Search results

Support 32x32 scaling for weights in MXFP8 weight quantization kernel

moe

mx

Status: Open.

#4185 In pytorch/ao;

· danielvegamyhre opened on Mar 26, 2026

·  

MXFP8 Training

fuse dynamic per group padding into cutedsl 2d MXFP8 quantization kernel

moe

mx

Status: Open.

#4184 In pytorch/ao;

· danielvegamyhre opened on Mar 26, 2026

·  

MXFP8 Training

[MXFP8 MoE] mx_block_rearrange_2d_M_groups_cuda fails for models with > 32 experts

module: trainingquantize_ api training flowquantize_ api training flow

moe

mx

Status: Open.

#4163 In pytorch/ao;

· yzhautouskay opened on Mar 24, 2026

·  

MXFP8 Training

# Feature Request: Support Async-Stream Transfer for [AffineQuantizedTensor] (Fix diffusers group_offload device mismatch)

Status: Open.

#4158 In pytorch/ao;

· zhangvia opened on Mar 24, 2026

mxfp8 RCEIL scaling should use optimized PTX instructions on CUDA

Status: Open.

#4140 In pytorch/ao;

· vkuzo opened on Mar 23, 2026

Buffer mutation tracking lost after convert_pt2e for index_put_

Status: Open.

#4138 In pytorch/ao;

· tom-arm opened on Mar 23, 2026

Linear+BatchNorm fusion produces incorrect model in most cases

Status: Open.

#4116 In pytorch/ao;

· StrycekSimon opened on Mar 19, 2026

torch.AcceleratorError: CUDA error: an illegal memory access was encountere

Status: Open.

#4114 In pytorch/ao;

· BiologicalExplosion opened on Mar 19, 2026

When will support be provided for customizing the momentum value and eps value in BN?

Status: Open.

#4107 In pytorch/ao;

· PhilCuriosity opened on Mar 18, 2026

./test/float8/test_everything_multi_gpu.sh failing in 4xH100 CI

float8

module: trainingquantize_ api training flowquantize_ api training flow

Status: Open.

Bug

#4106 In pytorch/ao;

· danielvegamyhre opened on Mar 18, 2026

·  

FP8 Rowwise Training

fp8 blockwise test failure on ROCM in CI

module: trainingquantize_ api training flowquantize_ api training flow

topic: rocm

Status: Open.

Bug

#4104 In pytorch/ao;

· danielvegamyhre opened on Mar 18, 2026

·  

FP8 Blockwise Training

Blockwise FP8 training seems to have much lower peak memory usage than others

float8

module: trainingquantize_ api training flowquantize_ api training flow

Status: Open.

#4101 In pytorch/ao;

· CFC87 opened on Mar 17, 2026

·  

FP8 Blockwise Training