com.microsoft.PackedMultiHeadAttention

com.microsoft · ONNX Runtime contrib operator · contrib since_version 1

Description

Multi-head self-attention over a padding-removed token stream. The token_offset and cumulative_sequence_length schedule maps every packed token to its sequence, and a token attends only the keys of that sequence. Query is either packed [T, N, 3, H] with key and value absent, or a separate [T, hidden] triple. The optional attention bias is indexed in PADDED coordinates. Float32 and float16 only; bfloat16 is not implemented.

See the ONNX Runtime PackedMultiHeadAttention contrib-operator spec for the reference semantics.

Inputs

Outputs

Attributes

Attributes and default values (overridable per request):

Type constraints

Implementation variants

One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers.

• packed_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• packed_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• packed_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• packed_attn_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• packed_attn_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• packed_attn_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• packed_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• packed_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• packed_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• packed_bias_attn_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• packed_bias_attn_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• packed_bias_attn_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• separate_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• separate_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• separate_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• separate_attn_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• separate_attn_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• separate_attn_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• separate_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• separate_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• separate_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.
• separate_bias_attn_bias_flash — Tiled attention with independent Q/K and V slices and subgroup dot reductions. Equal wide heads target four vec4 fragments per lane, bounded by subgroup/workgroup limits and native staging capacity. Unequal tiles cap work by token count and yield to serial or wider portable clusters when reuse is limited.
• separate_bias_attn_bias_flash_nosg — Portable tiled attention with independent query/key and value slices. Lanes cover the larger width, with guarded zero Q/K fragments and shared reductions. Device limits bound lane count and staging. Unequal-width tiles yield to serial attention below the query-reuse target.
• separate_bias_attn_bias_serial — Portable serial-key attention: one workgroup owns one (packed token, head), walks that token's own sequence and reduces each dot through workgroup memory. Head dimensions need no alignment and the value head dimension may differ.

Device requirements

Some implementation variants require subgroups. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype.

Files

• metadata.json — kernel metadata (id, digests, per-variant templates, provenance)
• manifest.json — the op contract (source of truth)
• test.json — correctness cases
• bench.json — benchmark cases
• attn-packed-varlen-cluster.wgsl.jinja
• attn-packed-varlen-scalar.wgsl.jinja

Use with @huggingface/kernels

npm install --save-exact @huggingface/kernels@0.0.1-preview.3

Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically.

The version: 1 option selects the published kernel contract; it is independent of any operator opset, contrib since_version, or model version. It follows the v1 branch as fixes land. To pin exact artifact bytes, pass a 40-character commit revision instead of version.

Replace each *Data placeholder with a typed array containing the corresponding input data.

import { getKernel } from "@huggingface/kernels";

const kernel = await getKernel("webgpu-kernels/com.microsoft.PackedMultiHeadAttention", { version: 1 });
const { outputT } = await kernel({
  queryT: { data: queryTData, shape: [1, 16] },
  keyT: { data: keyTData, shape: [1, 16] },
  valueT: { data: valueTData, shape: [1, 20] },
  tokenOffsetT: { data: tokenOffsetTData, shape: [1, 1] },
  cumulativeSequenceLengthT: { data: cumulativeSequenceLengthTData, shape: [2] },
}, {
  attrs: { num_heads: 1 },
});