numpy, which stays optional through pytest.importorskip; the sampler test reads platforms._sampler at call time since start_sampler replaces it. Also tuple parametrize names, split composite asserts, combined with statements, match= on pytest.raises, raise from None in fake async iterators, and a duplicate parametrize case removed.

Track llama-cpp chat and completion calls in the gguf integration

Ingest rejects the whole batch with 422 when a span attribute value is a list or dict, so one bad value dropped every unrelated event sent with it. SpanEvent.to_dict now sends such values as JSON strings and logs one warning per key. Closes #61

Send non-scalar span attributes as JSON strings

close() stopped the uploader without a final pass, so attachments captured in the last few seconds stayed on disk while their events were sent. With a timeout, close() and flush() now drain the attachment store in the time left after the event flush. Each request's timeout is capped at that deadline, and a lock keeps the background thread and the drain from sending the same attachment. Closes #62

…ests

Upload buffered attachments on close and flush with a timeout

main_graph, the default graph name from newer torch.onnx.export, wasn't in the generic list, so every Piper voice registered as one model called main_graph. It now falls back to the session's file name, which also gives a manually registered session an id. model_version 0, negative or INT64_MAX counts as unset. Closes #63

Improve README quickstart and add demo preview

Add compact trace and SQL previews to README

openai 1.30.0 passes proxies= to httpx, which httpx 0.28 removed.

…s a change Drops the llama.cpp line about automatic prompt and reply attachments, which the explicit-attachments change removed.

The fork hooks hold the drain lock across fork() so a child never inherits it locked.

init(dsn=...) now reuses the default client when the DSN points to the same project. A second client enabling integrations while another live client owns them logs a warning and leaves automatic tracking with the owner, as 0.2.2 did silently. wildedge run --strict-integrations keeps the RuntimeError.

Release 0.3.0

getattr on transformers' lazy modules imported unused model modules and logged a debug line for each one that failed to import.

Stop the Hugging Face hub patch from importing unused transformers modules

github-actions · 2026-10-21T17:49:59Z

Changelog preview for `v0.3.0`

Preview of the release notes that will be generated when this is tagged.

Added

mlx_whisper and faster_whisper integrations track local Whisper model
loads and transcriptions, including faster-whisper's
BatchedInferencePipeline. Transcriptions record duration, language, word
count and audio metadata. Audio and transcript attachments must be supplied
explicitly through the manual tracking API. See docs/whisper.md.
The OpenAI integration now tracks responses.create() for synchronous and
asynchronous clients, including streaming. Records input/output token usage,
cached and reasoning tokens, response status or incomplete reason, resolved
model metadata, and time to first output delta for streams. Older OpenAI SDKs
without Responses remain supported.
OpenAI sync and async embeddings.create() instrumentation records latency,
input token usage, batch size, vector dimensions, model metadata and errors.
Embedding vectors and raw inputs are never attached by this adapter.
EmbeddingOutputMeta now accepts optional tokens_in for analytics.
Anthropic native messages.stream() helpers are now instrumented for sync
and async clients, including event iteration, text_stream, and final-message
helpers, with the shared streaming lifecycle contract.
Process-wide integrations and hub patches now support sequential SDK client
replacement after the previous owner closes. Reinitialization replaces a
closed default. Existing models and provider clients keep their original
destination and must be recreated. Attempts to use a closed telemetry
destination warn once instead of silently dropping inference and error
events.

Changed

The llama.cpp Python integration is now named llama_cpp in Python, CLI,
and environment configuration. gguf remains a compatibility alias; all
selects the integration once under its new name. Model format metadata
remains gguf.
wildedge.init(dsn=...) reuses the existing default client when the DSN
points to the same project, for example dsn=os.environ["WILDEDGE_DSN"]
under wildedge run. It used to create a second client.
Enabling an integration or hub on a second client while another live client
owns process-wide instrumentation now logs a warning. Automatic events keep
going to the owner's project, as before, but this used to happen without a
word. With wildedge run --strict-integrations it raises RuntimeError.
Quantization labels are defined once in wildedge.quantizations and used
by every integration. Behaviour changes:
- MLX-LM reports N-bit weights as int4/int8, like Transformers, instead
  of q4/q8.
- GGUF models labelled from general.file_type report the type name (for
  example q4_k_m) instead of its number, and the i8 filename suffix
  reports int8.
- faster-whisper reports weight precision (int8, f16) instead of
  CTranslate2 compute types such as int8_float32.
The ultralytics integration now requires ultralytics 8.0.200 or later
(was 8.0.0). Earlier 8.0.x releases keep their helpers under
ultralytics.yolo.utils, and the integration turned itself off on them
without a word. instrument("ultralytics") now logs a warning naming the
installed version instead.

