Skip to content
KernelIndex
Search⌘K

Audio encoder to language model multimodal fusion

29 eligible runs
audio

Complete end-to-end multimodal fusion pipeline that processes raw audio features through a CTC encoder, applies windowed Q-Former downsampling with learnable queries, and scatters the resulting audio embeddings into language model token embeddings at audio token positions. This is the critical bottleneck that bridges audio and language modalities.

Fastest reported
71.0µsmean · 99% of SOL
GeometricLicense unknown · unknown

Reported evidence · last observed 2026-08-10. Reported by source; not independently reproduced. No public source. License unknown.

No public sourceRun detail →

Current records

Workloadsuite of 16 cases · mean latencynum_queries = 40 · window_size = 15 · qformer_head_dim = 64 · text_hidden_size = 4096 · encoder_input_dim = 80 · qformer_num_heads = 16 · encoder_hidden_dim = 512 · encoder_output_dim = 1024 · projector_hidden_size = 102416 cases

Not measured on H100 for this workload. Challenges →

Source-native comparison · GPU NVIDIA B200 · Workload suite of 16 cases · mean latency · Protocol SOL-ExecBench evaluation stack v1.1 · mean · 20 results · last observed 2026-08-10Record history →
#
Implementation
Latency
vs #1
Trust
Observed
1
71.0µs99% SOL
1.00×
Reported · license unknown · no source
2026-08-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
2
81.3µs98% SOL
1.15×
Reported · license unknown · no source
2026-08-09

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
3
86.3µs98% SOL
1.22×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
4
94.4µs94% SOL
1.33×
Reported · license unknown · no source
2026-07-14

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
5
115.8µs92% SOL
1.63×
Reported · license unknown · no source
2026-07-24

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
6
123.9µs94% SOL
1.75×
Reported · license unknown · no source
2026-07-16

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
7
127.3µs93% SOL
1.79×
Reported · license unknown · no source
2026-08-04

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
8
133.0µs92% SOL
1.87×
Reported · license unknown · no source
2026-07-16

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
9
135.8µs92% SOL
1.91×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
10
173.1µs85% SOL
2.44×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
11
176.9µs82% SOL
2.49×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
12
238.0µs73% SOL
3.35×
Reported · license unknown · no source
2026-08-05

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
13
316.9µs62% SOL
4.47×
Reported · license unknown · no source
2026-07-26

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
14
324.8µs66% SOL
4.58×
Reported · license unknown · no source
2026-07-24

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
15
441.2µs49% SOL
6.22×
Reported · license unknown · no source
2026-07-10

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
16
574.7µs40% SOL
8.10×
Reported · license unknown · no source
2026-07-25

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
17
607.1µs38% SOL
8.56×
Reported · license unknown · no source
2026-07-24

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
18
680.6µs35% SOL
9.59×
Reported · license unknown · no source
2026-07-23

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
19
690.9µs34% SOL
9.74×
Reported · license unknown · no source
2026-07-25

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →
20
691.8µs35% SOL
9.75×
Reported · license unknown · no source
2026-07-24

Measured exactly what you asked. Reported by source; not independently reproduced.

no sourcelicense unknownno install recipeRun detail →

Implementations

Implementation
Runtime
Best latency
Evidence
Availability
unknown
85.8µs
1.21×
Reported
license unknown · no source
unknown
94.4µs
1.33×
Reported
license unknown · no source
unknown
123.9µs
1.75×
Reported
license unknown · no source
unknown
126.8µs
1.79×
Reported
license unknown · no source
unknown
127.3µs
1.79×
Reported
license unknown · no source
unknown
115.8µs
1.63×
Reported
license unknown · no source
unknown
389.2µs
5.49×
Reported
license unknown · no source
unknown
135.8µs
1.91×
Reported
license unknown · no source
Show all 29 implementations ›
unknown
133.0µs
1.87×
Reported
license unknown · no source
unknown
171.9µs
2.42×
Reported
license unknown · no source
unknown
173.6µs
2.45×
Reported
license unknown · no source
unknown
173.1µs
2.44×
Reported
license unknown · no source
unknown
166.4µs
2.34×
Reported
license unknown · no source
unknown
176.9µs
2.49×
Reported
license unknown · no source
unknown
238.0µs
3.35×
Reported
license unknown · no source
unknown
324.8µs
4.58×
Reported
license unknown · no source
unknown
316.9µs
4.47×
Reported
license unknown · no source
unknown
331.7µs
4.68×
Reported
license unknown · no source
unknown
423.7µs
5.97×
Reported
license unknown · no source
unknown
441.2µs
6.22×
Reported
license unknown · no source
unknown
574.7µs
8.10×
Reported
license unknown · no source
unknown
607.1µs
8.56×
Reported
license unknown · no source
unknown
680.6µs
9.59×
Reported
license unknown · no source
unknown
691.8µs
9.75×
Reported
license unknown · no source
unknown
690.9µs
9.74×
Reported
license unknown · no source
unknown
1.26ms
17.81×
Reported
license unknown · no source
unknown
71.0µs
1.00×
Reported
license unknown · no source
unknown
81.3µs
1.15×
Reported
license unknown · no source
unknown
86.3µs
1.22×
Reported
license unknown · no source

Semantics

Inputs and outputs
input_featuresbf16 [batch_size, audio_seq_len, encoder_input_dim]
lm_embeddingsbf16 [batch_size, text_seq_len, text_hidden_size]
audio_token_positionsint64 [batch_size, num_audio_tokens]
encoder_input_weightbf16 [encoder_hidden_dim, encoder_input_dim]
encoder_input_biasbf16 [encoder_hidden_dim]
encoder_out_weightbf16 [encoder_output_dim, encoder_hidden_dim]
encoder_out_biasbf16 [encoder_output_dim]
encoder_out_mid_weightbf16 [encoder_hidden_dim, encoder_output_dim]
encoder_out_mid_biasbf16 [encoder_hidden_dim]
learnable_queriesbf16 [1, num_queries, projector_hidden_size]
qformer_q_proj_weightbf16 [projector_hidden_size, projector_hidden_size]
qformer_q_proj_biasbf16 [projector_hidden_size]
qformer_k_proj_weightbf16 [projector_hidden_size, encoder_hidden_dim]
qformer_k_proj_biasbf16 [projector_hidden_size]
qformer_v_proj_weightbf16 [projector_hidden_size, encoder_hidden_dim]
qformer_v_proj_biasbf16 [projector_hidden_size]
qformer_out_proj_weightbf16 [projector_hidden_size, projector_hidden_size]
qformer_out_proj_biasbf16 [projector_hidden_size]
projector_weightbf16 [text_hidden_size, projector_hidden_size]
projector_biasbf16 [text_hidden_size]
fused_embeddingsbf16 [batch_size, text_seq_len, text_hidden_size]
Axes and behavior
batch_sizevariable
num_queriesconstant = 40
window_sizeconstant = 15
text_seq_lenvariable
audio_seq_lenvariable
num_audio_tokensvariable
qformer_head_dimqformer_head_dim = projector_hidden_size // qformer_num_heads
text_hidden_sizeconstant = 4096
encoder_input_dimconstant = 80
qformer_num_headsconstant = 16
encoder_hidden_dimconstant = 512
encoder_output_dimconstant = 1024
projector_hidden_sizeconstant = 1024
determinismunspecified
constraintsNo mutation or aliasing
Identity
alias045_audio_encoder_to_language_model_multimodal_fusionmodelgranite-speech-3-3-8bsha256a13254c7df4c…
Sources: NVIDIA SOL-ExecBench (2026-08-10)last observed 2026-08-10How records are decidedJSON