Search results

Performance bottleneck in TransformerLayer and Attention when using attention mask

bugSomething isn't workingSomething isn't working

Status: Open.

#2703 In NVIDIA/TransformerEngine;

· yang-cx opened on Feb 25, 2026

[JAX] Optimize materialized attn mask to consume O(N) memory instead of O(N*N)

attention

jax

Status: Open.

Task

#2700 In NVIDIA/TransformerEngine;

· KshitijLakhani opened on Feb 23, 2026

[JAX] Expose sigmoid GLU activation type

jax

Priority = P1

Status: Open.

Enhancement

#2687 In NVIDIA/TransformerEngine;

· jberchtold-nvidia opened on Feb 17, 2026

JAX vmap issue with TE Attention

attention

bugSomething isn't workingSomething isn't working

field-request

jax

Status: Open.

#2685 In NVIDIA/TransformerEngine;

· jmukherjee-nvidia opened on Feb 17, 2026

Example of <code>quantized_model_init</code> for low-precision compute weights and fp32 main weights with fsdp2

enhancementNew feature or requestNew feature or request

Status: Open.

#2682 In NVIDIA/TransformerEngine;

· pstjohn opened on Feb 13, 2026

te.autocast() and fully_shard doesn't de-allocate quantized weights until the backwards pass

bugSomething isn't workingSomething isn't working

Status: Open.

#2681 In NVIDIA/TransformerEngine;

· pstjohn opened on Feb 13, 2026

Build failure in cublaslt_gemm.cu: "cuda" is ambiguous

bugSomething isn't workingSomething isn't working

Status: Open.

#2679 In NVIDIA/TransformerEngine;

· XueSongTap opened on Feb 13, 2026

[Inquiry] Support status and roadmap for MXFP8 on SM120 (Blackwell)

enhancementNew feature or requestNew feature or request

Status: Open.

#2668 In NVIDIA/TransformerEngine;

· Mnb66 opened on Feb 10, 2026

[Core] TE common nvte_grouped_gemm workspace is incorrect

bugSomething isn't workingSomething isn't working

MoE

Status: Open.

#2660 In NVIDIA/TransformerEngine;

· jberchtold-nvidia opened on Feb 9, 2026

[Core] TE common nvte_grouped_gemm treats output layout as column-wise instead of rowwise

bugSomething isn't workingSomething isn't working

MoE

Status: Open.

#2659 In NVIDIA/TransformerEngine;

· jberchtold-nvidia opened on Feb 9, 2026

[JAX] Upstream checkpointing of quantizations in TE/JAX to MaxText

field-request

fp4

FP8

jax

Status: Open.

#2658 In NVIDIA/TransformerEngine;

· jberchtold-nvidia opened on Feb 6, 2026

debug api scale_inv_min is always zero when using padding or divisibility requirements

bugSomething isn't workingSomething isn't working

Status: Open.

#2628 In NVIDIA/TransformerEngine;

· jomitchellnv opened on Jan 27, 2026