Skip to content

[minor] fixed code formatting doc - #896

Merged
zhyncs merged 1 commit into
sgl-project:mainfrom
etched-ai:main
Aug 2, 2024
Merged

zhyncs merged 1 commit into
sgl-project:mainfrom
etched-ai:main

Conversation

@min-xu-et

Copy link
Copy Markdown
Contributor

Motivation

Improve doc.

Modification

  • the doc was missing code block quotes, fixed.

Checklist

  1. Ensure pre-commit pre-commit run --all-files or other linting tools are used to fix potential lint issues.
  2. Confirm that modifications are covered by complete unit tests. If not, please add more unit tests for correctness.
  3. Modify documentation as needed, such as docstrings or example tutorials.

- the doc was missing code block quotes
@zhyncs
zhyncs merged commit 9319cd1 into sgl-project:main Aug 2, 2024
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 26, 2026
…e-policy knob's value came from) into the flip train

# Conflicts:
#	python/sglang/srt/managers/phase_policy.py
efschu pushed a commit to efschu/htsglang that referenced this pull request Aug 26, 2026
The merge-regression pass runs ruff with the repo's OWN selection
(.pre-commit-config.yaml: --select=F401,F821,UP037), not ruff's defaults --
the default run reports 575 findings on these files and measures a gate
nobody enforces.

Under the enforced selection: 0 F401, 0 F821, 22 UP037. Nineteen were
already there at 0cd27d9, measured rather than assumed by running the
same selection against the base blobs of the same files; three are new,
all in phase_policy.py, all forward references introduced by sgl-project#889/sgl-project#893
(effective_pp_exit_term, stranded_decode_s, superseded_pp_bound_warning).

Fixed all 22 rather than the 3. Both files carry `from __future__ import
annotations`, so the quotes are inert and their removal is the autofix the
hook itself applies; fixing only the new three would leave the file split
between two conventions, and leaving all 22 hands the next toucher a
pre-commit failure they did not cause.

91 passed across the suites over both files (sgl-project#889, sgl-project#893, sgl-project#896, the arm
outcome suite, sgl-project#838 layout conformance) after the rewrite.

Codespell, for the record, is NOT clean and is not made clean here: 12
hits on train-touched files against 11 at the same base. The single new one
is "oder" in test_one_chunk_tp_prefill_887.py:8 -- inside a verbatim German
user directive the test quotes as its premise, the same class as the
"egal" quote that has sat at phase_policy.py since before this train. A
quote is not a typo, and rewriting one to please a dictionary would falsify
the record it exists to preserve.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 13, 2026
…e reading now name their input

(23a) THE BOUNCE STATE, BESIDE THE BOUNCE TERM.

The exchange seat measured THREE distinct states that all price 2.16 GiB:
(depth=1, comparing=True), (depth=2, comparing=False), and (depth=2 under the
pre-sgl-project#1335 expression that under-priced by one widest layer). The figure alone
cannot tell a correctly priced shallow arm from the underpricing.

I proved that the hard way and it is why this commit exists rather than being a
tidy-up. I read 2.16 off my own dry run, told the exchange seat the cause was
`depth=1`, and was wrong: my branch forked from 9c22254 and simply did not
carry his fix. It was his THIRD case. Same number, three causes, wrong one
picked -- two messages after he had written the three cases out for me. The
wrong attribution had already gone to the boot seat and the coordinator.

Every ARM line now carries `bounce_depth=`, `bounce_slots=`, `bounce_inject=`
beside `xchg_bounce=`, and `bounce_state=absent` when the arm pins no bounce at
all -- an unpriced state and a state priced at defaults must not share a
spelling, the rule the ratchet fields already obey. `slots` comes from
`xchg_bounce.assemble_slots`, the one authority (sgl-project#1335), never a second
expression. `inject_mode` is resolved in the launcher's own process, which is
exactly the term a desk run and a boot can disagree about -- printing it makes
the disagreement readable instead of hidden inside an equal number.

(23b) WHICH INPUT PRODUCED S_D.

`_derive_d_l2_budget` lets a measured ownership vector in the sizing record
OVERRIDE `--d-cap-rank-share`, silently. The boot seat read `S_D=1` off its own
sizing table while the ARM line said `S_D=2`. Effective and unprinted: sgl-project#896.

The line now reads `S_D=2 (sd_share=0.3750 source=record)` or `source=flag`.
A terms dict written before this commit prints `source=unknown`, NOT `flag`:
absence cannot testify, and guessing the flag would be the same error as
reading an absent `pids` field as an empty one ([22-fix5b], four commits ago).

BOTH ARE PRINT-ONLY. Neither moves an arm. `xchg_bounce_prov` stops at
`choose` and never enters `price` -- a term only the printer needs must not be
threaded through the pricing path, where a later reader would have to prove it
does not move a number. A test asserts it is absent from `price`'s signature.

MEASURED, serving form, form key 24eb56724d73, quiet box (cards 4/4/4):
  ARM S=1 M=2400 S_D=4 (sd_share=0.4000 source=flag): anchors=9.49 rings=12.83
  overhead=0.89 xchg_bounce=2.86 bounce_depth=2 bounce_slots=3
  bounce_inject=shadow host_weights=46.40 ratchet_per_flip=8.70 ...
2.86 on the same invocation that printed 2.16 before -- which is how the
misattribution surfaced at all.

TESTS (7, red-first on c73c361: 5 failed / 2 passed). The load-bearing one
is `test_two_states_with_the_SAME_price_are_distinguishable`: two differently
priced states with an equal term must produce different lines. That is the
property whose absence produced the wrong attribution, asserted directly rather
than via the fields that happen to implement it today.
ruff clean.

(cherry picked from commit 20ddcde)
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.

2 participants