Add several features.

[jiayus-nvidia]

1. Add support for arbitrary mask.
2. Add support for paged kv in Ampere.
3. Add support for fp8 bwd with 5 quantization modes in Hopper.
4. Add support for all fp16/bf16 masks in fp8 fwd and bwd.

[netlify]

✅ Deploy Preview for pytorch-fbgemm-docs ready!

Name	Link
🔨 Latest commit	e01548d
🔍 Latest deploy log	https://app.netlify.com/projects/pytorch-fbgemm-docs/deploys/68ef3d48857cfa0008c49811
😎 Deploy Preview	https://deploy-preview-4998--pytorch-fbgemm-docs.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[meta-codesync]

@q10 has imported this pull request. If you are a Meta employee, you can view this in D84501311.

[q10]

Hi @jiayus-nvidia the default values parsed in generate_kernels.py appear to be broken, I see

raise ValueError("ARBITRARY_NFUNC must be odd")

when I run the script. Could you update the script so that reasonable defaults are applied when no environment variables are set?

[jiayus-nvidia]

Hi @q10, sorry for not considering this situation. Fixed now.

[q10]

Hi @jiayus-nvidia it appears there are some undefined symbols:

E   OSError: /home/ec2-user/miniconda/envs/build_binary/lib/python3.13/site-packages/fbgemm_gpu/experimental/hstu/fbgemm_gpu_experimental_hstu.so: undefined symbol: _Z13run_hstu_bwd_ILi90EN7cutlass12float_e4m3_tELi128ELb1ELb0ELb0ELb0ELb0ELb0ELb1ELi0EEvR15Hstu_bwd_paramsP11CUstream_st

Maybe the code generation step didn't generate all the template instantiations?

[jiayus-nvidia]

Hi @q10, you're right. That's because of the mismatch of instantiations generated and kernels called in main. Now it's ok to compile with no environment variables and run test.