FlashRT QKV Cache RoPE

This package provides FlashRT QKV split, Q/K RMSNorm, and RoPE kernels for Hugging Face Kernel Hub.

It is intended for VLA, VLM, and video-model attention staging where standard attention kernels can handle the attention core, but the model still needs a fast pre-attention QKV postprocess path. It also includes single-token decode helpers for Q staging and direct KV cache writes, plus a sequence GQA cache-write path for static decoder loops.

Kernels

qkv_split_bf16: split equal-head packed BF16 QKV into preallocated attention-ready Q/K/V tensors without normalization or RoPE.
qkv_split_norm_rope_bf16: split packed BF16 QKV, RMSNorm Q/K, apply RoPE to Q/K, and write BF16 Q/K tensors.
qkv_split_bias_norm_rope_v_bf16: add packed QKV bias, RMSNorm Q/K, apply RoPE to Q/K, and materialize BF16 Q/K/V tensors.
qkv_split_bias_norm_rope_v_cat_bf16: same video path, but writes directly into preallocated joint Q/K/V workspaces.
qkv_split_joint3_cat_bf16: VLA-oriented path that fuses video/action/und QKV postprocess and writes one attention-ready joint Q/K/V workspace.
qkv_split_bias_rope_bf16: add packed GQA/equal-head QKV bias, split Q/K/V, apply split-half RoPE to Q/K, and write attention-ready BF16 outputs.
qkv_split_bias_rope_fp16: same input contract, with FP16 Q/K/V output for FP16 attention islands.
qkv_split_per_head_norm_rope_bf16: split packed GQA QKV, apply per-head Q/K RMSNorm and rotate-half RoPE, and write BF16 Q/K/V.
qkv_split_rope_kvcache_bf16: split GQA packed QKV, apply interleaved RoPE to Q/K, and write K/V into preallocated sequence caches.
decode_q_norm_rope_stage_bf16: RMSNorm Q, apply rotate-half RoPE, and write a decode Q staging buffer.
decode_k_norm_rope_kvwrite_bf16: RMSNorm K, apply rotate-half RoPE, and write one K/V cache slot.
decode_k_norm_rope_kvwrite_devpos_bf16: same KV write, but selects the cache slot from a CUDA int32 cur_pos tensor for graph-friendly decode.

The decode APIs are fixed to head_dim == 128 and use BF16 (64,) cos/sin vectors. Unsupported shapes are rejected at the Tensor binding layer. The GQA sequence API uses BF16 interleaved (seq_len, head_dim) RoPE rows and supports different Q and KV head counts. The fused bias+RoPE API accepts FP32 half-width or full-width RoPE tables and supports even head dimensions from 8 through 256.