fix(reference-script): recognize the `sglang serve` entrypoint by jiejingzhangamd · Pull Request #1938 · AMD-AGI/Hyperloom

jiejingzhangamd · 2026-10-16T07:05:18Z

Description: --reference-script matched only sglang.launch_server for SGLang. The SGLang cookbook (and SGLang's own warning, "'sglang serve' is the recommended entrypoint") use sglang serve, so a recipe copied from the cookbook produced no flags and the CLI exited 2. The marker list now includes sglang serve, and the start of the flags is found by one rule for every framework (_flags_start): after python3 -m <module>, or after serve when the token before it is the vllm / sglang CLI. That replaces the vLLM-only branch in _extract_server_args and removes its now-unread framework parameter. Without the shared rule, sglang and serve would have been read as positionals and sglang recorded as the model.
Linked issue(s): Fixes --reference-script does not recognize the 'sglang serve' entrypoint #1933
Tests: added test_parse_sglang_serve_entrypoint_as_the_cookbook_writes_it (multi-line cookbook form: flags kept, --host/--port/--model-path dropped, model and env lifted) and test_parse_serve_entrypoint_does_not_take_the_cli_name_for_the_model; both fail on main. The 9 test files that touch reference scripts: 449 passed. ruff clean; code-metrics PASSED.
Real recipe check: the GLM-5.2 MXFP4 MI355X cookbook recipe used in an 8 h run, in sglang serve form, parses to the same 30 arg tokens, SGLANG_OPT_USE_TOPK_V2=1 and model as the hand-converted python3 -m sglang.launch_server form (identical: True); on main the serve form yields 0 tokens.
Size/complexity triggers crossed: none. _extract_server_args cognitive complexity drops 54 → 40; the baseline entry is lowered accordingly.
If this simplifies or refactors: the per-framework prefix branch in _extract_server_args is replaced by one entrypoint rule; existing vLLM / -m recipes parse unchanged (existing tests).
Observable effect: an operator can pass an SGLang cookbook command (sglang serve ...) to --reference-script as-is. Before, it was rejected with "yields neither a flag nor an export" (or carried exports only) and had to be rewritten to python3 -m sglang.launch_server by hand.
Breaking changes: no
PR addresses single concern: yes
Root cause is upstream, ticket filed: no

🤖 Generated with Claude Code

The SGLang cookbook and SGLang itself now launch with `sglang serve`, but --reference-script only matched `sglang.launch_server`, so a recipe copied from the cookbook yielded no flags and the CLI exited 2. Find where flags start with one rule for every framework: after `python3 -m <module>`, or after `serve` when the token before it is the vllm or sglang CLI. This replaces the separate vLLM branch, drops the now-unread framework parameter of _extract_server_args, and keeps `sglang` from being taken as the positional model. Fixes #1933 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HktBcjPwjE7QgkA8MgKH6q

github-actions · 2026-10-16T07:06:29Z

Code metrics gate: PASSED

Judged: the 3 files this change touches. Files it does not touch never fail it; anything found there is listed under Outside this change for information.

Dimension	Limit	Baseline entries	vs base
Cyclomatic complexity	fails > 20	72	+0
Cognitive complexity	fails > 30	284	+0
Function length (lines)	fails > 80	449	+0
Nested block depth	fails > 5	27	+0
Module length (lines)	warns > 800, fails > 1200	53	+0
Duplicated lines per file	fails > 0	118	+0
Dead code (vulture)	fails > 0	2	+0

Check	Findings	Effect
Comments	0
English only	0
Production code imports test code	0
Repeated literals	0
A1 Private name of another package	0
A8 Flag argument	0
A9 Tuple return	0
A10 Raw vocabulary comparison	0
A11 Hidden global write	0
A12 Pure forwarder	0
A12 Single-implementation abstraction	0

