Skip to content

ci: update container registry location - #3

Merged
saturley-hall merged 16 commits into
mainfrom
harrison/ci-test
Mar 4, 2025
Merged

saturley-hall merged 16 commits into
mainfrom
harrison/ci-test

Conversation

@saturley-hall

@saturley-hall saturley-hall commented Mar 4, 2025 •

Copy link
Copy Markdown
Member

What does the PR do?

Updates the container registry used for caching

Splits CodeQL and Copyright-Checks into separate workflows.

Update image that Copyright checks uses to a new location in GHCR.io

@github-actions

github-actions Bot commented Mar 4, 2025 •

Copy link
Copy Markdown
Contributor

Test Results

 2 files   2 suites   25s ⏱️
71 tests 71 ✅ 0 💤 0 ❌
89 runs  88 ✅ 1 💤 0 ❌

Results for commit c7118b9.

♻️ This comment has been updated with latest results.

jthomson04 added a commit that referenced this pull request Jun 4, 2026
The cache-realloc (#1) and metric-handle (#2/#3) commits left a few lines
unformatted — `longest_prefix_match`'s collapsed signature and the `merged`
binding in l1.rs, and the per-worker gauge assignment + a test assert in
metrics.rs. `cargo fmt --all --check` (run by the rust-tests CI job) flagged
them, failing the job. No behavior change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jthomson04 added a commit that referenced this pull request Jun 4, 2026
The cache-realloc (#1) and metric-handle (#2/#3) commits left a few lines
unformatted — `longest_prefix_match`'s collapsed signature and the `merged`
binding in l1.rs, and the per-worker gauge assignment + a test assert in
metrics.rs. `cargo fmt --all --check` (run by the rust-tests CI job) flagged
them, failing the job. No behavior change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
kangclzjc referenced this pull request in kangclzjc/dynamo Jun 4, 2026
… path (review #3)

On the orchestrator replay path (use_orchestrator=True) the regression
models were never installed: replay/main.py guarded the AIC benchmark-FPM
bootstrap behind ``adapter._sm is not None and not adapter._sm._is_easy``,
and ``_sm`` is None under use_orchestrator — so the whole block (incl. the
``load_benchmark_fpms`` calls) was skipped. ``get_regression`` then returned
None for the whole replay, the throughput regression stayed empty, and
orchestrator-replay scaling decisions diverged from PSM — contradicting the
adapter docstrings that claimed bootstrap_from_fpms→install_regressions was
wired.

Fix:
- engine_adapter: extract ``install_regressions_from_fpms`` (synchronous,
  builds the regressions from benchmark FPMs via the throwaway-PSM factory
  and installs them on the shared store — no plugin bootstrap).
  ``bootstrap_from_fpms`` now = install_regressions_from_fpms +
  bootstrap_plugins.
- replay_adapter: add path-agnostic ``install_benchmark_fpms`` — PSM path →
  ``PlannerStateMachine.load_benchmark_fpms``; orchestrator path →
  ``install_regressions_from_fpms`` (plugins were already bootstrapped at
  adapter construction, so this does NOT double-bootstrap). Corrected the
  ``_get_regression`` docstring to the real wiring.
- replay/main.py: guard is now ``not adapter._is_easy_mode()`` (path-
  agnostic) and the two ``adapter._sm.load_benchmark_fpms(...)`` calls become
  ``adapter.install_benchmark_fpms(...)``.

Test: install_benchmark_fpms on the orchestrator path makes
get_regression("agg") non-None (was None pre-fix).

840 planner tests pass (+1). NOTE: replay/main.py's end-to-end path can't be
exercised in this env (its Rust _core symbol predates the installed
extension); the edit is py_compile-validated and the adapter method it calls
is unit-tested.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jthomson04 added a commit that referenced this pull request Jun 4, 2026
The cache-realloc (#1) and metric-handle (#2/#3) commits left a few lines
unformatted — `longest_prefix_match`'s collapsed signature and the `merged`
binding in l1.rs, and the per-worker gauge assignment + a test assert in
metrics.rs. `cargo fmt --all --check` (run by the rust-tests CI job) flagged
them, failing the job. No behavior change.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: jthomson04 <jwillthomson19@gmail.com>
zhongdaor-nv added a commit that referenced this pull request Jun 8, 2026
During each cell's measurement window, take a few lightweight per-node
snapshots (background thread, foreground sub.wait() unperturbed) capturing the
event-plane's resource footprint so we can quantify the brokerless-ZMQ vs NATS
cost trade-off:

  - established TCP count via `ss -tn state established` (ZMQ direct mesh is
    O(p*s); NATS is O(p+s) to the broker) -- the headline connection-cost diff.
  - per-process CPU% / RSS / open-fd / nproc for the key classes
    (mocker, event_plane_bench_sub, nats-server, frontend, zmq_broker, loadgen)
    via /proc/<pid>/{stat,status} + ls /proc/<pid>/fd.
  - nats-server broker cost: active during NATS cells vs ~idle during ZMQ
    cells (ZMQ has no broker -- the saving to quantify).

