deepseek-ocr2-moe-experts

DeepSeek-OCR-2 MoE experts

Fused routed-experts MoE for DeepSeek-OCR-2 (deepseek-community/DeepSeek-OCR-2 text model): 64 routed experts, top-6, plain SiLU SwiGLU, router-weight scale + scatter-add. Routing (fp32 softmax + greedy top-6 scaled by routed_scaling_factor, DeepseekOcr2TextMoe.route_tokens_to_experts) and the 2 shared experts happen outside the op; both entry points consume (hidden_states, top_k_index, top_k_weights): deepseek_ocr2_moe_experts (grouped GEMM, prefill shapes) and deepseek_ocr2_moe_experts_decode (fused gather-GEMV for n_tokens ~ 1–4 — two persistent-shape kernels, no per-expert cuBLAS loop, no D2H sync). Constant-swap port of aymous/cohere2-moe-experts.

Built with kernel-builder as native torch.library ops (AOT, loadable via the kernels library). Reference op: transformers DeepseekOcr2TextExperts.

Benchmark (NVIDIA GB10 / sm_121, aarch64, bf16)

Mean latency per workload — op_eager / op_compile are the torch reference (eager and torch.compile); custom is this kernel. Hidden 1280, per-expert intermediate 896, 64 experts top-6.

Prefill shape — n_tokens = 2048, deepseek_ocr2_moe_experts: 4.261 ms, 2.26× vs op_compile, verified.

Decode shape — n_tokens = 1, deepseek_ocr2_moe_experts_decode: 0.191 ms, 4.35× vs op_compile, verified.

Usage

import torch
from kernels import get_kernel

kernel = get_kernel("aymous/deepseek-ocr2-moe-experts", version=1)
# gate_up_proj: [E, 2F, H], down_proj: [E, H, F]; top_k from the router (outside the op)
top_k = top_k_index.size(1)
# prefill (grouped GEMM):
out = kernel.deepseek_ocr2_moe_experts(
    hidden_states, gate_up_proj, down_proj,
    top_k_index.int(), top_k_weights.float(), top_k,
)
# decode (fused gather-GEMV, n_tokens ~ 1-4):
out = kernel.deepseek_ocr2_moe_experts_decode(
    hidden_states, gate_up_proj, down_proj,
    top_k_index.int(), top_k_weights.float(), top_k,
)

Note: the published build targets aarch64 / CUDA / sm_121 (NVIDIA GB10). Other platforms need the full build matrix.