97 findings in files this change does not touch

Outside this change (informational): 97

Unit	Dimension	Baseline	Now
`src/hyperloom/orchestrator/enablement/recipe/projections.py` `project_launch_evidence`	Cognitive complexity	absent	36
`src/hyperloom/orchestrator/state/shared_state.py` `resolve_graded_comparison`	Cognitive complexity	absent	34
`src/hyperloom_kb/collect/mapping.py` `_experience`	Cognitive complexity	absent	31
`src/hyperloom_kb/schema.py` `FieldDeclaration.__post_init__`	Cognitive complexity	absent	32
`src/kernelforge/cli/forge_loop_command.py` `_prepare_measurement_driver`	Cognitive complexity	absent	36
`src/kernelforge/cli/forge_loop_command.py` `_read_kb_warm_start`	Cognitive complexity	absent	40
`src/kernelforge/loop/runner.py` `IterationLoop._collect_session_candidate`	Cognitive complexity	absent	44
`src/kernelforge/loop/runner.py` `IterationLoop._rule_on_attempt`	Cognitive complexity	absent	38
`src/hyperloom_kb/tests/test_cli.py`	Duplicated lines per file	absent	18
`src/hyperloom_kb/tests/test_global_identity_and_pull.py`	Duplicated lines per file	absent	13
`src/hyperloom_kb/tests/test_local_state.py`	Duplicated lines per file	absent	18
`src/hyperloom/inference_optimizer/cli/__init__.py` `_reexport_session_environment`	Function length (lines)	absent	112
`src/hyperloom/inference_optimizer/cli/__init__.py` `_resolve_agent_lineup`	Function length (lines)	absent	89
`src/hyperloom/inference_optimizer/cli/__init__.py` `_resolve_run_topology`	Function length (lines)	absent	92
`src/hyperloom/inference_optimizer/cli/__init__.py` `_resume_session`	Function length (lines)	absent	171
`src/hyperloom/inference_optimizer/cli/__init__.py` `_seed_fresh_session`	Function length (lines)	absent	183
`src/hyperloom/inference_optimizer/multi_node/cli.py` `_restart_multinode`	Function length (lines)	absent	107
`src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py` `_build_parser`	Function length (lines)	absent	84
`src/hyperloom/inference_optimizer/multi_node/scripts/launch_multinode.py` `_spawn_rank_actors`	Function length (lines)	absent	90
`src/hyperloom_kb/http_service.py` `RequestHandler._dispatch`	Function length (lines)	absent	91
`src/hyperloom_kb/sync.py` `GlobalSync.pull`	Function length (lines)	absent	86
`src/kernelforge/cli/forge_loop_command.py` `_RunReporter.build_result`	Function length (lines)	absent	91
`src/kernelforge/cli/forge_loop_command.py` `_prepare_measurement_driver`	Function length (lines)	absent	140
`src/kernelforge/cli/forge_loop_command.py` `_read_kb_warm_start`	Function length (lines)	absent	154
`src/kernelforge/loop/knowledge_integration.py` `_measure_warmstart_field`	Function length (lines)	absent	83

...and 72 more; the job log lists every row.

Verdict computed by the gate scripts in /home/runner/work/_temp/code-metrics-gate, not by this tree's copy.

_{Thresholds and sources: [tool.hyperloom.code_metrics] in pyproject.toml. Tools: complexipy 8.0.1, jscpd 5.4.1, ruff 0.16.2, vulture 2.16.}

chaojhou · 2026-10-16T15:12:06Z

PR #1938 -- fix(reference-script): recognize the `sglang serve` entrypoint

What it does: --reference-script only matched sglang.launch_server for SGLang, so a recipe copied from the SGLang cookbook in its recommended sglang serve form yielded no flags and the CLI exited 2. sglang serve is now an entrypoint marker, and one rule (_flags_start) finds where the flags begin for every framework: after python3 -m <module>, or after serve when the token before it is the vllm / sglang CLI. This replaces the vLLM-only branch in _extract_server_args and removes its unread framework parameter; cognitive complexity drops 54 -> 40 and the baseline entry is lowered. Run against base and head, sglang serve ... and uv run sglang serve ... now parse to their flags and model, while vllm serve ..., a path-qualified vllm serve, python3 -m sglang.launch_server ... and ATOM recipes parse exactly as before.

Blocking issues: none

Checked: reference_script.py (_entrypoint_markers, _find_entrypoint_line, _flags_start, _extract_server_args, parse_reference_script); cli/bootstrap.py _resolve_reference_recipe and every consumer of the parsed model (SharedState.reference_model in the current_setting.sh render, enablement/lane.py:582); every caller of _extract_server_args; the new tests and the description | Ran: parse_reference_script from extracted base and head trees on vLLM, SGLang (-m and serve forms, path-qualified, uv run, docker run, numactl) and ATOM recipes, comparing flags and model; CI at bf66a4c (green) | Base: a3fc0ba | Head: bf66a4c

LGTM

jiejingzhangamd requested a review from a team as a code owner October 16, 2026 07:05

chaojhou merged commit dd6a92d into main Oct 16, 2026
40 checks passed

chaojhou deleted the fix/reference-script-sglang-serve branch October 16, 2026 15:12

Navigation Menu

Repository navigation

fix(reference-script): recognize the `sglang serve` entrypoint - #1938

fix(reference-script): recognize the `sglang serve` entrypoint#1938
chaojhou merged 1 commit into
mainfrom
fix/reference-script-sglang-serve

jiejingzhangamd commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

Outside this change (informational): 97

Uh oh!

chaojhou commented Oct 16, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jiejingzhangamd commented Oct 16, 2026

Uh oh!

github-actions Bot commented Oct 16, 2026

Code metrics gate: PASSED

Outside this change (informational): 97

Uh oh!

chaojhou commented Oct 16, 2026

PR #1938 -- fix(reference-script): recognize the sglang serve entrypoint

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

PR #1938 -- fix(reference-script): recognize the `sglang serve` entrypoint