Repository navigation
Conversation
c4da2fd to
4e7730c
Compare
4e7730c to
1b8e14a
Compare
1b8e14a to
c3d17a7
Compare
3f2c787 to
4c9c948
Compare
5c8ca12 to
03863ab
Compare
049a6b5 to
7a7ca65
Compare
caa16c6 to
ded29b8
Compare
ded29b8 to
aa39dd9
Compare
aa39dd9 to
caad901
Compare
caad901 to
dfc51ef
Compare
dfc51ef to
425cf2d
Compare
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 425cf2d. Configure here.
| stream.seek(offset) | ||
| while chunk := stream.read(65536): | ||
| output.write(chunk) | ||
| offset = stream.tell() |
There was a problem hiding this comment.
Log streaming reprints on each poll
Medium Severity
stream_log stores a text-mode tell() cookie in offset and later compares it to stat().st_size. That cookie is not a byte position, and progress-bar \r or a partial UTF-8 sequence at EOF makes it much larger than the file, so the next poll rewinds to the start and reprints the whole log. LogOutputStreamer already tracks a binary byte offset for this.
Reviewed by Cursor Bugbot for commit 425cf2d. Configure here.
425cf2d to
379f11e
Compare
379f11e to
e645b57
Compare
e645b57 to
e42ae4b
Compare
|
Closing because these failure-log and runtime-version changes were merged upstream in NVIDIA#476. |


Changes
The log-reading issue was identified during MI355X DeepSeek V4 SGLang validation: after a decode worker failed, strict UTF-8 decoding prevented printing its AITER log tail. This fixes failure reporting, not the underlying worker failure.
Upstream #476 and fork #14 contain the same changes. The fork PR sits on the branch-based AMD/MoRI/ATOM stack. Those three runtime layers have already merged upstream.