chore(fork): prune NVIDIA-only CI and record the upstream base

[taufeeque9]

Written by Claude (on Taufeeque's behalf).

Sets this repo up as a FAR.AI fork: prunes the upstream CI that cannot run here and rolls out the fork-base manifest and tooling, mirroring nemo-gym#16 (which rolled out the pattern from megatron-bridge-internal#3).

Changes

Area	Change
Workflow prune	Removes all 34 GitHub workflows (plus workflows/config/): every one is NVIDIA-org automation — gated on github.repository == 'NVIDIA/Megatron-LM', or needing FW-CI-templates, copy-pr-bot, self-hosted runners, or org secrets (PAT, NVIDIA_INFERENCE_*, TWINE_PASSWORD, Slack webhooks). Also CODEOWNERS (names @NVIDIA/* teams that do not exist here) and copy-pr-bot.yaml.
GitLab CI	Left in place, deliberately. Megatron-LM's real test CI is GitLab (.gitlab-ci.yml, .gitlab/), which does not run on GitHub — it is inert here, and deleting it would only add conflict surface on every upstream sync. Same for the workflow helpers under .github/actions/ and .github/scripts/.
.fork-base.json	New manifest: upstream repo, branch, and base commit (d12f6c8c). Changes only when upstream is merged in.
tools/fork_base.py	--write / --check / --print. Dropped in unchanged from the nemo-rl copy.
tools/sync_upstream.sh	make sync-upstream — merges newer upstream onto a sync/upstream-<date>-<sha> branch and regenerates the manifest. Unchanged from the other forks.
.github/workflows/fork-base.yml	Runs --check on every PR and push to farai/main. Publishes the status context fork-base. Follows nemo-gym's variant: checkout pinned by SHA, no PAT (this repo has no private submodules).
Makefile, README.farai.md	make help lists the commands; the README carries the fork workflow and patch list. Upstream ships no Makefile, so this one is fork-only.

Out of scope

No unit-test-suite bring-up (nemo-gym#16's sharded-suite step has no analogue here): Megatron-LM's tests are GPU-bound, so there is nothing runnable on GitHub-hosted runners.

Relationship to #1

#1 (tf-at/moe-lora-stack-on-d12f6c8c) is the pending expert-LoRA GEMM stack + hybrid-recompute guard series and is listed as such in README.farai.md. This PR should merge first; once it does, #1's merge ref carries the manifest and the fork-base check applies to it.

Verification

• farai/main sits exactly on upstream commit d12f6c8c9aff51e166d872fd70151687a8e3f375 ("Add generic interface for SSM inference (Add generic interface for SSM inference NVIDIA/Megatron-LM#5382)"): git merge-base origin/farai/main upstream/main returns it, and merge-base --all returns exactly one commit.
• python3 tools/fork_base.py --check passes locally, in both plain and PR (--against origin/farai/main) modes — the latter reports the expected bootstrap note (target has no manifest yet).
• bash -n tools/sync_upstream.sh passes; fork-base.yml parses as YAML.

🤖 Generated with Claude Code