Fixed

Every error event now caps error_message at 500 characters, keeping the
start and the end of a longer message. Integrations used to cut it to the
first 200, which lost the cause in messages that end with it, such as
mlx-whisper audio errors that start with the ffmpeg banner. Manual
track_error() calls had no limit at all. provider_error_code is capped
at 64 characters for every event.
The llama.cpp integration now tracks Llama.create_chat_completion() and
Llama.create_completion(), streaming or not. It used to patch only
Llama.__call__, so chat apps on llama-cpp-python reported model loads but
no inference events. It now patches create_completion, which every entry
point runs, so each call is tracked once.
llama.cpp inference events now include the stop reason for non-streamed
calls, the generation config (temperature, top_p, top_k, max_tokens,
repeat_penalty, penalties, seed, stop sequence count), and, for chat
calls, prompt_type="chat" and turn_index measured on the last user
message. Prompts, chat messages and replies are never attached. As before,
attachments upload only when passed to track_inference(attachments=[...]).
import wildedge no longer imports every installed ML framework and API
client. Each library is imported when its integration is activated or a
model from it is registered, so SDK import is faster and a broken unrelated
package no longer stops the SDK from starting.
Changing a model's metadata (for example register_model(..., version="v2"))
now takes effect after a restart. With offline persistence on (the default),
the model registry saved by an earlier process kept the first metadata ever
registered for each model on that machine, so later changes were ignored.
This includes anthropic models already saved as "unknown".
MLX-LM models now report the device they run on (mps for the Apple GPU)
instead of always cpu. Accelerator names are defined once in
wildedge.accelerators, and PyTorch, Transformers, Ultralytics,
TensorFlow and Keras share their device detection.
timm models now use the requested architecture/variant name for model identity
instead of the Python class name, so models such as resnet18 and resnet50 no
longer share telemetry. Manual registration falls back to timm architecture
metadata, and explicit model IDs retain priority.
Ultralytics model.predict() now records inference telemetry. Instrumentation
patches predict, which model(...) delegates to, so both entry points are
tracked exactly once.
Streaming wrappers now finalize once across exhaustion, explicit closure,
cancellation, and iterator errors. Early closure/cancellation records an
unsuccessful inference with partial metadata when available. TTFT ignores
lifecycle/role-only chunks; telemetry callbacks cannot break iteration and
retain the trace context captured at stream creation. Remote API streams
ending without their completion marker record STREAM_INCOMPLETE.
A span attribute whose value is a list, dict or other non-scalar no longer
gets the whole event batch rejected. Ingest only accepts scalar attribute
values and answered 422 for the batch, dropping unrelated loads, inferences
and spans. Such values are now sent as JSON strings, with one warning per
key.
close(timeout=...) and flush(timeout=...) now upload buffered
attachments in the time left after sending events. close() used to stop
the attachment uploader without a final upload, so a short-lived process
lost the attachments of its last few seconds (the uploader runs every 5 s)
while its events still referenced them. close() without a timeout still
leaves them on disk for the next start.
ONNX models exported with the default main_graph graph name (newer
torch.onnx.export, Piper TTS voices) are now named after their file, so
different voices no longer show up as one model called main_graph. A
session created without an id falls back to the file name too, and
model_version 0, negative or INT64_MAX (the exporters' "unset") is sent as
unknown instead of 9223372036854775807.
transformers pipelines called with a string or a list record one inference
event per call instead of two. pipeline() loads the model before
Pipeline.__init__ runs, so the model was registered with its own forward
hooks, and each pipeline call was tracked by the pipeline and again by the
model inside it. That model's hooks now skip forward passes inside a
tracked call of its pipeline; other models the pipeline runs, and direct
calls to the model, are still tracked. Generator and dataset inputs, which
run lazily, are not covered yet (transformers pipelines with generator or dataset inputs record events before inference runs #78).
flush() no longer sends a batch the background sender is already sending.
Both used to send the same events and then each removed a batch from the
queue, so events were duplicated and, depending on timing, later events
could be removed without being sent. This showed up when a process started
with events saved by an earlier run and flushed soon after.
Enabling the huggingface hub no longer imports unused transformers model
modules. To find references to hf_hub_download and snapshot_download,
the hub tracker looked up the name on every loaded module, and transformers'
lazy modules import code on lookup. With transformers loaded this imported
about 140 extra modules, took about 0.15 s and logged about 430 "failed to
patch" debug lines. The tracker now reads each module's namespace instead.

