【训练营】小模型训练支持 - #220

Open

accelerator-llc wants to merge 16 commits into

InfiniTensor:masterfrom

accelerator-llc:feat/small-model-cnn

Commits on Sep 18, 2026

fix: use shared_ptr for MNIST network to avoid bad_weak_ptr
accelerator-llc
committed
6334835
feat(conv): add Conv2d operator for CPU with autograd and unit tests
accelerator-llc
committed
ac49bd6
feat(conv): add CUDA Conv2d kernels with autograd and unit tests
accelerator-llc
committed
67dace8
fix(conv): guard Conv2d CPU kernels against non-FP32 inputs
accelerator-llc
committed
93a20d8
feat(relu): add ReLU activation with autograd for CPU and CUDA
accelerator-llc
committed
e141a59
fix(mnist): compute dataset sample stride from the normalized dtype
accelerator-llc
committed
bbe2163
feat(mnist): add CNN demo selectable via --model flag
accelerator-llc
committed
c064d06
feat(mnist): add DDP distributed training via infini_run
accelerator-llc
committed
3ebb99e
test(linear): add bias-gradient regression tests
accelerator-llc
committed
3c7780a
refactor(mnist): invoke modules through operator()
accelerator-llc
committed
de76ce8
test(conv): merge device-split conv tests into parameterized files
accelerator-llc
committed
abb73d2
perf(conv): batch the conv GEMMs and drop redundant col buffer init
accelerator-llc
committed
380d33f
test(conv): merge the device-split conv train tests
accelerator-llc
committed
bbf6d4c
chore(kernels): fix include hygiene in conv and relu
accelerator-llc
committed
1e662ec
feat(mnist): report test metrics every epoch and log the training step
accelerator-llc
committed
acfbe87
fix(tests): adapt linear backward test to shared test utility changes
accelerator-llc
committed
c8de4c8

Navigation Menu

【训练营】小模型训练支持 - #220

【训练营】小模型训练支持#220
accelerator-llc wants to merge 16 commits into
InfiniTensor:masterfrom
accelerator-llc:feat/small-model-cnn

Commits on Sep 18, 2026

fix: use shared_ptr for MNIST network to avoid bad_weak_ptr

feat(conv): add Conv2d operator for CPU with autograd and unit tests

feat(conv): add CUDA Conv2d kernels with autograd and unit tests

fix(conv): guard Conv2d CPU kernels against non-FP32 inputs

feat(relu): add ReLU activation with autograd for CPU and CUDA

fix(mnist): compute dataset sample stride from the normalized dtype

feat(mnist): add CNN demo selectable via --model flag

feat(mnist): add DDP distributed training via infini_run

test(linear): add bias-gradient regression tests

refactor(mnist): invoke modules through operator()

test(conv): merge device-split conv tests into parameterized files

perf(conv): batch the conv GEMMs and drop redundant col buffer init

test(conv): merge the device-split conv train tests

chore(kernels): fix include hygiene in conv and relu

feat(mnist): report test metrics every epoch and log the training step

fix(tests): adapt linear backward test to shared test utility changes