Codestin Search App

ousamabenyounes · 2026-08-12T03:34:54Z

Summary

detect whether the installed llama-server exposes --reasoning
use --reasoning on|off for enable_thinking on modern builds
retain --chat-template-kwargs for older builds, independent preserve_thinking state, and reasoning_effort-style templates
preserve inherited LLAMA_ARG_REASONING overrides from unsloth start

Test verification (RED → GREEN)

Rebased PR head with the original regression test only:

FAILED test_modern_launch_honors_reasoning_env_override[True-off]
assert '--reasoning' not in ['llama-server', '--reasoning', 'on']

Patched branch:

4 passed in 1.50s
review-fix diff coverage: 3/3 changed executable lines covered

Follow-up test-only guards from Apoze's reviewed commit were cherry-picked and kept credited. After rebasing onto current main, the rebased baseline needed the launch test double updated for current keyworded GPU/health probes before the file could run:

baseline ./run-ci.sh: 1 failed, 3 passed
final ./run-ci.sh: 6 passed

One current-main Python 3.13 sentinel drift was then corrected test-only after CI exposed it:

test_non_gguf_reload_settings.py::TestNonGgufStatusReportsWhatTheLoadAskedFor::test_the_non_gguf_status_branch_publishes_them
1 passed

One current-main repo CPU sentinel drift was also corrected test-only after CI exposed the model picker's centralized selectMeta load-id shape:

tests/studio/test_model_picker_contracts.py::test_a_pinned_cached_row_loads_from_the_id_the_backend_pinned
1 passed

Validation

./run-ci.sh upstream baseline: 296 backend tests + 6 CLI reasoning tests passed
./run-ci.sh patched branch: 300 backend tests + 6 CLI reasoning tests passed
./run-ci.sh after this follow-up commit: git diff --check, Ruff, and tests/test_launch_reasoning_args.py passed (6 passed)
selected adjacent launch/reasoning backend tests plus the Python 3.13 and repo CPU sentinel failures passed (545 passed)
Ruff and git diff --check: passed

mahiatlinux · 2026-08-12T06:29:29Z

@codex review

chatgpt-codex-connector · 2026-08-12T06:32:30Z

Codex Review: Didn't find any major issues. Keep them coming!

Reviewed commit: 94baeb68f7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b0c0af4b9e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

Apoze · 2026-08-23T20:56:57Z

Thank you for updating this PR. GitHub now reports the branch as mergeable on current main, and the implementation correctly covers the native flag, the old-binary fallback, the #9096 preserve default, and the inherited environment override.

A separately validated current-main implementation remains available as one commit here:

Apoze@502668e

It keeps the change generic and avoids regressions introduced since this PR was opened:

uses native --reasoning on|off when the installed llama-server supports it;
preserves the separately calculated _preserve_thinking_default added by Default preserve thinking on for Qwen3.8 #9096;
retains the existing --chat-template-kwargs fallback for older binaries;
leaves an inherited LLAMA_ARG_REASONING operator override in control.

Current-main validation: 9 focused launch-reasoning tests passed; 1,659 adjacent capability, placement, and launch tests passed; Ruff, the repository format script, and git diff --check passed. Five additional local Windows failures were platform-only (resource is Unix-only and one simulated Metal case reads the Windows host-memory guard), not reasoning failures.

Two regression guards from that commit are still worth carrying into this PR before merge:

set _preserve_thinking_default = True in a modern-path test and assert {"preserve_thinking": true} remains in --chat-template-kwargs; the current fixture always sets it to False, so it would not catch the original Default preserve thinking on for Qwen3.8 #9096 regression;
use reasoning_effort style with a modern-capable binary and assert it stays in --chat-template-kwargs rather than being translated to the boolean --reasoning flag.

I am intentionally not opening a competing PR while this one remains active.

ousamabenyounes · 2026-08-23T23:54:06Z

Addressed the inherited reasoning override on 148c8e5: modern launches now leave LLAMA_ARG_REASONING in control, preserve the separately computed preserve-thinking default, and retain the legacy kwargs fallback. Validation reproduced RED (1 failed, 2 passed), then focused GREEN (3 passed), full baseline/final 296 → 299 passed, 100% diff coverage, and final CI completed with 39 successes and 0 failures. Co-authored with Apoze; attribution is recorded in the commit trailer.

ousamabenyounes · 2026-08-24T06:36:59Z

Thank you @Apoze for taking the time to investigate this and share your fix.
I had been struggling to find a clean path forward on this PR, and your implementation was very helpful.
In particular, it identified the missing environment-override case and provided a clear, well-tested path to the final solution.

Incorporated the relevant fix and tests, and made sure your contribution was properly credited through the Co-authored-by trailer in commit 4348739.

Apoze · 2026-08-24T08:51:54Z

I added the two remaining test-only regression guards on top of the current head 148c8e5d43decbe780dd78f699fed769c52377d0:

Qwen3.8 preserve_thinking=True remains in --chat-template-kwargs on the modern --reasoning path.
reasoning_effort-style templates keep their launch default in --chat-template-kwargs and never convert it to --reasoning.

Commit: Apoze@e0d0066

Validation: 323 adjacent launch/llama-server reasoning tests passed; Ruff check and git diff --check passed. Independent review found no issues. The commit changes only studio/backend/tests/test_launch_reasoning_args.py and is ready to cherry-pick:

git cherry-pick e0d0066193353d5d5abbe796eb10f6eabcf07d5f

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 70a2843125

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 54142a3adf

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

ousamabenyounes · 2026-09-02T03:48:25Z

Follow-up pushed in e4c893eb081e493e6000621493b90f4e547474ff.

