Add empty cpu/gpu/training extras_require slots (Phase 4)

[Shi-Dong]

Summary

• Add three empty placeholder slots to setup.py's extras_require: cpu, gpu, training. Each one is documented inline so the intent is obvious to future contributors.
• No runtime change yet (slots are empty lists). The full categorization is in the audit table below; populating the lists is Phase 6's job, after the patched-fork wheels exist (Phase 5).

Why

The pip install miles roadmap (Phase 4) calls for declaring the extras_require taxonomy now so:

1. Future commits that populate it land in well-named slots without bikeshedding.
2. A user can already type pip install miles[gpu] and not get a KeyError on the typo — they just get the install_requires baseline (same as no extras), and the dependency list grows transparently as Phases 5–6 land.
3. The setup.py inline comments serve as the single source of truth for "what goes where", removing the ambiguity that scattered comments across Dockerfile / requirements.txt / docs would create.

Phase 4 audit: where each current install lands

Already in install_requires (from requirements.txt)

20 packages — these are the always-installed dependencies of miles. All install cleanly from PyPI without a CUDA toolchain.

Package	Notes
accelerate	Brings in torch transitively (key for the eventual CPU path).
blobfile, datasets, httpx[http2], mcp[cli], memray, omegaconf, pillow, pybase64, pylatexenc, pytest-asyncio, pyyaml, qwen_vl_utils, tensorboard, transformers, wandb	Standard PyPI installs.
ray[default]	Pure-Python; runs on CPU and GPU.
ring_flash_attn; platform_system == "Linux"	Conditionally Linux-only. Importable on CPU but exercises GPU code paths.
sglang-router>=0.2.3	Pure-Python router; today comes from a radixark/sgl-router-for-miles pre-built wheel in the image, also published to PyPI.
torchft-nightly==2026.4.3; platform_system == "Linux"	Linux-only nightly.

extras_require["gpu"] — heavy GPU stack (Phase 6 target)

Today installed by docker/Dockerfile, not setup.py. Each needs CUDA toolchain at install time, a matching torch version, or a pre-built wheel.

Source	Notes
flash-attn	Pre-built wheel from yueming-yuan/miles-wheels release. Build-from-source is multi-hour.
flash-attn-3 (hopper)	Hopper-only (sm_90a). Pre-built wheel + sources flash_attn_interface.py from Dao-AILab/flash-attention raw.
flash-linear-attention==0.4.2	PyPI.
tilelang	Nightly from https://tile-ai.github.io/whl/nightly/cu128/.
causal-conv1d==1.6.1	PyPI, CUDA source build.
mamba-ssm==2.3.1	PyPI, CUDA source build.
transformer_engine[pytorch]==2.10.0 / transformer_engine==2.12.0 + transformer_engine_cu13	PyPI, CUDA build. ENABLE_CUDA_13 branch in Dockerfile picks variants.
apex	Pre-built wheel; from-source typically requires manual NVCC env.
nvidia-modelopt[torch]>=0.37.0	PyPI, CUDA.
torch_memory_saver (fzyzcjy/torch_memory_saver @ pinned SHA)	git+url today; eventually point at a stable wheel or pin via a Phase-3-style build arg.
mbridge (ISEEKYAN/mbridge @ pinned SHA)	Same situation as torch_memory_saver.
nvidia-cudnn-cu12==9.16.0.29 or cu13	PyPI. CUDA-version-specific.
numpy<2	Post-install pin needed for Megatron compatibility. Belongs in install_requires once we figure out the proper version constraint.

extras_require["training"] — full training stack (Phase 5 + 6 target)

Everything in gpu plus the patched forks. These don't exist on PyPI yet; Phase 5 publishes them to a private index.

Source	Notes
radixark-sglang (from third_party/sglang)	Published from the submodule on submodule pointer bump. Eventual name TBD.
radixark-megatron (from third_party/Megatron-LM)	Published from the submodule.
radixark-megatron-bridge (Phase 3-pinned SHA)	Currently pip install git+...@<sha> in Dockerfile; Phase 6 may switch to a published wheel for symmetry.

Docker-only / vendor-built / system-level (not exposed via pip)

What	Why we leave it alone
apt install nvtop rsync dnsutils ethtool	System-level diagnostics, not Python deps.
nccl-tests (built from NVIDIA/nccl-tests source)	System binary in /usr/local/bin/.
int4_qat (fake_int4_quant_cuda wheel)	Custom-built wheel; users get it from the image.
sgl-model-gateway binary	Rust-built native binary. Goes in /usr/local/bin/.
sglang_router wheel	Has both a pre-built wheel path and a Rust-source path. The PyPI version is in install_requires; the image-specific wheel handles edge cases.

extras_require["cpu"]

Currently empty. Likely stays empty or gets minor additions for things like docstring-rendering, pretty CLI, etc. that aren't worth bloating install_requires for. Will revisit in Phase 6 once we know what the CPU-only entry point needs.

Test plan

• python3 -c 'import ast; ast.parse(open("setup.py").read())' — syntax OK.
• On a fresh python:3.11 container: pip install miles[cpu] — should succeed (against an empty list — equivalent to pip install miles for now).
• This PR is intentionally a no-op for runtime; the meaningful tests come once Phase 5 lands and Phase 6 fills the slots.

Stacked on

radixark/miles PR #1194 (Phase 3: pin Megatron-Bridge SHA). Once that merges, this PR's base auto-updates.

What's NOT in this PR

• No new extras_require entries with actual packages — that's Phase 6 (after the patched-fork wheels exist).
• No CI / Dockerfile changes — those depend on Phase 6 too.
• No new doc files — keeping the audit in the PR description rather than committing a .md to the repo.

[gemini-code-assist]

Code Review

This pull request adds placeholder extra requirements (cpu, gpu, and training) to setup.py as part of the roadmap for pip install miles. These placeholders are currently empty and will be populated in future phases. There are no review comments, and I have no feedback to provide.

[Shi-Dong]

Closing as obsolete. This PR is superseded by a thin packaging slice cut fresh off current `main`:

• radixark/miles#PRA — re-add `third_party/sglang` + `third_party/Megatron-LM` submodules
• radixark/miles#PRB — `setup.py` bundling + `publish-pypi.yml` + `miles-rl` 0.0.3

Why this one is no longer needed: see the per-PR note below. The branch is left in place; nothing is lost.