New ressample.py is a stdlib-only host-level sampler (runs on the node HOST,
not inside pyxis, so its ss/`/proc` scan sees every enroot process + all node
TCP). bench.py fires it on every node via a new launcher.popen_host() (srun
with no --container-* flags), collects the echoed JSON snapshots, folds them
into compact summary stats, writes a per-cell *_res.json, and attaches the
summary to every summary.json row. Gated behind DYN_BENCH_RESSAMPLE (default
on); snapshot count via DYN_BENCH_RESSAMPLE_N (default 2).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
zhongdaor-nv added a commit that referenced this pull request Jun 8, 2026
The sampler snapshots are read back from the srun-streamed stdout logs, not a
follow-up `cat` srun; correct the docstring to match the implementation.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
zhongdaor-nv added a commit that referenced this pull request Jun 8, 2026
#3)

Three fixes found in the lego-c2 sanity run:

1. RSS read 0 for in-container processes. Viewed from the node host, enroot/
   pyxis processes report VmRSS=0 in /proc/<pid>/status (and 0 resident in
   statm); only /proc/<pid>/smaps_rollup `Rss:` reports the true resident set.
   Read smaps_rollup first, fall back to VmRSS for host procs.

2. nproc inflation. The `srun` client + slurmstepd that launch nats-server /
   event_plane_bench_sub carry the full `-- nats-server ...` command in their
   argv, so the cmdline-substring match counted them too (nats nproc=3 instead
   of 1, with 0 RSS dragging the sum). Skip launcher/wrapper cmdlines
   (srun/slurmstepd/bash -c/` -- `/the sampler itself) so each class counts
   only the real target process.

3. Worker-node samplers crashed with FileNotFoundError: the node0-created
   LOGDIR doesn't exist on worker nodes, so `open(--out)` died before the
   snapshot was emitted -> only node0 reported, losing the cross-node
   connection counts. Now echo the JSON to stdout FIRST (the orchestrator
   reads snapshots from the srun-streamed stdout, not the --out file), then
   makedirs + best-effort write the node-local --out copy.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
BenHamm added a commit that referenced this pull request Aug 20, 2026
- SPDX headers on synthesize-user-workload openai.yaml and all .codex files (#1)
- CODEOWNERS entries for .codex/, agents/, agent-docs/ (#2)
- lycheeignore the two agent-docs links that 404 until this PR merges (#3)
- .gitignore: remove local-dev leftovers that ignored two tracked skill
  dirs; keep an anchored /runs/ for EXP_ROOT artifacts (#4-7)
- vllm.md: scope the parallelism constraint to gpus_per_replica (#10)
- tensorrt-llm.md: drop engine-era/C++-executor KV fields and use_cuda_graph
  in favor of the surfaces real configs use (#11)
- vllm.md: normalize flag spellings to hyphenated form; hedge the
  unverifiable v0.18 scheduler-default claim (#12)
- kubernetes-recipe-workflow.md: define DEPLOY_ROOT before use (#14)

agent-reported label created in the repo (#13). Findings 8-9 (stop-request
record location) tracked separately pending a design decision.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tzulingk added a commit that referenced this pull request Sep 3, 2026
Follow-up to #12991 addressing the architecture review of scale_elastic_ep:

- #1 Outer timeout backstop: wrap the scale in asyncio.wait_for
  (DYN_SCALE_EP_TIMEOUT_S, default 600s). Without it, a scale that hangs
  past vLLM's own limits leaves _scale_ep_in_progress stuck True and wedges
  the endpoint forever. A timed-out scale is a wedged engine -> fail fast.
- #2 Warm the heavy 'ray' import off the event loop (asyncio.to_thread) so
  the first scale on a worker does not block the loop.
- #3 Document that the ray.util.state.list_nodes patch mutates a
  process-global (safe: serialised by _scale_ep_in_progress, restored in
  finally).
- #4 Document that an externally-cancelled scale (shutdown / client
  disconnect) intentionally does not force a restart.

Adds test_hung_grow_times_out_and_restarts.

Signed-off-by: Tzu-Ling <tzulingk@nvidia.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
copy-pr-bot Bot pushed a commit that referenced this pull request Oct 13, 2026
fix(operator): refine DGDR v1beta2 API contracts

This branch was previously deployed

1 inactive deployment
GITLAB — c7118b9f Deployed Mar 4, 2025 by saturley-hall
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants