From dd5ff9dcef200b630abbf0786c482d8d94ffe3dc Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=87=91=E9=BB=84=E8=89=B2=E8=91=A1=E8=90=84=E7=90=83?= =?UTF-8?q?=E5=90=9B=E5=90=9B?= Date: Sat, 14 Feb 2026 18:50:05 +0800 Subject: [PATCH] ROCm: Add gfx950 (MI355X/CDNA4) to is_cdna() (#4051) MI355X (gfx950) has the same 1024-thread workgroup limit as MI300X (gfx942), but was missing from is_cdna(), causing all Triton kernels to use num_warps=32 (2048 threads) instead of 16 (1024 threads), resulting in OutOfResources crash. Tested on: 8x AMD Instinct MI355X (gfx950), ROCm 7.1 --- unsloth/kernels/utils.py | 1 + 1 file changed, 1 insertion(+) diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py index 5dcc7c232c..0b13c04532 100644 --- a/unsloth/kernels/utils.py +++ b/unsloth/kernels/utils.py @@ -82,6 +82,7 @@ def is_cdna(): "gfx940", "gfx941", "gfx942", + "gfx950", # CDNA4 (MI350/MI355X) )