sgl-flash-attn3: build against the Torch stable ABI

[moon-bot-app]

sgl-flash-attn3 is currently built against a specific Torch version ([torch] src = [...torch-ext/torch_binding.cpp...], no stable-abi). SGLang has moved to Torch 2.11, which we are deprecating in the kernel builder (see huggingface/kernels#732, this discussion). Rather than keeping a legacy Torch version alive just for this kernel, it would be better to switch it to the Torch stable ABI, like we already do for flash-attn3 and vllm-flash-attn3 — then a single build works across Torch versions.

The upstream kernel already has everything needed: it is in sync with the upstream stable ABI sources, cf. hopper/flash_api_stable.cpp.

So this should mostly be build configuration changes, mirroring flash-attn3:

• In sgl-flash-attn3/build.toml, set stable-abi in the [torch] section and point src at flash_api_stable.cpp instead of the torch-ext/ bindings + shim.
• Adjust sgl-flash-attn3/flake.nix if needed (cf. flash-attn3/flake.nix).
• Verify that the op signatures exposed by flash_api_stable.cpp stay in sync with flash_api.cpp modulo stable API/ABI differences, so the Python-side API does not change for SGLang users.

cc @danieldk @sayakpaul

---

Requested by Daniël de Kok - Slack thread - Agent trace

[danieldk]

CC @adarshxs

[nathon-lee]

@danieldk @sayakpaul May I take this issue on? I’m happy to help with it.

[sayakpaul]

@adarshxs did you already start it?

[adarshxs]

Yes here: #1054