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

Struct_ep_dispatch_block_scaled_nv_dyn_global_scales

struct Struct_ep_dispatch_block_scaled_nv_dyn_global_scales

Registers the 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, skip_a2a: Bool, allreduce_world_size: Int, //, target: StringSpan[ImmStaticOrigin]](output_tokens: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_tokens.static_spec], output_scales: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_scales.static_spec], output_rowwise_scales: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_rowwise_scales.static_spec], row_offsets: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=row_offsets.static_spec], scales_offsets: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=scales_offsets.static_spec], expert_ids: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=expert_ids.static_spec], src_info: ManagedTensorSlice[IOSpec[_, _].Output, static_spec=src_info.static_spec], atomic_counters: ManagedTensorSlice[IOSpec[_, _].MutableInput, static_spec=atomic_counters.static_spec], input_tokens: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input_tokens.static_spec], topk_ids: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=topk_ids.static_spec], send_ptrs: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=send_ptrs.static_spec], recv_ptrs: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=recv_ptrs.static_spec], recv_count_ptrs: ManagedTensorSlice[IOSpec[_, _].Input, static_spec=recv_count_ptrs.static_spec], context: DeviceContext)

Execute the fused Expert Parallelism NVFP4 dispatch kernel with dynamic global scales.

Each token is quantized against its own global scale, 2688 / rowmax, rather than a caller-supplied input scale. The scale's BF16 inverse travels with the token and lands in output_rowwise_scales, so a received row dequantizes as fp4 * block_scale * rowwise_scale.

Parameters:

• ​input_dtype (DType): DType of the input tokens before dispatch (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).
• ​skip_a2a (Bool): Whether to skip the all-to-all communication and send tokens only within the current device (inferred).
• ​allreduce_world_size (Int): Number of ranks participating in the allreduce following dispatch (inferred).
• ​target (StringSpan[ImmStaticOrigin]): Compile-time device target.

Args:

• ​output_tokens (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_tokens.static_spec]): Output tensor storing the received tokens in NVFP4 format.
• ​output_scales (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_scales.static_spec]): Output tensor storing the NVFP4 block scales for the received tokens.
• ​output_rowwise_scales (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=output_rowwise_scales.static_spec]): Output tensor storing each received token's BF16 inverse global scale, rowmax / 2688. Shape [max_recv_tokens], one entry per row of output_tokens.
• ​row_offsets (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=row_offsets.static_spec]): Output tensor storing the row offsets for the received tokens.
• ​scales_offsets (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=scales_offsets.static_spec]): Output tensor storing the offsets into the scales buffer for the received tokens.
• ​expert_ids (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=expert_ids.static_spec]): Output tensor storing the expert ID for each received token.
• ​src_info (ManagedTensorSlice[IOSpec[_, _].Output, static_spec=src_info.static_spec]): Output tensor recording the originating rank and token index for each received token. Shape [num_tokens, 2].
• ​atomic_counters (ManagedTensorSlice[IOSpec[_, _].MutableInput, static_spec=atomic_counters.static_spec]): Atomic counters coordinating work across thread blocks during the dispatch phase.
• ​input_tokens (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=input_tokens.static_spec]): Input tokens to dispatch to experts. Shape [num_tokens, hidden_size].
• ​topk_ids (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=topk_ids.static_spec]): Input tensor of top-k expert IDs per token. Shape [num_tokens, top_k].
• ​send_ptrs (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=send_ptrs.static_spec]): Send buffer pointers for the dispatch phase.
• ​recv_ptrs (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=recv_ptrs.static_spec]): Receive buffer pointers for the dispatch phase.
• ​recv_count_ptrs (ManagedTensorSlice[IOSpec[_, _].Input, static_spec=recv_count_ptrs.static_spec]): Receive count buffer pointers tracking tokens received per expert.
• ​context (DeviceContext): GPU device context for the current device.