NVFP4 4over6 CUDA 13 ARM64 (436719d)

ARM64 / Python 3.12 / CUDA 13 / torch 2.13.0 wheel built locally on GB200 from 436719d. Contains flash_moe, mxfp8_moe, and nvfp4_moe.

Adds optional four_over_six=True to NVFP4 weight/activation quantizers and grouped GEMM. Defaults to False. Enabled arithmetic matches FlashInfer's default 4/6 recipe: 448 normalization, MAE, strict error scoring and default approximate candidate arithmetic.

SHA256: e8311b13da602def10c4cc88ac6abad2c8d5090799d04a14a3c89421257a0d68

The installed wheel passed 8 numerical tests, 15 exact comparisons against FlashInfer/vLLM defaults, compiled forward/backward equality, and a Prime-RL MoE forward/backward smoke with non-zero router/expert gradients. All registered kernels report available. No GitHub Actions build was used.

NVFP4 CUDA 13 (ec62fa7)

ARM64 CUDA 13 / PyTorch 2.13 kernel wheel built locally on GB200 from ec62fa7c8ce1885a49a21f5ab53b260399105703.

The NVFP4 operator is prime_rl::grouped_nvfp4_gemm; quantization and dequantization helpers also use the prime_rl namespace with NVFP4-qualified names. The wheel includes the existing Flash MoE and MXFP8 kernels.

Installed-wheel GPU numerical validation: 4 passed. The registered operator schema was checked directly. No GitHub Actions build was used.

SHA256: b31d69270f06ba0238e202b2316427cca10aca2576b6fddf84f61cba223c9bcd.

NVFP4 GB200 development wheels (e6b525a)

NVFP4 development wheel built locally on NVIDIA GB200 (aarch64), CUDA 13.0.88 and PyTorch 2.13.0+cu130.

Source: e6b525a on feat/nvfp4-moe. Includes the MXFP8 runtime required by current Prime-RL and the NVFP4 grouped GEMM port. Build dependencies are declared per kernel in kernels.toml.

Validation: four NVFP4 GPU numerical tests passed against the installed wheel. Both backward recipes match the intended BF16 operands. This release contains the ARM64 wheel for the GB200 integration run.

NVFP4 CUDA 13: traceable grouped GEMM

Built locally on GB200 with CUDA 13 and PyTorch 2.13 from prime-kernels d9a5ceb. Removes torch.compiler.disable from the NVFP4 grouped-GEMM wrapper.

Validation: compiled forward, dgrad, and wgrad match eager exactly for both backward modes; existing GPU numerical tests: 4 passed.

The wheel includes Flash MoE, MXFP8 and NVFP4 kernels. CUDA computation is unchanged. No GitHub Actions build was used.

SHA256: dc73e91ce1ae0dd0174618cb89c4a603b11232dce02a4978d81ac7e979bcb350.

NVFP4 CUDA 13 (564904f)

ARM64 CUDA 13 / PyTorch 2.13 kernel wheel built locally on GB200 from 564904f7c1755237e5a97f3f5378d57cfe1736cf.

The NVFP4 operator is prime_rl::grouped_nvfp4_gemm; quantization and dequantization helpers also use the prime_rl namespace with NVFP4-qualified names. The wheel includes the existing Flash MoE and MXFP8 kernels.

Installed-wheel GPU numerical validation: 4 passed. The registered operator schema was checked directly. No GitHub Actions build was used.

SHA256: 7bb6d4db14cb599de97514c78b74aa72de53d77ebd801232f3e1b461429cee10.

v0.1.0-5405233

Kernel wheels built from 5405233.

v0.1.1

Kernel wheels built from 1abf4c7.

v0.1.0-5d17dda

Kernel wheels built from 5d17dda.

v0.1.0-1abf4c7

Kernel wheels built from 1abf4c7.

v0.1.0-6128055

Kernel wheels built from 6128055.