flashinfer.fused_moe.trtllm_bf16_moe

flashinfer.fused_moe.trtllm_bf16_moe(routing_logits: Tensor, routing_bias: Tensor | None, hidden_states: Tensor, gemm1_weights: Tensor, gemm2_weights: Tensor, num_experts: int, top_k: int, n_group: int | None, topk_group: int | None, intermediate_size: int, local_expert_offset: int, local_num_experts: int, routed_scaling_factor: float | None = None, routing_method_type: int = 0, use_shuffled_weight: bool = True, weight_layout: int = WeightLayout.BlockMajorK, do_finalize: bool = True, enable_pdl: bool = True, tune_max_num_tokens: int = 8192, activation_type: int = 3, norm_topk_prob: bool = True, routing_replay_out: Tensor | None = None) → List[Tensor] | Tensor

BF16 MoE operation with autotuning support.

Implements a bfloat16 Mixture of Experts layer using the TensorRT-LLM backend with automatic performance tuning for optimal tile-size selection.

Parameters:

  • routing_logits (torch.Tensor) – [seq_len, num_experts] tensor of routing logits. float32 or bfloat16.

  • routing_bias (Optional[torch.Tensor]) – Optional [num_experts] tensor of routing bias. Must be bfloat16 if provided.

  • hidden_states (torch.Tensor) – [seq_len, hidden_size] tensor of input hidden states. Must be bfloat16.

  • gemm1_weights (torch.Tensor) – [num_experts, M // 128, hidden_size // 128, 128] first-layer weights, bfloat16. M equals 2 * intermediate_size for gated activations and intermediate_size for non-gated activations.

  • gemm2_weights (torch.Tensor) – [num_experts, hidden_size // 128, intermediate_size, 128] second-layer weights, bfloat16.

  • num_experts (int) – Total number of experts.

  • top_k (int) – Number of experts to route to per token.

  • n_group (Optional[int]) – Number of expert groups.

  • topk_group (Optional[int]) – Number of groups to consider for top-k routing.

  • intermediate_size (int) – Size of the intermediate layer.

  • local_expert_offset (int) – Offset of local experts in the global expert space.

  • local_num_experts (int) – Number of experts handled by this device.

  • routed_scaling_factor (Optional[float]) – Scaling factor for routing (may be None for some methods).

  • routing_method_type (int) –

    Routing method (default 0). Selects the routing-kernel pipeline; matches flashinfer.tllm_enums.RoutingMethodType.

    • 0 Default — Softmax → TopK.

    • 1 Renormalize — TopK → Softmax.

    • 2 DeepSeekV3 — Sigmoid → RoutingBiasAdd → Top-2 in group → Top-topk_group groups → Top-top_k experts from the selected groups.

    • 3 Llama4 — Top-1 → Sigmoid.

    • 4 RenormalizeNaive — Softmax → TopK → Renormalize (Qwen3 style).

    • 5 TopK — TopK only (no softmax/sigmoid).

    • 6 SigmoidRenorm — Sigmoid → TopK → Renormalize (divide by the sum of the top-K weights).

    • 7 MiniMax2 — Sigmoid + Bias → TopK → ScaledSumNormalize (routeScale = 1.0, epsilon = 1e-20).

    • 8 Sigmoid — Sigmoid → TopK (no renormalization).

    • 9 Unspecified — reserved.

  • use_shuffled_weight (bool) – Whether to use the shuffled weight layout (default True).

  • weight_layout (int) –

    Weight layout for gemm1_weights / gemm2_weights; matches flashinfer.tllm_enums.WeightLayout. This BF16 MoE entry point requires BlockMajorK — passing any other value raises a runtime error. Default WeightLayout.BlockMajorK.

    • 0 MajorK — K-major, logical shape [Mn, K]. Not supported by this function.

    • 1 MajorMn — M-major (A) / N-major (B), logical shape [K, Mn]. Not supported by this function.

    • 2 BlockMajorK — Blocked along K, logical shape [K / blockK, Mn, blockK] (blockK is fixed at 128 B).

  • do_finalize (bool) – Whether to finalize the output (default True).

  • enable_pdl (bool) – Whether to enable Programmatic Dependent Launch. Auto-enabled for SM90+ when True.

  • tune_max_num_tokens (int) – Maximum number of tokens for autotuning (default 8192).

  • activation_type (int) – Activation type (default 3 — Swiglu). 3 Swiglu; 6 Relu2 (non-gated).

  • norm_topk_prob (bool) – Whether to normalize the top-k probabilities (default True).

  • routing_replay_out (Optional[torch.Tensor]) – Optional int16 tensor of shape (num_tokens_or_larger, top_k) used to capture the selected expert IDs during routing. Column order matches topk_indices. When None (default) the kernel skips the write entirely. The buffer may be larger than num_tokens for CUDA-graph pre-allocation; only rows [0, num_tokens) are written.

Returns:

If do_finalize is True returns the final MoE output (deprecated scalar return; will become [output] in v0.8.0). Otherwise returns [gemm2_output, expert_weights, expanded_idx_to_permuted_idx].

Return type:

torch.Tensor or List[torch.Tensor]