Full Changelog: v0.2.2...v0.3.0

github-actions · 2026-10-21T17:50:19Z

Coverage report

Total: 90% (12173/13583 lines)

Top uncovered files

wildedge/integrations/transformers.py: 63.2% (111 lines missed)
wildedge/integrations/mlx.py: 43.0% (98 lines missed)
wildedge/integrations/ultralytics.py: 67.8% (84 lines missed)
wildedge/integrations/pytorch.py: 64.5% (70 lines missed)
wildedge/platforms/macos.py: 57.1% (67 lines missed)

The integrations cut error messages to ERROR_MSG_MAX_LEN and provider codes to PROVIDER_ERROR_CODE_MAX_LEN before calling track_error(). Manual calls sent the full text. ErrorEvent now applies both limits when serializing, so every error event gets them.

Measured across 37 typical failures, 9 messages were over 200 characters. Several put their cause at the end. mlx-whisper audio errors start with the ffmpeg banner, and torch state_dict errors end with the shapes. The integrations kept only the first 200 characters. ErrorEvent now keeps the start and the end of a message over 500 characters, with an ellipsis between. The integrations and decorators no longer cut the message themselves, so the limit is applied once.

The notes listed every commit subject since the previous tag, including merge commits and release housekeeping such as changelog moves. They are now the curated CHANGELOG.md section for the version. A missing or empty section fails the build instead of publishing empty notes.

Cap error fields in every error event and build release notes from CHANGELOG.md

0.2.x already uploaded only explicit attachments. Automatic GGUF capture was added and removed during 0.3.0 development and never shipped, so the entry described no change for users upgrading.

piotrekno1 and others added 30 commits September 30, 2026 12:07

Merge pull request #58 from wild-edge/release/0.2.2

0e3404a

Merge release/0.2.2 into devel

README: PyPI badge, links to the app, docs and blog, and where the DS…

1e8e19c

…N comes from Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Merge pull request #60 from wild-edge/docs/readme-dsn-links

53a5ff1

README: PyPI badge, links to the app, docs and blog, and where the DS…

Cover the gguf integration's fallback and error paths

f48992b

Pin setuptools for ultralytics 8.0.0 on Python 3.12 and 3.13

14a7ce2

ultralytics 8.0.0 imports pkg_resources without declaring setuptools, and Python 3.12+ environments no longer include it. The old compat test skipped here; with skips now failing, the full run caught it.

Run compat on pull requests into main and release branches only

0270e0c

Every devel pull request running the matrix is more than needed; devel gets the nightly run. Pull requests into main or a release branch run the full matrix, so the separate pr profile is gone.

Docs: the gguf compat model is downloaded

d74d1d3

Merge pull request #69 from wild-edge/ci/compat-current-pins

24a2202

Fix the compat matrix and make it test each integration

Merge remote-tracking branch 'origin/devel' into fix/gguf-track-creat…

97bd484

…e-completion

Make the gguf example a multi-turn chat

b88d048

The example only called llm(...). It now runs three streamed create_chat_completion turns and one create_completion, so it exercises the chat tracking from this PR.

