JohnQinAMDandClaude Opus 5 ef5f7cd119 [ROCm][DSv4.1][Perf] Dequantize the MXFP8 weight once when dot_scaled cannot be used (#56560) ...

Signed-off-by: JohnQinAMD <yanyuan.qin@amd.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

2026-09-15 09:07:26 +08:00

..

cute_dsl

[Kimi-K3][Perf] Prefetch ll_bf16 router weights for M=1 (#53524)

2026-09-01 21:40:19 +00:00

mixed_precision

[ROCm][Perf] W4A16: keep skinny GEMM zero-points packed 4-bit (#54965)

2026-09-14 17:40:25 +08:00

mxfp4

[Quantization] Select linear backends per quantization (#51204)

2026-09-14 16:24:18 -04:00

mxfp6

[Docs] Fix two docs build warnings (#51014)

2026-08-04 05:29:04 -07:00

mxfp8

[ROCm][DSv4.1][Perf] Dequantize the MXFP8 weight once when dot_scaled cannot be used (#56560)

2026-09-15 09:07:26 +08:00

nvfp4

[Fast Start] Support fp4 (#55465)

2026-09-10 13:22:17 +00:00

scaled_mm

[Bugfix][CPU] Fix DeepSeek-R1 (FP8 MLA + MoE) correctness on CPU backend (#56773)

2026-09-14 17:30:28 +08:00

__init__.py

[Quantization] Select linear backends per quantization (#51204)

2026-09-14 16:24:18 -04:00

base.py

[Refactor] Remove dead code for quantization (#52221)

2026-08-16 13:10:16 -04:00

zentorch_utils.py

[CPU][Zen] Route BF16 MoE inference through zentorch on AMD (#44201)

2026-08-10 23:20:05 -07:00