I verified Apoze's 2026-08-24 comment and test-only commit e0d0066193353d5d5abbe796eb10f6eabcf07d5f through the GitHub API, cherry-picked the two regression guards cleanly, and kept Apoze credited on that test commit with the requested co-author trailer.

CI then exposed two current-main source-sentinel drifts, both fixed as separate test-only commits. The latest push also addresses the three launcher review threads: explicit LLAMA_ARG_REASONING=on/off now updates the recorded reasoning_default, auto stays on the model-derived default path, and inconclusive probes no longer emit a conflicting fallback enable_thinking kwarg when an explicit env override is present.

Local validation:

RED for the new reviewer regressions: 3 failed, 5 passed before the launcher fix.
Final ./run-ci.sh: 8 passed.
Selected adjacent launch/reasoning/sentinel suite: 547 passed.

Remote PR checks are green on the latest SHA; the only non-pass rows are the expected skipped GPU/gate/file-edit/resume jobs.

danielhanchen · 2026-09-02T10:44:08Z

The launcher change itself is contained, but this branch also rewrites the sentinels in studio/backend/tests/test_non_gguf_reload_settings.py and tests/studio/test_model_picker_contracts.py, which loosens guards unrelated to reasoning. Could you split those two into their own change and keep this one to the llama-server reasoning flag?

for more information, see https://pre-commit.ci

Co-authored-by: Apoze <[email protected]>

for more information, see https://pre-commit.ci

Add regression coverage for preserving model thinking defaults on modern --reasoning launches and keeping reasoning_effort-style templates in chat-template kwargs. Adapt the launch test backend double to current main's keyworded health and GPU probes after rebasing. Co-authored-by: Apoze <[email protected]>

Keep LLAMA_ARG_REASONING=auto on the model-derived launch default path, synchronize explicit on/off overrides with reasoning_default, and avoid emitting fallback enable_thinking kwargs when an inconclusive probe leaves an explicit environment override as the only trusted launch control.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: eab442f62d

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

chatgpt-codex-connector · 2026-09-08T10:40:49Z

+        if explicit_env_reasoning is not None:
+            self._reasoning_default = explicit_env_reasoning == _REASONING_ON


Keep the recorded default aligned on old binaries

When the capability probe conclusively identifies an older binary (supports_reasoning_flag=False and mtp_probe_inconclusive=False) and LLAMA_ARG_REASONING is explicitly on or off, this assignment records the environment value even though the fallback below still emits the model-derived enable_thinking kwarg. Older managed binaries intentionally ignore that environment variable (unsloth_cli/commands/start.py:1271-1274), so the running template uses thinking_default while omitted Anthropic requests consult the contradictory reasoning_default at routes/inference.py:4239, potentially leaking genuine reasoning or misclassifying literal think markup. Only synchronize this property when the environment override can actually control the selected binary.

Useful? React with 👍 / 👎.

ousamabenyounes requested a review from danielhanchen as a code owner August 12, 2026 03:34

ousamabenyounes force-pushed the fix/issue-7526 branch from 94baeb6 to b0c0af4 Compare August 16, 2026 16:54

chatgpt-codex-connector Bot reviewed Aug 16, 2026

View reviewed changes

Comment thread studio/backend/core/inference/llama_cpp.py Outdated

ousamabenyounes force-pushed the fix/issue-7526 branch 2 times, most recently from a71f987 to 148c8e5 Compare August 23, 2026 22:53

ousamabenyounes force-pushed the fix/issue-7526 branch from 148c8e5 to 70a2843 Compare September 2, 2026 01:16

chatgpt-codex-connector Bot reviewed Sep 2, 2026

View reviewed changes

Comment thread studio/backend/core/inference/llama_cpp.py Outdated

Comment thread studio/backend/core/inference/llama_cpp.py Outdated

chatgpt-codex-connector Bot reviewed Sep 2, 2026

View reviewed changes

Comment thread studio/backend/core/inference/llama_cpp.py Outdated

danielhanchen added the refactor# label Sep 2, 2026

ousamabenyounes and others added 6 commits September 8, 2026 10:18

fix(studio): use reasoning launch flag (unslothai#7526)

247c989

[pre-commit.ci] auto fixes from pre-commit.com hooks

f0ca24b

for more information, see https://pre-commit.ci

fix(studio): honor reasoning environment override

d294b45

Co-authored-by: Apoze <[email protected]>

[pre-commit.ci] auto fixes from pre-commit.com hooks

041f6cf

for more information, see https://pre-commit.ci

ousamabenyounes force-pushed the fix/issue-7526 branch from e4c893e to eab442f Compare September 8, 2026 10:34

ousamabenyounes requested a review from oobabooga as a code owner September 8, 2026 10:34

chatgpt-codex-connector Bot reviewed Sep 8, 2026

View reviewed changes

danielhanchen added important# and removed refactor# labels Sep 8, 2026

		if explicit_env_reasoning is not None:
		self._reasoning_default = explicit_env_reasoning == _REASONING_ON

Uh oh!

Uh oh!

Conversation

ousamabenyounes commented Aug 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test verification (RED → GREEN)

Validation

Uh oh!

mahiatlinux commented Aug 12, 2026

Uh oh!

chatgpt-codex-connector Bot commented Aug 12, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Apoze commented Aug 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ousamabenyounes commented Aug 23, 2026

Uh oh!

ousamabenyounes commented Aug 24, 2026

Uh oh!

Apoze commented Aug 24, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

ousamabenyounes commented Sep 2, 2026

Uh oh!

danielhanchen commented Sep 2, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Sep 8, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

ousamabenyounes commented Aug 12, 2026 •

edited

Loading

Apoze commented Aug 23, 2026 •

edited

Loading