Build miles as a pure wheel that imports FSDP and torchtitan without Megatron by HJSang · Pull Request #4308 · radixark/miles

HJSang · 2026-10-22T17:04:43Z

Stack (merge in order): #4308 → #4315 → #4316 → #4311 (needs radixark/miles-wheels#4).

First step toward pip install miles for the FSDP and torchtitan backends, without Megatron.

Changes

Pure wheel. Drop the custom bdist_wheel that tagged a pure-Python package cp3X-cp3X-manylinux*; it now builds one py3-none-any wheel.
Ship what was missing. The fixed chat templates (templates/*.jinja, read from beside the module) go in package_data. MANIFEST.in adds requirements.txt, without which the sdist could not build.
torchtitan extra. torchtitan==0.3.0 (the commit the image pins) plus torch>=2.13, which 0.3.0 needs but does not declare.
wheel-smoke workflow + tests/ci/wheel_smoke.py: build the sdist and wheel, install wheel[torchtitan] into a clean venv, hide Megatron / mbridge / TransformerEngine / apex, then import train.py, train_async.py, both backend actors, and the modules loaded by name at run time (the default rollout function, the session server), and parse their args.
.github/actions/resolve-dependency-refs. The CPU and CUDA workflows carried identical inline copies of the ci-sglang-pr: / release-lock.json / branch-default resolution. All three jobs now share one action. Both local actions (this one and install-megatron-bridge) load from github.sha so /rerun-test on an older PR head and release CI on an existing branch still find it.

Testing

wheel-smoke passes on GitHub-hosted Linux (torch 2.14.1). Negative checks: a planted import megatron.core fails it, and so does running from inside the checkout.
End to end on an H200 devbox with the wheel stack, no Megatron: Qwen3-0.6B GRPO trains on FSDP and on torchtitan (with Compute log probs with one plain-torch vocab-parallel cross entropy, without Megatron #4315 and Pass packed-document boundaries to HF attention so FSDP runs on FlashAttention-4 #4316).

Not in this PR

A hosted sglang-miles wheel and the exact sglang pin (Build the sglang-miles wheel hourly and keep every version #4311, then a follow-up).
Moving Megatron-only requirements (onnxscript, nvidia-resiliency-ext, torchft-nightly) to an extra; that touches every Dockerfile and the image fingerprint.

🤖 Generated with Claude Code

HJSang · 2026-10-22T17:04:48Z

@claude review always

HJSang · 2026-10-22T17:04:49Z

@codex review

chatgpt-codex-connector · 2026-10-22T17:05:06Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-22T17:09:30.413760Z	`1b10595`	Manual request

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" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1b1059589b

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 · 2026-10-22T17:09:27Z

+          if [ -f release-lock.json ]; then
+            ref=$(python3 -c "import json; print(json.load(open('release-lock.json'))['sglang_commit'])")
+          else
+            ref=sglang-miles


Honor the PR's SGLang override

When a coordinated Miles/SGLang change declares ci-sglang-pr: in the PR body, the existing CPU and CUDA workflows resolve that override before the release lock or branch default, but this new workflow always chooses only release-lock.json or sglang-miles. The wheel check therefore tests a different SGLang tree from the rest of CI and can either block a valid coordinated change or pass against code the PR does not target; reuse the canonical dependency-ref resolution here. .claude/rules/single-source-of-truth.mdL5-L6

Useful? React with 👍 / 👎.

Done in 8768466. Thanks, nice catch.

claude

Findings marked 🟡 are optional suggestions and need no follow-up push.

claude

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

HJSang · 2026-10-22T21:05:17Z

@claude review

claude

Code review found no issues

No high-confidence issues detected in this change.

claude

No new blocking issues. 1 optional suggestion (a nit or a note on pre-existing code) was found and not posted. Nothing in this review needs a push before merging.

claude

Code review found no issues

No high-confidence issues detected in this change.

…hout Megatron The wheel was tagged per interpreter and platform (cp312-cp312-manylinux1_x86_64) although Miles has no compiled code in it, so it is now a single py3-none-any wheel. The fixed chat templates that tito_tokenizer reads from next to the module were missing from the wheel, and the sdist could not build because it lacked requirements.txt; both now ship. The torchtitan extra installs the 0.3.0 release the Docker image pins by commit, with the torch>=2.13 floor validate_torchtitan_args already enforces at runtime. The wheel-smoke workflow installs the built wheel into a clean environment and runs tests/ci/wheel_smoke.py, which blocks Megatron, mbridge, TransformerEngine and apex on import, loads train.py and train_async.py, imports both backend actors and parses arguments for each. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

wheel-smoke picked only release-lock.json or sglang-miles, so a PR pairing with an SGLang PR through ci-sglang-pr: would smoke against the wrong tree. The resolution block the CPU and CUDA jobs carried as identical inline copies is now the resolve-dependency-refs action, and all three jobs use it. wheel_smoke.py now hides Megatron by setting sys.modules entries to None instead of a raising meta-path finder, so importlib.util.find_spec probes return None as they would for a package that is really absent. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