Import inspect at module level in the gguf tests

aea50b1

Merge pull request #59 from wild-edge/fix/gguf-track-create-completion

3992cdf

Track llama-cpp chat and completion calls in the gguf integration

Cover the str() fallback for span attributes JSON can't encode

0b89bb6

Import the events common module at module level in the tests

b97136a

Merge pull request #66 from wild-edge/fix/span-attribute-scalars

9b15261

Send non-scalar span attributes as JSON strings

Import constants and WildEdge at module level in the new attachment t…

50fab2b

…ests

Merge pull request #67 from wild-edge/fix/close-uploads-attachments

5fb8f30

Upload buffered attachments on close and flush with a timeout

piotrekno1 added 20 commits October 21, 2026 13:11

Remove release documentation bundle note from README

9448483

Clarify README setup, integration activation and telemetry

68c8ff6

Make integration overview compact with expandable reference

c69008b

Move API-specific README details into provider documentation

3c9667d

Merge pull request #94 from wild-edge/docs/readme-local-quickstart

e9c6038

Improve README quickstart and add demo preview

Show compact trace and SQL previews with extra views collapsed

68580ce

Show computed sample SQL results as stacked model volumes

d98df57

Merge pull request #95 from wild-edge/docs/readme-trace-sql-gallery

ba7b637

Add compact trace and SQL previews to README

Merge branch 'main' of github.com:wild-edge/wildedge-python into devel

bf806c9

Fix whisper example span kind and add uv script header

d7fc6a6

Pin httpx 0.27 for the openai min compat rows

faa5f77

openai 1.30.0 passes proxies= to httpx, which httpx 0.28 removed.

Merge duplicate changelog groups and list instrumentation ownership a…

f1225f5

…s a change Drops the llama.cpp line about automatic prompt and reply attachments, which the explicit-attachments change removed.

Serialize consumer drains so flush() cannot resend an in-flight batch

cb23ef1

The fork hooks hold the drain lock across fork() so a child never inherits it locked.

Add streaming and whisper docs to llms.txt

0e24083

Release 0.3.0

d45c19a

Merge pull request #96 from wild-edge/release/0.3.0-fixes

397a9bd

Release 0.3.0

Scan module dicts when patching Hugging Face download functions

8adbaca

getattr on transformers' lazy modules imported unused model modules and logged a debug line for each one that failed to import.

Merge pull request #97 from wild-edge/fix/hf-hub-patch-lazy-imports

90cb8cd

Stop the Hugging Face hub patch from importing unused transformers modules

Move the Hugging Face hub fix into the 0.3.0 changelog

6c3a54e

piotrekno1 mentioned this pull request Oct 22, 2026

Cap error fields in every error event and build release notes from CHANGELOG.md #99

Merged

piotrekno1 added 5 commits October 22, 2026 15:57

Merge pull request #99 from wild-edge/fix/truncate-error-fields

b98d2b0

Cap error fields in every error event and build release notes from CHANGELOG.md

Correct the 0.3.0 attachments changelog entry

e8e70f9

0.2.x already uploaded only explicit attachments. Automatic GGUF capture was added and removed during 0.3.0 development and never shipped, so the entry described no change for users upgrading.

Date the 0.3.0 release 2026-10-24

61fb799

piotrekno1 merged commit 3de0b85 into main Oct 24, 2026
95 checks passed

Navigation Menu

Repository navigation

Release 0.3.0 - #98

Release 0.3.0#98
piotrekno1 merged 113 commits into
mainfrom
release/0.3.0

piotrekno1 commented Oct 21, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 21, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 21, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

piotrekno1 commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Verification on this branch

Uh oh!

github-actions Bot commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changelog preview for v0.3.0

Added

Changed

Fixed

Uh oh!

github-actions Bot commented Oct 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Coverage report

Top uncovered files

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

piotrekno1 commented Oct 21, 2026 •

edited

Loading

github-actions Bot commented Oct 21, 2026 •

edited

Loading

Changelog preview for `v0.3.0`

github-actions Bot commented Oct 21, 2026 •

edited

Loading