Skip to content

Follow-up to #28: patch 0003 subject and README speaker tags - #38

Merged
plotarmordev merged 2 commits into
mainfrom
fix/pr28-followups-clean
Nov 1, 2026
Merged

plotarmordev merged 2 commits into
mainfrom
fix/pr28-followups-clean

Conversation

@plotarmordev

Copy link
Copy Markdown
Owner

This replaces #30 with the same content and cleaner commit text. The net diff against main is byte-identical to the reviewed head of #30 (bfe684aa): git diff origin/main..HEAD | sha256sum here and git diff origin/main...bfe684aa | sha256sum both give 99ebebd99f845c821ff4e8ca2182e5f7a67ee073ae0b8a0d0477e762eb924729.

I rebuilt it as two commits on current main so the history that lands has accurate first-person messages, without the merge commits and without a message describing a README hunk and a patch that are no longer in the diff.

What it changes (3 files, +6/-6):

  • engine/nvfp4/patches/0003-dialogue-speakers.patch: I give the patch the series' server: subject prefix. Only the mail header changes; the patched code is the same.
  • engine/nvfp4/patches/SHA256SUMS: I update the 0003 hash to match.
  • README.md: I list speaker tags 0 to 15 beside per-request sampling as features that need no launch flag, carried by CUDA NVFP4 0002-0003 and Metal NVFP4 0004-0006.

Follow-up to #28 by @krunkosaurus.

Checks I ran at this head:

  • sha256sum -c SHA256SUMS in engine/nvfp4/patches: 4/4 OK
  • sha256sum -c SHA256SUMS in engine/nvfp4/macos/patches: 6/6 OK
  • python3 -m unittest discover -s recipe -p 'test_*.py': 51 tests OK (15 skipped)
  • python3 bench/stream_check.py --self-test: passed

No GPU check is needed: the diff touches only README text and a patch mail header; the patched code is unchanged.

I change the patch 0003 mail header subject from "Drama speech:
preserve dialogue speakers across long-form requests" to "server:
preserve dialogue speakers across long-form requests" so it matches
the "server:" prefix used by the rest of the CUDA NVFP4 series. Only
the Subject header changes; the patched code is untouched.

I update the SHA256SUMS line for 0003-dialogue-speakers.patch to the
new hash.

Follow-up to #28 by @krunkosaurus.
The README's full-model paragraph named only per-request sampling
(`temperature`, `top_p`, `top_k`, `seed`) as needing no launch flag.
I add speaker tags 0 to 15 in request text to that list, and I widen
the carrying series: the CUDA NVFP4 series grows from patch 0002 to
0002, 0003 and the Metal NVFP4 series from 0004, 0005 to 0004-0006,
so both now carry the same two request-level features. The sentence
saying each feature's notes are in the engine tree's `CHANGELOG.md`
or `docs/` keeps its wording and only rewraps onto its own line.
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