GitHub loads a local action from the checked-out tree, and the reusable CI workflows check out inputs.ref, which can predate the action: /rerun-test on an older PR head, or release CI on an existing release branch. Both now check out .github/actions at github.sha, the commit the workflow runs from. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The reasons they carried are in the PR description. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…he workflow's commit wheel_smoke imported the drivers and actors, but the default rollout function and the session server are loaded by name at run time. install-megatron-bridge had the same old-ref problem as resolve-dependency-refs. Found by an adversarial review of this PR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

guapisolo · 2026-10-27T21:16:30Z

+MEGATRON_ONLY = ("megatron", "mbridge", "transformer_engine", "apex")
+BACKENDS = ("fsdp", "torchtitan")
+# loaded by name at run time, so importing the drivers does not reach them
+RUNTIME_MODULES = ("miles.rollout.sglang_rollout", "miles.rollout.session.server")


cc cmt

[P2] RUNTIME_MODULES checks the legacy rollout, not the default one. miles.rollout.sglang_rollout is selected only when MILES_USE_LEGACY_ROLLOUT_V1=1 (resolve_rollout_function_paths in miles/utils/arguments.py). The default is InferenceRolloutFn, which imports inference_rollout_train inside _call_train and inference_rollout_eval inside _call_eval, so this smoke never imports either module. Nothing on that path imports Megatron at import time today, so this is a missing guard rather than a current crash. Impact: if a Megatron-only import lands in the default train or eval rollout path, wheel-smoke stays green and a wheel install fails at its first rollout. It also contradicts the PR body, which says the smoke covers "the default rollout function".

Suggested fix:

RUNTIME_MODULES = ( "miles.rollout.inference_rollout.inference_rollout_train", "miles.rollout.inference_rollout.inference_rollout_eval", "miles.rollout.session.server", )

Keep miles.rollout.sglang_rollout too if the legacy v1 path should stay covered.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

claude

Code review found no issues

No high-confidence issues detected in this change.

guapisolo

cc cmt

Approving #4308 at e6ded55. The review ran at 319eb84; the rebase onto main keeps those five commits unchanged, and the new commit raises python_requires to 3.11.

[P2] Still open, not blocking: tests/ci/wheel_smoke.py imports the legacy miles.rollout.sglang_rollout rather than the default inference_rollout_train / inference_rollout_eval, so a Megatron-only import added to the default rollout path would keep wheel-smoke green and fail a wheel install at its first rollout. Nothing on that path imports Megatron today. Details and a 3-line fix: #4308 (comment). It can land here or in a follow-up.

Verification: read-only review of this increment at 319eb84 (native review, a Codex cross-harness review, a code-quality check), a range-diff to this head, and a read of the new commit; wheel-smoke is green at this head. Not covered: training from the wheel alone.

chatgpt-codex-connector Bot reviewed Oct 22, 2026

View reviewed changes

claude Bot reviewed Oct 22, 2026

View reviewed changes

Comment thread .github/workflows/wheel-smoke.yml Outdated

Comment thread tests/ci/wheel_smoke.py Outdated

HJSang force-pushed the hjsang/pip-wheel branch from 8768466 to 453b573 Compare October 22, 2026 21:04

HJSang marked this pull request as ready for review October 22, 2026 21:04

HJSang requested review from guapisolo, yueming-yuan and yushengsu-thu as code owners October 22, 2026 21:04

claude Bot reviewed Oct 22, 2026

View reviewed changes

Comment thread .github/workflows/_run-ci.yml Outdated

claude Bot reviewed Oct 22, 2026

View reviewed changes

HJSang changed the title ~~Build Miles as a pure wheel and check FSDP/torchtitan run without Megatron~~ Build miles as a pure wheel that imports FSDP and torchtitan without Megatron Oct 22, 2026

claude Bot reviewed Oct 22, 2026

View reviewed changes

claude Bot reviewed Oct 26, 2026

View reviewed changes

HJSang and others added 5 commits October 27, 2026 10:43

Trim the comments in the pip-wheel change

394203b

The reasons they carried are in the PR description. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

guapisolo reviewed Oct 27, 2026

View reviewed changes

Require Python 3.11, the oldest sglang supports

e6ded55

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

HJSang force-pushed the hjsang/pip-wheel branch from 319eb84 to e6ded55 Compare October 27, 2026 21:34

claude Bot reviewed Oct 27, 2026

View reviewed changes

guapisolo approved these changes Oct 27, 2026

View reviewed changes

yueming-yuan approved these changes Oct 28, 2026

View reviewed changes

Conversation

HJSang commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Changes

Testing

Not in this PR

Uh oh!

HJSang commented Oct 22, 2026

Uh oh!

HJSang commented Oct 22, 2026

Uh oh!

chatgpt-codex-connector Bot commented Oct 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

HJSang Oct 22, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Claude Code Review

Uh oh!

HJSang commented Oct 22, 2026

Uh oh!

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

guapisolo Oct 27, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

guapisolo left a comment

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

HJSang commented Oct 22, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Oct 22, 2026 •

edited

Loading