Search results

MXFP8 uses logical block size 32, but some ROCm training paths still hard-require block_len == 128

bugSomething isn't workingSomething isn't working

status: triage

Status: Open.

#539 In ROCm/TransformerEngine;

· ramgandikota opened on Apr 13, 2026

[BUG] gfx1201 (RDNA4 / Radeon AI PRO R9700) not in AITER arch table — FP8 WMMA silently falls back to FP32

status: triage

Status: Open.

#520 In ROCm/TransformerEngine;

· mohanapgvb opened on Apr 6, 2026

ValueError: ROCm is not supported by requested frameworks: ['pytorch']

Status: Open.

#396 In ROCm/TransformerEngine;

· AmdSampsa opened on Dec 5, 2025

Enable FP8 (E4M3/E5M2) kernel execution for Radeon AI PRO R9700 (gfx1201, RDNA4)

enhancementNew feature or requestNew feature or request

status: triage

Status: Open.

#359 In ROCm/TransformerEngine;

· meven3000 opened on Nov 2, 2025

nvfp4

enhancementNew feature or requestNew feature or request

Status: Open.

#358 In ROCm/TransformerEngine;

· ehartford opened on Nov 1, 2025

[Issue]: Build system flaws report

Status: Open.

#215 In ROCm/TransformerEngine;

· Arech8 opened on Jun 25, 2025

[Out of Box Experience]: ROCm Transformer Engine Should Be Included in AMD Pytorch Images

Under Investigation

Status: Open.

#82 In ROCm/TransformerEngine;

· functionstackx opened on Oct 18, 2024

MI300X FP8 TE.Linear 2x Slower than AMP BF16 F.Linear

Status: Open.

#73 In ROCm/TransformerEngine;

· functionstackx opened on Oct 13, 2024

[1xMI300X] GPT-2 XL 1.5B FP8 Training ~30% slower than H100 FP8

Status: Open.

#72 In ROCm/TransformerEngine;

· functionstackx opened on Oct 13, 2024