Skip to content

Docs: replace an unreproducible dsv4 bind-phase baseline - #2034

Merged
ChaoWao merged 1 commit into
hw-native-sys:mainfrom
ChaoWao:docs/dsv4-bind-phase-current-numbers
Aug 26, 2026
Merged

ChaoWao merged 1 commit into
hw-native-sys:mainfrom
ChaoWao:docs/dsv4-bind-phase-current-numbers

Conversation

@ChaoWao

@ChaoWao ChaoWao commented Aug 26, 2026

Copy link
Copy Markdown
Collaborator

Summary

docs/dfx/hbg-bind-phases.md corrects two rows of its pinned table — dsv4's
args and host_view_close — in a paragraph that tells the reader what to
expect on a current tree. Both of its figures were unusable for that:

  • The baseline was f830f13c3 plus a then-uncommitted child-memory change,
    so it cannot be reproduced from any commit in the tree.
  • host_view_close was measured before Fix: prefer staging views for host tensor access #1973 removed the halHostRegister
    side, leaving it an order of magnitude high. A reader A/B-ing against
    0.014–0.028 ms today reads a 10x regression where there is none.

Both now come from the merged tree at dcf7559e8, 12 binds over two ranks at
--rounds 6: args 0.036–0.075 ms, host_view_close 0.0012–0.0030 ms with
count=0 bytes=0. #2022 established that giving a phase's counters the span its
clock covers changes what the numbers mean and not their values, so these remain
the current figures and were not re-measured on top of it.

Peak host RSS is recorded beside them, because the rows being corrected are
per-byte costs over what a bind stages: 1.31 GiB across the whole process
tree under --skip-golden and 23.4 GiB when the fixture is streamed in,
against the ~45.5 GB per rank the pinned row cost. That also retires the local
note that dsv4 was SIGKILLed at --rounds 6 — both modes now complete six
rounds with exit 0.

The pinned table itself is untouched. It is anchored to 777d4171 on
purpose and says so ("For orientation, not thresholds … Re-measure rather than
trusting this table"), so refreshing its rows would work against its intent.
Only the paragraph that speaks for the current tree is changed.

Testing

  • markdownlint-cli2 --config tests/lint/.markdownlint.yaml — 0 errors
  • tests/lint/check_retired_names.py — clean
  • Figures measured on dcf7559e8 via the doc's own recipe, two arms,
    task-submit --device auto --device-num 2, 12 binds each

Docs-only; no code path is touched.

@coderabbitai

coderabbitai Bot commented Aug 26, 2026

Copy link
Copy Markdown

Warning

Review limit reached

Next included review available in 19 minutes.

View limit details

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1292deaf-00cf-4a4f-901d-cb44991faaf4

📥 Commits

Reviewing files that changed from the base of the PR and between f40cacf and 8c67082.

📒 Files selected for processing (1)
  • docs/dfx/hbg-bind-phases.md

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

The paragraph correcting dsv4's `args` and `host_view_close` rows cited
measurements taken on `f830f13c3` plus a then-uncommitted child-memory
change, so its baseline cannot be reproduced from any commit in the tree.
Its `host_view_close` figure was also measured before hw-native-sys#1973 removed the
`halHostRegister` side, leaving it an order of magnitude high: a reader
A/B-ing against 0.014-0.028 ms today would read a 10x regression where
there is none.

Both figures now come from the merged tree at `dcf7559e8`, 12 binds over
two ranks at `--rounds 6`: `args` 0.036-0.075 ms, `host_view_close`
0.0012-0.0030 ms with `count=0 bytes=0`. hw-native-sys#2022 established that giving a
phase's counters the span its clock covers changes what the numbers mean
and not their values, so those remain the current figures.

The same run's peak host RSS is recorded beside them, since the rows they
correct are per-byte costs over what a bind stages: 1.31 GiB across the
whole process tree under `--skip-golden`, 23.4 GiB when the fixture is
streamed in, against the ~45.5 GB per rank the pinned row cost.

The pinned table itself is untouched. It is anchored to `777d4171` on
purpose and says so.
@ChaoWao
ChaoWao merged commit ca9b6ad into hw-native-sys:main Aug 26, 2026
16 checks passed
@ChaoWao
ChaoWao deleted the docs/dsv4-bind-phase-current-numbers branch August 26, 2026 12:33
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.

1 participant