perf: faster GPU decode: one SSM_CONV for the depthwise conv, less host work per step by kittyandrew · Pull Request #102 · mudler/parakeet.cpp

test: check both depthwise conv paths against NeMo's conv

kittyandrew

committed

89f316d

perf: skip the attention mask when it masks nothing

kittyandrew

committed

5059837

test: check that the attention mask is skipped only when it masks nothing

kittyandrew

committed

0b34ba1

perf: build the relative position encoding once per length

kittyandrew

committed

74fe432

test: check the cached relative position table against a fresh one

kittyandrew

committed

9a64194

perf: share GpuMel's DFT basis between instances

kittyandrew

committed

6748975

test: check GpuMel's shared DFT basis, and GpuMel on a GPU

kittyandrew

committed

69c6ca7

perf: wait once per compute on the persistent backend

kittyandrew

committed

f6e161c

test: check the single-wait readback on GPUs

kittyandrew

committed

44e5a1e

perf: drop the copies of the LSTM gate views

kittyandrew

committed

6ed7b75

perf: run the prediction step and the joint as one graph

kittyandrew

committed

29f5112

test: check the fused decoder step against the separate graphs

kittyandrew

committed

d7dd3f2

perf: frame GpuMel's input without copying the signal twice

kittyandrew

committed

690050b

test: check GpuMel's framing against the two-copy framing

kittyandrew

committed

b8a30f6

perf: read the relative shift through a strided view

kittyandrew

committed

7dd6027

test: check the strided rel_shift view against NeMo's rel_shift

kittyandrew

committed

7e03f14

Navigation Menu

Repository navigation

perf: faster GPU decode: one SSM_CONV for the depthwise conv, less host work per step - #102

perf: faster GPU decode: one SSM_CONV for the depthwise conv, less host work per step#102
kittyandrew wants to merge 17 commits into
mudler:masterfrom
kittyandrew:depthwise-ssm-conv

Commits on Oct 18, 2026

perf: run the conformer's depthwise conv as one SSM_CONV on GPU backends

Commits on Oct 24, 2026

test: check both depthwise conv paths against NeMo's conv

perf: skip the attention mask when it masks nothing

test: check that the attention mask is skipped only when it masks nothing

perf: build the relative position encoding once per length

test: check the cached relative position table against a fresh one

perf: share GpuMel's DFT basis between instances

test: check GpuMel's shared DFT basis, and GpuMel on a GPU

perf: wait once per compute on the persistent backend

test: check the single-wait readback on GPUs

perf: drop the copies of the LSTM gate views

perf: run the prediction step and the joint as one graph

test: check the fused decoder step against the separate graphs

perf: frame GpuMel's input without copying the signal twice

test: check GpuMel's framing against the two-copy framing

perf: read the relative shift through a strided view

test: check the strided rel_shift view against NeMo's rel_shift