For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo struct

DistributedEPDispatchBlockScaledNVDynGlobalScales

struct DistributedEPDispatchBlockScaledNVDynGlobalScales

Registers the mo.distributed.ep.dispatch.block.scaled.nv.dyn_global_scales graph op with the graph compiler.

Implemented traits

AnyType, Deinitable, Movable

Methods

execute

static def execute[input_dtype: DType, dispatch_dtype: DType, dispatch_scale_dtype: DType, hidden_size: Int, top_k: Int, n_experts: Int, max_token_per_rank: Int, n_gpus_per_node: Int, n_nodes: Int, fused_shared_expert: Bool, //, target: StringSpan[ImmStaticOrigin], _trace_name: StringSpan[ImmStaticOrigin]](output_tokens: VariadicTensors[IOSpec[_, _].Output, static_specs=output_tokens.static_specs], output_scales: VariadicTensors[IOSpec[_, _].Output, static_specs=output_scales.static_specs], output_rowwise_scales: VariadicTensors[IOSpec[_, _].Output, static_specs=output_rowwise_scales.static_specs], row_offsets: VariadicTensors[IOSpec[_, _].Output, static_specs=row_offsets.static_specs], scales_offsets: VariadicTensors[IOSpec[_, _].Output, static_specs=scales_offsets.static_specs], expert_ids: VariadicTensors[IOSpec[_, _].Output, static_specs=expert_ids.static_specs], src_info: VariadicTensors[IOSpec[_, _].Output, static_specs=src_info.static_specs], input_tokens: VariadicTensors[IOSpec[_, _].Input, static_specs=input_tokens.static_specs], topk_ids: VariadicTensors[IOSpec[_, _].Input, static_specs=topk_ids.static_specs], send_ptrs: VariadicTensors[IOSpec[_, _].Input, static_specs=send_ptrs.static_specs], recv_ptrs: VariadicTensors[IOSpec[_, _].Input, static_specs=recv_ptrs.static_specs], recv_count_ptrs: VariadicTensors[IOSpec[_, _].Input, static_specs=recv_count_ptrs.static_specs], atomic_counters: VariadicTensors[IOSpec[_, _].MutableInput, static_specs=atomic_counters.static_specs], dev_ctxs: DeviceContextArray)

Multi-device fused Expert Parallelism NVFP4 dispatch with dynamic global scales.

Launches the EP dispatch kernel on all devices simultaneously via _ep_launch_device_collective. Each device routes its tokens to experts based on top-k IDs, quantizes each token to NVFP4 against its own global scale, 2688 / rowmax, and sends it to the appropriate peer devices. The receivers store the scales' BF16 inverses in output_rowwise_scales.

Parameters:

• ​input_dtype (DType): DType of the input tokens before quantization (inferred).
• ​dispatch_dtype (DType): DType used for the quantized token payload during dispatch (inferred).
• ​dispatch_scale_dtype (DType): DType of the block scales accompanying the dispatched tokens (inferred).
• ​hidden_size (Int): Size of the model's hidden dimension (inferred).
• ​top_k (Int): Number of experts each token is routed to (inferred).
• ​n_experts (Int): Total number of experts across all GPUs (inferred).
• ​max_token_per_rank (Int): Maximum number of tokens per GPU (inferred).
• ​n_gpus_per_node (Int): Number of GPUs per node (inferred).
• ​n_nodes (Int): Number of physical nodes (inferred).
• ​fused_shared_expert (Bool): Whether a shared expert is fused into the dispatch kernel (inferred).
• ​target (StringSpan[ImmStaticOrigin]): Compile-time device target.
• ​_trace_name (StringSpan[ImmStaticOrigin]): Trace label for this op.

Args:

• ​output_tokens (VariadicTensors[IOSpec[_, _].Output, static_specs=output_tokens.static_specs]): Output variadic tensors storing the dispatched NVFP4-quantized tokens, one per device.
• ​output_scales (VariadicTensors[IOSpec[_, _].Output, static_specs=output_scales.static_specs]): Output variadic tensors storing the NVFP4 block scales, one per device.
• ​output_rowwise_scales (VariadicTensors[IOSpec[_, _].Output, static_specs=output_rowwise_scales.static_specs]): Output variadic tensors storing each received token's BF16 inverse global scale, one per device.
• ​row_offsets (VariadicTensors[IOSpec[_, _].Output, static_specs=row_offsets.static_specs]): Output variadic tensors storing the row offsets for the received tokens, one per device.
• ​scales_offsets (VariadicTensors[IOSpec[_, _].Output, static_specs=scales_offsets.static_specs]): Output variadic tensors storing the offsets into the scales buffer, one per device.
• ​expert_ids (VariadicTensors[IOSpec[_, _].Output, static_specs=expert_ids.static_specs]): Output variadic tensors storing the expert ID for each received token, one per device.
• ​src_info (VariadicTensors[IOSpec[_, _].Output, static_specs=src_info.static_specs]): Output variadic tensors recording the originating rank and token index for each received token, one per device.
• ​input_tokens (VariadicTensors[IOSpec[_, _].Input, static_specs=input_tokens.static_specs]): Input variadic tensors of tokens to dispatch, one per device.
• ​topk_ids (VariadicTensors[IOSpec[_, _].Input, static_specs=topk_ids.static_specs]): Input variadic tensors of top-k expert IDs per token, one per device.
• ​send_ptrs (VariadicTensors[IOSpec[_, _].Input, static_specs=send_ptrs.static_specs]): Send buffer pointers for the dispatch phase, one per device.
• ​recv_ptrs (VariadicTensors[IOSpec[_, _].Input, static_specs=recv_ptrs.static_specs]): Receive buffer pointers for the dispatch phase, one per device.
• ​recv_count_ptrs (VariadicTensors[IOSpec[_, _].Input, static_specs=recv_count_ptrs.static_specs]): Receive count buffer pointers tracking tokens received per expert, one per device.
• ​atomic_counters (VariadicTensors[IOSpec[_, _].MutableInput, static_specs=atomic_counters.static_specs]): Atomic counters coordinating work across thread blocks, one per device.
• ​dev_ctxs (DeviceContextArray): List of GPU device contexts, one per device.