[refactor] add hardware-specific installation instruction

[FrankLeeeee]

Motivation

SpecForge pinned a single CUDA torch in its base dependencies and kept a separate requirements-rocm.txt for AMD, so installs on other hardware needed manual workarounds and the docs had to spell out each variant by hand. SGLang 0.5.18 only ships CUDA 13 wheels, so the installation guidance also needed to say so clearly. This PR makes the install path hardware-aware in pyproject.toml and lets users pick their setup in the docs and get the right command.

Modifications

• Move torch out of the base dependencies into cuda and rocm extras. The cuda extra also pins sglang-kernel and mooncake-transfer-engine-cuda13, and [tool.uv.sources] routes each extra to its wheel index (cu130, rocm7.2).
• Fold openai into the base dependencies and remove the data extra; update the data preparation doc and the regenerate script hint accordingly.
• Delete requirements-rocm.txt. ROCm and Ascend NPU install with --no-deps inside the vendor stack, as the docs already recommended.
• Rewrite the installation guide around an interactive configurator (hardware, version, installer, extras) that generates the install command, with per-hardware notes below it. The landing-page selector is extended and reused for this.

the installation selector in doc:

Related Issues

Accuracy Test

Benchmark & Profiling

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://sgl-fru7574.slack.com/archives/C09784E3EN6 to discuss your PR.

[maocheng23]

LGTM, but have we tested this on different hardwares?

[curnane-lab]

LGTM, but have we tested this on different hardwares?

From the Ascend NPU side: I'll run the install flow on Ascend NPU 910C (A3 node) (driver + CANN + torch_npu, then pip install -e . --no-deps with SGLang 0.5.18 NPU build) and post the results here shortly

[FrankLeeeee]

@maocheng23 , CUDA works well. still testing on AMD and NPU.

[FrankLeeeee]

I have talked to @Fridge003 and SGLang is deprecating the support for CUDA 12.9 as well, so SpecForge can go directly with CUDA 13. Nonetheless, I will provide a CUDA 12.9 docker image for specforge since sglang still has it.

[curnane-lab]

NPU test results from the Ascend side (following up on my comment above and @maocheng23's question on hardware testing):

Tested on a 910C (A3) node, sglang_v0.5.18 NPU container, 16 NPUs (6 capture + 10 trainer), config qwen3.5-4b-dflash-disaggregated-npu.yaml.

With the container's vendor stack (torch 2.10.0+cpu + torch_npu 2.10.0), training comes up fine: capture servers, rollout workers and the 10-rank trainer all connect, epoch 1 is running and features are going into the mooncake store.

I also tried the latest torch_npu release (26.1.1, torch_npu 2.12.0.post2, paired with PyTorch 2.12.0). On CANN 9.0.0 this fails: triton_ascend 3.2.0's Ascend backend compiles npu_utils.cpp against RT_LIMIT_TYPE_SIMT_WARP_STACK_SIZE, which CANN 9.0.0's base.h has renamed to RT_LIMIT_TYPE_SIMT_DVG_WARP_STACK_SIZE, and the compile error takes down all 6 capture servers. Note this combo officially pairs with CANN 9.1.X, so I'll upgrade CANN and re-run — will report back here.

[FrankLeeeee]

thanks @curnane-lab

[maocheng23]

LGTM, thanks a lot!
But this PR is still a draft, shall we merge it after combining #843 ?